spatial decision support systems for large arc routing...
TRANSCRIPT
Spatial Decision Support Systems
for Large Arc Routing Problems
Peter Bernard Keenan BComm MMangtSc
Dissertation presented to the Department of Management Information
Systems, Faculty of Commerce, University College Dublin, in partial
fulfilment of the requirements of the degree of Doctor of Philosophy.
June 2001
Supervisor: Professor H. C. Harrison
This dissertation is dedicated to my parents, for their encouragement and
support from the earliest days of my education.
i
Acknowledgements
I would like to express my appreciation to my supervisor, Professor
Harold Harrison for his involvement from the inception to the completion
of this dissertation. I would also like to thank my colleagues in the MIS
Department, especially Dr Cathal Brugha, for their support and
encouragement. Recognition is also due to everyone in UCD and
University Software Systems who has worked with me on arc routing
applications over the years.
ii
Abstract
The transport sector is a vital component of business operations and is of
considerable economic importance in modern economies. Better decision
making in this sector can make a significant contribution to business
operations. Information technology can greatly facilitate decision making,
especially in the form of Decision Support Systems (DSS). Transport
problems generally, and routing problems in particular, represent an
important area of DSS application. Within the routing field, most work
has centred on node routing problems; with much less attention paid to
arc routing problems. This dissertation reviews developments in DSS for
routing problems generally, and discusses in detail the design issues for a
DSS for large arc routing problems. This class of problem is especially
relevant for arc routing applications on rural road networks in Ireland.
Geographic Information Systems (GIS) technology is identified as
providing the means to enhance routing DSS. An extensive review is
provided of the GIS field and its relationship to other information
systems. This is followed by a comprehensive overview of the relationship
between routing and GIS. This analysis suggests that the integration of
GIS with routing models can allow the development of a Spatial Decision
Support System (SDSS) to provide superior decision support for this class
of problem.
The specific issues of building a GIS based DSS for large arc time
capacitated routing problems are discussed. Existing arc routing lower
bound procedures are modified for the time-capacitated problem. Three
potential heuristic modelling approaches for this problem are identified;
one approach was rejected as unpromising. The other two heuristics are
extensively tested against the modified lower bounds for the problem and
are evaluated from the point of view of probable user acceptability. This
work indicated that these heuristics could form the basis of an effective
DSS for this class of problem.
iii
Contents
Acknowledgements............................................................................. i
List of Tables .................................................................................... ix
List of Figures .................................................................................. xi
Glossary .......................................................................................... xiv
CHAPTER 1 : INTRODUCTION ............................................................. 1
1.1 Introduction................................................................................. 1
1.1.1 Preface.................................................................................................................... 1
1.1.2 Information Technology........................................................................................ 1
1.1.3 Growth of Information Systems (IS) .................................................................... 3
1.2 Decision Support Systems .......................................................... 4
1.2.1 Decision Structure................................................................................................. 4
1.1.2 Definitions of DSS................................................................................................. 6
1.1.3 The DSS field ...................................................................................................... 11
1.1.4 Contributions of DSS to decision-making ......................................................... 13
1.3 DSS Technology......................................................................... 14
1.3.1 Building DSS ...................................................................................................... 14
1.3.2 DSS generator approach..................................................................................... 15
1.3.3 ROMC approach.................................................................................................. 17
1.3.4 Current DSS technology ..................................................................................... 17
1.4 The nature of this research in DSS.......................................... 19
1.4.1 The current state of DSS..................................................................................... 19
1.4.2 The structure of this dissertation ....................................................................... 19
CHAPTER 2 : DECISION SUPPORT FOR ROUTING PROBLEMS
..................................................................................................................... 21
2.1 Routing DSS .............................................................................. 21
2.1.1 The nature of routing problems.......................................................................... 21
2.1.2 Suitability of routing problems for DSS ............................................................ 22
2.1.3 Early routing Software ....................................................................................... 23
2.2 Traditional Vehicle Routing DSS............................................. 26
2.2.1 Requirement for user interaction........................................................................ 26
2.2.2 The use of graphics in routing systems .............................................................. 27
iv
1.3 Impact of DSS Developments on Routing DSS ....................... 30
1.1.1 DSS developments............................................................................................... 30
1.1.2 Artificial Intelligence .......................................................................................... 31
1.1.3 User Defined algorithms..................................................................................... 32
1.1.4 Visual Interactive techniques ............................................................................. 33
1.4 Arc Routing Systems................................................................. 34
1.4.1 Characteristics of Arc Routing DSS................................................................... 34
1.4.2 Specific arc routing applications........................................................................ 35
1.4.3 Comprehensive arc routing software packages.................................................. 36
1.5 The Future of Routing DSS ...................................................... 39
1.5.1 Trends in Routing DSS....................................................................................... 39
1.1.2 Geographic Information Systems ....................................................................... 39
CHAPTER 3 : GEOGRAPHIC INFORMATION SYSTEMS.............. 41
3.1 Geographic Information Systems (GIS)................................... 41
3.1.1 Development of GIS Technology......................................................................... 41
1.1.2 GIS Data.............................................................................................................. 42
1.2 GIS Software ............................................................................. 44
1.2.1 Current developments in GIS software .............................................................. 44
1.2.2 Components of GIS.............................................................................................. 45
1.3 GIS and DSS.............................................................................. 47
1.3.1 Relationship between GIS and DSS research.................................................... 47
1.3.2 Is a GIS a DSS?................................................................................................... 49
1.4 Spatial Decision Support Systems ........................................... 52
1.4.1 SDSS decision makers ........................................................................................ 52
1.4.2 GIS as a DSS Generator ..................................................................................... 55
1.4.3 Extending GIS to a broader community ............................................................ 57
1.5 Current Spatial Decision Support Systems Technology ......... 60
1.5.1 Suitability of GIS Software for building DSS................................................... 60
1.5.2 Commercial GIS software................................................................................... 64
1.5.3 Future directions ................................................................................................. 66
CHAPTER 4 : MODELLING ROUTING PROBLEMS IN GIS ......... 67
4.1 Introduction .............................................................................. 67
4.1.1 GIS and routing .................................................................................................. 67
4.1.2 Spatial Decision Support Systems for routing .................................................. 68
v
4.2 Information Requirements for Routing ................................... 69
4.2.1 Categories of routing data .................................................................................. 69
4.2.2 Interdependence of routing parameters ............................................................. 73
4.2.3 Spatial interactions in routing problems........................................................... 78
4.3 The Role of GIS in Supporting Routing Problems................... 83
4.3.1 Routing problems supported by traditional DSS.............................................. 83
4.3.2 Routing problems requiring GIS support .......................................................... 85
4.3.3 Spatially complex routing problems .................................................................. 86
4.4 Implementing routing SDSS .................................................... 88
4.4.1 Data requirements for routing system implementation .................................... 88
4.4.2 Using GIS data in routing DSS ......................................................................... 89
4.5 The future of GIS and routing.................................................. 92
4.5.1 Types of Routing Software.................................................................................. 92
4.5.2 The use of GIS for routing .................................................................................. 93
CHAPTER 5 : ARC ROUTING PROBLEMS ....................................... 96
5.1 Routing Problems...................................................................... 96
5.1.1 Background ......................................................................................................... 96
5.1.2 Types of routing problem .................................................................................... 96
5.1.3 Problem complexity ............................................................................................. 98
5.2 Arc Routing Problems ............................................................. 100
5.2.1 The Chinese Postman Problem......................................................................... 100
5.2.2 The Rural Postman Problem ............................................................................ 102
5.2.3 Other Uncapacitated Arc Problems.................................................................. 103
5.3 The Capacitated Arc Routing Problem (CARP)..................... 104
5.3.1 Definition of CARP............................................................................................ 104
5.3.2 Representing Arc Routing problems as a TSP................................................. 105
5.4 Linear programming formulations of CARP ......................... 106
5.4.1 Golden and Wong formulation ......................................................................... 106
5.4.2 Belenguer and Benavent formulation .............................................................. 108
5.5 Lower bounds for the CARP ................................................... 109
5.5.1 Early graph theory bounds ............................................................................... 109
5.5.2 Node Duplication Lower Bound ....................................................................... 110
5.5.3 Bound LB1......................................................................................................... 114
5.5.4 Bounds exploiting cuts away from the depot ................................................... 117
5.5.5 Cutting Plane Bound ........................................................................................ 125
vi
CHAPTER 6 : SOLUTIONS FOR THE CARP................................... 126
6.1 Heuristic Solutions for the Capacitated Arc Routing Problem
........................................................................................................ 126
6.1.1 Single Pass heuristics ....................................................................................... 126
6.1.2 Route-first cluster-second heuristics ................................................................ 128
6.1.3 Cluster-first route-second heuristics ................................................................ 129
6.1.4 New approaches to CARP ................................................................................. 130
6.1.5 Real world network representation issues ....................................................... 131
6.2 Branching approaches to CARP............................................. 133
6.2.1 Branching techniques ....................................................................................... 133
6.2.2 The Tour Construction Algorithm.................................................................... 134
6.2.3 Practical feasibility of branch and bound algorithm...................................... 140
CHAPTER 7 : CARP ON IRISH RURAL NETWORKS ................... 142
7.1 Postal delivery of rural networks ........................................... 142
7.1.1 Irish Rural Road Networks .............................................................................. 142
7.1.2 Rural Postal Delivery........................................................................................ 146
7.2 CARP on large sparse networks............................................. 147
7.2.1 The Large Sparse Capacitated Arc Routing Problem (LSCARP)................... 147
7.2.2 Time Capacitated Arc Routing Problems ........................................................ 149
7.3 Modified bounds for LSCARP................................................. 150
7.3.1 Lower bounds for the LSCARP......................................................................... 150
7.3.2 Depot based lower Bounds for TCARP ............................................................ 150
7.3.3 Bounds for TCARP based on the entire graph................................................. 152
7.3.4 Modification to LB2 cutset strategy.................................................................. 154
7.4 Computational Results for TLB2 ........................................... 155
7.4.1 Example of new cutset strategy......................................................................... 155
7.4.2 Simple Computational Example for TCARP................................................... 157
7.5 Comparison of lower bound and optimal results................... 161
7.5.1 Time based branch and bound procedure........................................................ 161
7.5.2 Computational experiments.............................................................................. 164
7.5.3 Implications of computational results ............................................................. 166
CHAPTER 8 : SOLUTION PROCEDURES FOR TCARP............... 167
8.1 Introduction............................................................................. 167
8.1.1 Solutions for TCARP......................................................................................... 167
vii
8.1.2 Algorithm requirements for the LSCARP ........................................................ 167
8.2 Heuristic approaches .............................................................. 168
8.2.1 CARP solution techniques................................................................................. 168
8.2.2 Route-First Cluster-Second approach .............................................................. 169
8.2.3 Cluster-First Route-Second Approaches .......................................................... 169
8.3 Route first, cluster second algorithm..................................... 170
8.3.1 Tour construction .............................................................................................. 170
8.3.2 Computational example for the RFCS algorithm............................................ 171
8.3.3 Evaluation of the RFCS algorithm .................................................................. 173
8.4 Tree based Approach to clustering for the LSCARP............ 173
8.4.1 Clustering on rural road networks................................................................... 173
8.4.2 The shortest path tree clustering algorithm..................................................... 174
8.4.3 Computational example of tree clustering algorithm...................................... 175
8.4.4 Evaluation of shortest path tree clustering approach ..................................... 177
8.5 Insertion heuristic for clustering for the LSCARP................ 178
8.5.1 Justification for using insertion procedure...................................................... 178
8.5.2 Operation of the algorithm ............................................................................... 179
8.5.3 Computational example .................................................................................... 180
8.5.4 Evaluation of Insertion heuristic ..................................................................... 182
8.6 Refinements to heuristic algorithms...................................... 182
8.6.1 Network simplification ..................................................................................... 182
8.6.2 Route refinement ............................................................................................... 184
CHAPTER 9 : DECISION SUPPORT FOR THE LSCARP ON IRISH
ROAD NETWORKS. .............................................................................. 185
9.1 Specific DSS Characteristics .................................................. 185
9.1.1 Network characteristics .................................................................................... 185
9.1.2 Management Requirements .............................................................................. 185
9.1.3 Address structure .............................................................................................. 185
9.2 ROMC approach...................................................................... 188
9.2.1 Principles of the ROMC approach.................................................................... 188
9.2.2 Representations ................................................................................................. 188
9.2.3 Operations ......................................................................................................... 190
9.2.4 Memory Aids...................................................................................................... 191
9.2.5 Controls.............................................................................................................. 191
9.3 TransCAD GIS ........................................................................ 192
viii
9.3.1 TransCAD Software characteristics................................................................. 192
9.3.2 GDK ................................................................................................................... 193
9.3.3 User Interface Features..................................................................................... 194
9.4 Initial Computational Experiments with routing algorithms
........................................................................................................ 195
9.4.1 Early testing of RFCS and Tree based approaches ......................................... 195
9.4.2 Assessment of initial testing ............................................................................. 195
9.5 Final Computational Experiments with routing algorithms 196
9.5.1 Data sets used.................................................................................................... 196
9.5.2 Case One ............................................................................................................ 197
9.5.3 Case One solutions ............................................................................................ 199
9.5.4 Case Two............................................................................................................ 200
9.5.5 Case Three ......................................................................................................... 202
9.6 Conclusion ............................................................................... 205
9.6.1 Decision Support for Routing ........................................................................... 205
9.6.2 Modelling issues ................................................................................................ 205
9.6.3 Further Research on Modelling Techniques .................................................... 206
9.6.4 Developments in SDSS ..................................................................................... 208
References...................................................................................... 209
APPENDIX A ...........................................................................................A1
A.1.1 Case One Solution 1 ...........................................................................................A1
A.1.2 : Case One Solution 2 .........................................................................................A2
A.1.3 : Case One Solution 3 .........................................................................................A3
A.1.4 : Case One Solution 4 .........................................................................................A4
A.1.5 : Case One Solution 5 .........................................................................................A5
A.1.6 : Case One Solution 6 .........................................................................................A6
A.1.7 : Case One Solution 7 .........................................................................................A7
A.1.8 : Case One Solution 8 .........................................................................................A8
A.1.9 Case One Solution 9 ...........................................................................................A9
A.1.10 : Case One Solution ........................................................................................A10
APPENDIX B ..........................................................................................A11
B.1.1 : Case Two Solution..........................................................................................A11
APPENDIX C ..........................................................................................A12
C.1.1 : Case Three Solution A12
ix
List of Tables
Table 1-1 : Decision Structure (McCosh and Scott Morton, 1978,Page 8) 6
Table 1-2 : Alter’s Taxonomy of DSS (Alter, 1980) 9
Table 1-3 : DSS design guidelines (Barbosa and Hirko, 1980) 10
Table 2-1: Features of GeoRoute (Georoute) 38
Table 3-1 : Data operations in GIS 45
Table 3-2 : Computerised Support for Decision-making (adapted from
Turban, (1995 Page 19) 53
Table 3-3 : DSS Generator Features 56
Table 3-4 : Contextual Information in SDSS (Keenan, 1998b) 59
Table 3-5 : Software integration techniques for building SDSS (Keenan,
1998c) 63
Table 4-1 : Main Characteristics of Routing Systems (Keenan, 1998a) 69
Table 4-2 : Constraints in Vehicle Routing Problems (adapted from Bodin
and Golden) (1981) 70
Table 4-3 : GPS applications in routing 74
Table 4-4 : Types of Location Data 79
Table 4-5 : Types of Path Data 80
Table 4-6 : Example of an agricultural routing SDSS 91
Table 4-7 : Support Requirements of Routing Problems 93
Table 5-1 : Classification in Vehicle Routing and Scheduling (Bodin and
Golden, 1981). 97
Table 5-2 : Summary of LB2 Algorithm 123
Table 6-1 : Heuristic performance (adapted from Assad and Golden(1995))
128
Table 6-2 : Turn Penalties (Bodin and Kursh, 1979) 132
Table 6-3 : Outline Code of the Tour Construction Algorithm: 140
Table 7-1 : The sets at each iteration using the original LB2 procedure156
Table 7-2 : The sets at each iteration using the modified cutset selection
157
Table 7-3 : TLB2 example 158
Table 7-4 : Summary of Time based bound 160
x
Table 7-5 : Computational results for 35 arc network 164
Table 8-1 : Pseudo code for RFCS approach 171
Table 8-2 : Infeasible route for RFCS procedure 171
Table 8-3 : Routes generated by RFCS algorithm 173
Table 8-4 : Pseudo code of tree clustering algorithm 175
Table 8-5 : Routes from Tree Clustering approach 176
Table 8-6 : Pseudo code for insertion algorithm 179
Table 8-7 : Insertion heuristic route 182
Table 9-1 : Examples of GIS operations to facilitate use of townland
structure 187
Table 9-2 : Summary of ROMC features 191
Table 9-3 : Initial computational results 195
Table 9-4 : Region data 197
Table 9-5 : Case One - total time summary 199
Table 9-6 : Case Two - total time summary 202
Table 9-7 : Case Three - total time summary 204
xi
List of Figures
Figure 1-1 : Components of a DSS (Sprague, 1980) 7
Figure 1-2 : DSS Technology Levels (Sprague 1980) 15
Figure 2-1 : Use of barriers to prevent routes crossing geographic features
24
Figure 2-2: Graphic display of ROVER software 29
Figure 2-3 : GeoRoute interface 37
Figure 3-1 : Use of spatial techniques to identify neighbouring regions 46
Figure 3-2 : Example of GIS use to identify street segments close to a
route (Keenan, 1998c) 47
Figure 3-3 : Building a SDSS by integrating models with GIS 57
Figure 4-1 : Area served from bus stop will include neighbouring streets.
76
Figure 4-2 : The total service area is mapped on to a limited number of
locations. 77
Figure 4-3 : Network constrained route avoiding passing within a certain
distance of a point location. 81
Figure 4-4 : Patrol area around irregular boundary of sensitive
installation 82
Figure 5-1 : The Königsberg bridge problem (Euler, 1736) 100
Figure 5-2 : Graph with four odd points (C,D,E,F) and addition of
redundant arcs to make all nodes even (Kwan, 1962) 101
Figure 5-3 : Introducing new nodes for each original arc (Pearn, Assad
and Golden, 1987) 105
Figure 5-4 : Illegal and legal subtours (Golden and Wong(1981)) 108
Figure 5-5 : The Original Graph 111
Figure 5-6 : The Transformed Graph 112
Figure 5-7 : Initial NDLB MCPM on the 9-arc example 113
Figure 5-8 : Optimal Matching of Hs 116
Figure 5-9 : The Cutset from LB1 118
Figure 5-10 : The second cut 120
xii
Figure 6-1 : Block Design for snow ploughing (Gendreau, Laporte et al.,
1997) 132
Figure 6-2 : Network enhancement by addition of penalties at junctions
(Roy and Rousseau, 1989) 133
Figure 6-3 : Postman paths on the 9-arc example 136
Figure 6-4 : Branch and Bound Sub-Problems 138
Figure 6-5 : Fathomed Sub-Problems and Optimal Solution, Sub-Problem
G 138
Figure 7-1 : Network with arc-node ratio of 2 143
Figure 7-2 : Network with high arc/node ratio 143
Figure 7-3 : Extract from Irish Rural road network 144
Figure 7-4 : Extract from Dublin City main road network 145
Figure 7-5 : Extract from New York City road network 146
Figure 7-6 : Multiple nodes in U′ at each iteration 154
Figure 7-7 : Original LB2 cutsets 155
Figure 7-8 : Selecting Nodes one at a time, in increasing order of
connectivity 156
Figure 7-9 : 35 arc network 163
Figure 7-10 : Maximum TLB2 cut for vehicle capacity of 35 165
Figure 8-1 : Simplified network for heuristic examples 172
Figure 8-2 : Treelike structure of rural road network 174
Figure 8-3 : Routes derived from clusters 177
Figure 8-4 : Seed Arcs for Insertion heuristic 180
Figure 8-5 : Routes 1&2 using insertion heuristic 181
Figure 8-6 : Multiple arcs in a single road section 183
Figure 8-7 : Complex Cul-de-sac road section 183
Figure 9-1 : Overlay of townland boundaries and road network 187
Figure 9-2 : Removal of irrelevant detail in interface and algorithmic
representations 188
Figure 9-3 : Representation of arc attributes in different colours and line
weights 189
Figure 9-4 : Only sections of road visited more than once are shown 190
xiii
Figure 9-5 : TransCAD interface with TransCAD arc routing data 193
Figure 9-6 : Case One - road network with population density 198
Figure 9-7 : Case One - reduced network 198
Figure 9-8 : Case Two - road network and reduced network 201
Figure 9-9 : Case Three - road network and reduced network 203
xiv
Glossary
Arc Routing Problem An OR/MS problem concerned with visiting a
sequence of arcs (edges) in a graph.
Cluster-First Route-Second Approach (CFRS) Routing problem
where the the allocation of arcs or nodes to vehicles is established prior to
the routing sequence.
Decision Support System (DSS) A computer-based system consisting
of an interface, a database and a problem processing system whose
purpose is to support decision-making activities.
DSS Generator - Computer software that provides tools and capabilities
that help a developer quickly and easily build a specific DSS.
Capacitated Arc Routing Problem An OR/MS problem concerned with
a vehicle of limited capacity visiting a sequence of arcs (edges) in a graph.
Chinese Postman Problem (CPP) An arc routing problem where the
objective is to visit each arc at least once.
Geographic Information System (GIS) A computer-based system for
the storage and processing of spatial information.
Global Positioning System (GPS) A satellite based system for
establishing the location of a point on the Earth's surface.
Large Sparse Capacitated Arc Routing Problem (LSCARP) The arc
routing problem on large sparse networks that is the specific problem
discussed in this dissertation.
Minimum Cost Perfect Matching (MCPM) The minimum set of non
adjacent edges which joins all of the nodes in a graph.
ROMC (Representation, Operations, Memory Aids, Control)
Design Approach A systematic user-oriented approach for developing
large-scale DSS.
xv
Route-First Cluster-Second Approach (RFCS) Routing problem
where the routing sequence is established prior to the allocation of arcs or
nodes to vehicles.
Rural Postman Problem (RPP) is an extension of the CPP where only
a subset of arcs (edges) from the network must be traversed.
Spatial Decision Support System (SDSS) Decision Support System
for processing spatial information based on GIS technology.
Time Capacitated Arc Routing Problem (TCARP) Arc routing
problem where the vehicles routes are limited to a maximum time.
TransCAD The GIS software designed for working with transportation
problems that is used for the prototype systems discussed in this
disseratation.
Spatial Decision Support Systems for Large Arc Routing Problems
1
Chapter 1 : Introduction
1.1 Introduction
1.1.1 Preface
The arrival of digital computer technology has led to rapid and continuing
change in many aspects of human activity, especially that of business. As
with other technological innovations, the use of Information Technology
(IT) has allowed business to perform its operations more effectively.
However, the greatest contribution of the introduction of IT has been in
the improved planning of business operations. To allow this improvement
to take place, a variety of techniques have been developed to better
exploit the capacity of IT. In some cases, these computer-based
approaches have been quite different from the manual methods
previously employed. As research continues into how best to approach
problem solving using IT, new and distinctive computer based techniques
have been introduced. This research led to investigation into the
reconciliation of the problem representations used by the human decision-
makers with those used by the digital computer. This dissertation reports
one such piece of research. It looks at how computer technology and
computerised techniques can be synthesised with human problem solving
abilities to provide a superior approach to the management of a specific
problem, namely that of Decision Support Systems for large arc routing
problems.
1.1.2 Information Technology
Digital computer technology was introduced in the 1940s, its
development having been greatly accelerated by the needs of the Second
World War. Early computers were extremely limited in power by modern
standards, but they did enormously increase the rate at which
Spatial Decision Support Systems for Large Arc Routing Problems
2
mathematical computation could take place, hence the name computer.
These machines were used initially for quantitative applications where
fast computation was essential, for instance code-breaking or ballistic
calculations. Early civilian applications were also essentially numerical
in nature and included payroll processing and census enumeration. The
introduction of these machines, capable of fast computation, led to the
development of techniques that exploited this ability. In particular
computer technology enabled the emergence of the field known as
Operational Research in the United Kingdom and Operations Research or
Management Science (OR/MS) in the United States. The field made a
great contribution by developing solution techniques that exploited the
rapidly growing power of computers. This allowed solutions to be derived
in a reasonable time, in circumstances where manual computation was
infeasible.
As computer technology developed, computational performance increased
exponentially. This was accompanied by enormous increases in the
storage capacity of computers. These developments meant that computers
became convenient ways to store and retrieve large amounts of data. The
potential usefulness of IT was greatly increased when this enhanced
storage capacity was allied to the rapidly accelerating computational
power of the machines. The full exploitation of these developments in
computer hardware required the design of appropriate computer
languages and new computer science techniques.
As time progressed, the peripheral devices associated with computers also
underwent rapid development. New input and output devices allowed
greater ease of use, for example, terminals, printers, plotters, keyboards,
etc. These developments became possible because the rapid increase in
computer speed allowed more resources be devoted to managing
input/output and achieved greater user-friendliness. This trend
culminated with the introduction of the personal computer (PC), which
offered colour screens and new input devices such as the mouse.
Spatial Decision Support Systems for Large Arc Routing Problems
3
Computer technology has become very widely used and familiar to
business managers and, with the use of appropriate software, and now
plays an important role in management.
1.1.3 Growth of Information Systems (IS)
The rapid advancement in IT hardware capabilities provided the
potential for new IT applications, however this potential could be
released only by the use of appropriate software. Early mainstream
business applications of computer technology, such as payroll processing,
were known as Transaction Processing Systems (TPS). These applications
were characterised by the automation of repetitive clerical tasks and
exploited the computational abilities of the computer. These early
applications allowed cost savings in the clerical departments in which
they were implemented, but had no direct implications for management
planning in other areas. As computer power increased, it became
apparent that some information useful to management could be produced.
For example, summary reports of sales or purchasing could be derived
from the computerisation of these functions. This led to the introduction
of Management Information Systems (MIS) which provided routine
reports in standard formats for management use. However, not all
functional areas utilised the type of routine procedures embodied in these
systems. MIS reports could only be altered by programmers and were
therefore of little use for ad-hoc or unexpected decisions where
information needs could not be predicted far in advance.
In parallel with these developments, the growing computational power of
IT was also being exploited for mathematical modelling applications.
These techniques address problems such as the optimal allocation of
resources in production planning (linear programming), and inventory
management (economic order quantity). These modelling applications
were capable of addressing management problems, but could only be used
easily by specialists. Early computer technology required punched card
data input and had very limited facilities for presenting output, making it
Spatial Decision Support Systems for Large Arc Routing Problems
4
extremely difficult for managers to operate these modelling applications
themselves. The limitations of the technology not only made computer
use inconvenient, but also imposed restricted computer-orientated
problem representations on the process. These problem representations
differed greatly from that familiar to mainstream managers. The impact
of IT on business has been hindered by this division between managers
with an understanding of the business problems and the technically
literate specialists who could best employ computer technology.
With the increase in IT capabilities, researchers began to propose a more
comprehensive use of the technology to support management needs. By
the early 1970s, obvious progress had been made in the application of
technology and in the use of problem solving techniques. A variety of in-
formation systems were proposed to meet the diverse needs of users
(Mentzas, 1994). One of the most important types of system introduced
was the Decision Support System (DSS), which focussed on better man-
agement performance rather than on the replacement of clerical labour.
1.2 Decision Support Systems
1.2.1 Decision Structure
Much of the early work on the concept of a DSS took place at the
Massachusetts Institute of Technology (Gorry and Scott-Morton, 1971;
Little, 1971). Gorry and Scott-Morton built on a management framework
introduced by Simon (1977), that identified decision structure as critical
to the process of decision-making. Gorry and Scott-Morton saw the
identification of the degree of decision structure as the starting point for
the design of an appropriate information system. At one end of a
spectrum were structured decisions. These were repetitive and routine
and so were sometimes called programmed decisions. For structured
decisions, a definite procedure exists to solve the problem and that can be
applied routinely to any new decision. By contrast, unstructured or
unprogrammed decisions were novel and had no clear-cut procedure to
Spatial Decision Support Systems for Large Arc Routing Problems
5
solve them. Instead, a general adaptive strategy was used to solve such
problems. While many structured activities exist at the operational level
in an organisation, they may also be found at the tactical and strategic
level of decision-making. Unstructured decisions may also be found at all
levels of the organisation, although on balance long term strategic
decisions are more likely to be unstructured than those at lower levels of
the organisation
Much of the early contribution by the use of IT was made in structured
operational problems. However, many important problems are semi-
structured and this class of problem requires a system that provides for
easy interaction between man and machine. Examples include DSS
applications that aim to assist (support) the solving of problems using
both human and computer techniques.
Discussion of structure is complicated by the fact that many problems are
structured in principle, but are addressed in practice using management
judgement. A major reason for manual intervention in modelling occurs
where the problem is so complex as to be computationally infeasible.
Many problems are inherently structured, but are so complex that with-
out IT only simplified problem representations could be used. The advent
of new modelling techniques, coupled with the rapidly increasing power of
computers, meant that structure can be directly modelled for an
increasing range of problems. One example is linear programming, which
can be used for decision-making in fields that previously required great
management experience. Likewise, inventory models can model the struc-
ture inherent in reordering decisions, which previously required
management expertise (McCosh and Scott Morton, 1978,Page 10)
In addition to introducing three levels of decision structure (Table 1-1),
Simon proposed a structure for problem solving with three main phases.
! Intelligence : searching the environment
! Design : the development of possible courses of action
! Choice : selecting a particular course of action
Spatial Decision Support Systems for Large Arc Routing Problems
6
In a given problem, some or all of these phases may be structured.
Problems may have one relatively structured phase with the other phases
much less structured. For instance, the environmental conditions may be
known, but there may not be any clarity in the courses of action available
to solve the problem. The choice phase may be totally structured, for
example if monetary value is used, but the previous stages may be more
ambiguous. It is desirable that a DSS support all phases of decision-
making.
1.2.2 Definitions of DSS
DSS became feasible in the 1970s following developments in computer
technology. As input/output devices improved, machines were introduced
which used a keyboard for input and which employed graphical terminals
capable of displaying computer output. This meant that a manager could
Table 1-1 : Decision Structure
(McCosh and Scott Morton, 1978,Page 8)
Operationalcontrol
Managementcontrol
StrategicPlanning
Structured Order entry
Accounts
receivable
Inventory control
Variance analysis
Short-term
forecasting
Tanker fleet structure
Factory location
Warehouse location
Unstructured Short term cash
management
Job-shop
scheduling
Advertising
selection
Budget
preparation
Sales planning
Production
planning
Price setting
New product planning
Mergers and
acquisitions
Spatial Decision Support Systems for Large Arc Routing Problems
7
interact directly with the computer. The development of improved
computer storage and retrieval technology meant that managers could be
given access through such a terminal to large amounts of data. Increased
user control meant that managers could extract only that subset of data
of interest at any given point in time. When DSS were first proposed,
computer performance had advanced to the point where sophisticated
models could be solved in real time. The early definitions of DSS reflected
these aspects of technology in that they tended to define DSS as flexible
systems combining database and model components aimed at less
structured decisions (Sprague, 1980; Sprague and Carlson, 1982). These
modelling and database components are under the control of the user
through an interface or dialogue system (Figure 1-1).
Database Model base
DBMS MBMS
DGMS
The DSS
Task Environment
User
Figure 1-1 : Components of a DSS (Sprague, 1980)
Spatial Decision Support Systems for Large Arc Routing Problems
8
The dialogue component of this system itself consists of three
components.
! The Database Management System (DBMS) allows easy
access to the data and the ability to alter and reorganise the
contents of the database.
! The Modelbase Management System (MBMS) allows the
user access to the models and provides the ability to intervene
in the modelling process.
! The Dialogue Generation Management System allows the
user control the presentation of information and the interaction
with the system.
Other authors have proposed broadly similar generic descriptions of DSS,
for example Bonczek, Holsapple and Whinston (1981). They identified a
DSS as comprising three parts, a definition that is broadly similar to that
of Sprague above.
! The Language System is the means by which the user
interacts with the DSS and is analogous to the user interface.
! The Knowledge System provides system access to domain
relevant information other than that in the mind of the user.
This is generally a DBMS but these authors use a deliberately
broad definition to include other knowledge representations
such as the rule base in an expert system.
! The Problem-processing System can interpret the knowledge
used in the knowledge system. This would include traditional
models but also techniques such as inference in an expert
system.
Spatial Decision Support Systems for Large Arc Routing Problems
9
The modelling (problem-processing) component of DSS was able to exploit
the many techniques developed in the OR/MS and financial fields.
However, the nature of the definition of DSS allows varying degrees of
combination of database and modelling techniques. There is a lack of
agreement on what constitutes a DSS; some systems regarded as DSS by
one author might be excluded by another. Alter (1980) produced an
influential taxonomy of DSS that proposed seven subdivisions ranging
from data driven systems to model driven systems (Table 1-2). While
most DSS definitions would include categories at the centre of Alter’s
classification, many would exclude those at its extremes. For example,
File Drawer systems do not contain any models and therefore would be
excluded by Sprague’s definition. Similarly Suggestion systems could be
said to support decisions rather than make them and so could be excluded
from a definition of DSS.
The argument of what constitutes a DSS is an ongoing one. Keen (1986)
noted that there was a tendency for any system that contributes to
decision-making to be called a DSS. Stabell (1986) identified four schools
Table 1-2 : Alter’s Taxonomy of DSS (Alter, 1980)
System type Example
File Drawer Query Systems (MIS)
Airline Reservation
Data Analysis Database Management Systems
Analysis Information Systems Spreadsheets
Accounting Models Monetary Simulations
Representational Models Simulation
Optimisation Models Linear Programming
Suggestion Models AI Expert Systems
Spatial Decision Support Systems for Large Arc Routing Problems
10
of DSS thought; these include the influences of modelling and decision
analysis. Silver (1991) proposed a classification of DSS that emphasised
the degree of guidance provided. Against this background of disagreement
on the definition of DSS we see many “decision support systems” that are
essentially OR/MS models with minimal database or interface features.
Other systems are little more than DBMS with some retrieval facilities
and little modelling capability. A recent comprehensive review of DSS
applications identified a large number of articles with the descriptor
“decision support systems” (Eom, Lee, Kim and Somarajan, 1998).
However, less than one-fifth of these articles were actually included in
the survey, as most systems did not meet their definition of a DSS.
In looking at the potential for incorporating OR/MS techniques in a DSS,
Barbosa and Hirko (1980) put forward a succinct set of guidelines for the
design of OR/MS based DSS (Table 1-3). Many systems described as being
DSS by their authors do not meet the requirements of these guidelines.
Table 1-3 : DSS design guidelines (Barbosa and Hirko, 1980)
DSSfeature
Design Guideline
Interface No unnecessary distractions, user should not have to laboriously enter control
parameters
Control parameters should be expressed in terms with which the user is
familiar
Control System should support a spectrum of control, e.g. manual and fully automated
operation
User should be able to provide information as required
Flexibility Algorithmic and manual operations should be interchangeable
Feedback User should be aware of the state of the process at all times
User feedback on usability should be used to improve the system
Spatial Decision Support Systems for Large Arc Routing Problems
11
1.2.3 The DSS field
In the twenty-five years since the concept of DSS was first introduced, the
importance of DSS research has been widely recognised. For example, one
survey showed that more than one-third of IS researchers were working
in the DSS area (Teng and Galletta, 1990). The same survey found that
DSS was the topic that the most researchers felt was deserving of greater
research attention, with almost a quarter of researchers surveyed holding
this view. Another recent survey (Lee, Gosain and Im, 1999) suggested
that DSS was frequently discussed in academic journals, but was less
often the subject of attention in practitioner publications.
Several surveys have attempted to identify the nature of the DSS field.
Reviews of the field tend to indicate the use of DSS at operational levels,
while seeing the need for further development in applications directed at
higher levels of management. In one of the few contributions by an Irish
based author, Er (1988) notes the widespread use of DSS for short term
decisions and the less frequent use of DSS for longer term decision-
making at higher levels in the organisation. In a bibliographic survey of
DSS applications over a sixteen-year period (Eom and Lee, 1990), OR/MS
applications were to the fore, primarily in the operations management
and routing areas. Later work by the same authors (Eom, Lee, Somarajan
and Kim, 1997; Eom, Lee et al., 1998) found that OR/MS models were
still dominant, although other fields such as artificial intelligence were
making an impact. It is notable that mainstream OR/MS publications
represented four of the top five journals identified as being important for
DSS applications (Interfaces, European Journal of Operational Research,
Operations Research, and Computers and Operations Research).
Other work has reflected this distribution of DSS output between OR/MS
and IS journals (Abraham and Wankel, 1995). However, it is probable
that many applications in these publications, although described as DSS,
were discarded by this survey because they were not comprehensive
systems. Nevertheless, the close relationship between OR/MS and DSS is
Spatial Decision Support Systems for Large Arc Routing Problems
12
clear. In the past some in the OR/MS community have seen DSS as
merely a subset of OR/MS, for instance see the paper by Naylor (1982)
and the counter argument by Watson and Hill (1983). Other authors have
identified related classes of system, such as the decision insight systems
reviewed by Golden, Hevner and Power (1986). Many of these systems
would be regarded as DSS by other authors.
DSS therefore exploits the vast range of modelling approaches found in
OR/MS. However, in a DSS, the model is only one part of a system and
other disciplines are of relevance. One study (Eom, Lee and Kim, 1993)
identified psychology, artificial intelligence, OR/MS, computer science,
organisational sciences and functional management theory as being the
reference disciplines for DSS. This work looked at author cocitation as a
means of classifying DSS research. The authors found that distinct
groups of researchers existed, notable fields being those of group DSS,
multi-criteria decision-making, routing, database management and other
OR/MS applications. They also noted that other themes in DSS research,
such as model management, had little influence on the DSS applications
reviewed.
In the context of this dissertation, it is notable that transportation
applications are a recognised area of application of DSS. Many survey
and review articles of DSS mention the importance of and give examples
of this type of application. DSS textbooks frequently use transportation
applications as examples, for instance Sauder and Westerman (1993).
Sprague (1980) discusses the use of the Geodata Analysis and Display
System (GADS) which was used to build routing DSS. GADS employed
map display as an important component of decision support and can be
regarded as a form of Spatial DSS, this class of DSS is discussed further
in Section 3.4. In his review, Er (1988) notes the importance of
transportation applications. Analyses of the literature (Eom and Lee,
1990; Eom, Lee et al., 1993) have identified routing as the most
important area of DSS application in business. The authors of these
Spatial Decision Support Systems for Large Arc Routing Problems
13
reviews suggest that the relative importance of routing DSS reflects the
nature of routing problems. They suggest that these are less suitable for
the expert systems approach used in other business applications and
indicate the importance of skilled user intervention for problem solving in
the routing field.
1.2.4 Contributions of DSS to decision-making
A DSS may contribute in various ways to improving decision-making
(Forgionne, 1999). DSS use may improve the process of making a
decision. This might occur where by reducing the time taken to make the
decision. The increasing speed of modern computers makes this an
obvious advantage. DSS use may improve the outcomes of the decision
process by leading to better decisions being made. This may result from
more information being brought to bear on the problem, with a
consequent improvement in the intelligence phase of the decision. The
language system (interface) used in the DSS may facilitate easier
assimilation of the information presented, for example if graphics are
used. This dissertation focuses on mapping information as a form of
graphic representation of particular importance arc routing problems.
The information storage and retrieval components of the DSS may
contribute to this improvement. DSS use may lead to consideration of
more courses of action, leading to improvements in the choice phase of the
decision. This might result from the reduced time needed to model each
alternative. Modern computer technology allows extensive what-if
analysis take place in a relatively short time. DSS may lead to an
improved assessment of the value of those courses of action examined,
this might arise through the computer modelling of more complex system
models rather than alternative overly simplified manual techniques.
DSS use can also lead to long term benefits. DSS is recognised as
providing user learning benefits, although not all systems fully exploit
this potential gain (Santana, 1995). DSS use can also assist organisations
Spatial Decision Support Systems for Large Arc Routing Problems
14
by allowing the capture of additional information relating to the decision-
making process. If a common DSS is used by different decision-makers, it
can help spread common practice throughout the organisation and assist
in the spread of knowledge. However, this type of benefit may be difficult
to realise where individual decision-makers have greatly different
approaches to decision-making.
1.3 DSS Technology
1.3.1 Building DSS
The emergence of DSS as a concept was greatly influenced by changes in
technology. In the 1970s, computer technology had advanced to the point
where it was usable by non-IT professionals and where costs had declined
to the point where greater use of the technology had become cost effective.
This trend was enhanced by the introduction of PCs at the end of the
decade; this technology was given respectability with the introduction of
the IBM PC in 1981. There has been rapid progress in technology since
then, to the point where modern microcomputers are some five hundred
times as powerful as the original IBM PC. Modern machines are provided
with two hundred times as much electronic memory and perhaps one
thousand times as much disk storage space. Yet, the real price of modern
machines is about one quarter of the original IBM PC. Against this
background the cost of the technology has become less of a limitation and
the growth of DSS is now determined by the ability of designers to
propose systems of genuine benefit to managers.
When the concept of DSS was first introduced, it had distinctive
characteristics differentiating it from earlier types of information system.
These earlier systems, such as MIS, were built by teams of IT
professionals, in departments removed from the mainstream
management in the organisation. Large MIS projects used a formal
process, such as the Systems Development Lifecycle (SDLC). These
systems were developed from complex specifications and were not easily
Spatial Decision Support Systems for Large Arc Routing Problems
15
changed after completion. DSS, on the other hand, required a much closer
association with the decision-maker. Consequently, there was a greater
need for user input into the system design. The definition of DSS also
required that the user had good control over the operation of the system,
which raised the possibility of the user being able to alter the system.
Indeed, with the greater ease of use brought about by improvements in
IT, in some cases the user could develop the DSS his or herself.
1.3.2 DSS generator approach
An early comprehensive framework for the development of DSS was
provided by Sprague (1980; Sprague and Carlson, 1982). In Sprague's
framework, a DSS may be built from individual software components
called tools that were then combined to form a DSS. These could include
programming languages, programming libraries and small specialised
applications. At a higher level in Sprague's framework are DSS
generators, from which a specific DSS can be quickly built. Generators
might be built from lower level tools (Figure 1-2). Sprague envisioned
DSSGenerator
Specific Applications
DSS Tools
DSSGenerator
Figure 1-2 : DSS Technology Levels (Sprague 1980)
Spatial Decision Support Systems for Large Arc Routing Problems
16
that different specific DSS applications would require different
combinations of the generator and tools. Sprague used GADS (Grace,
1977), as an example of a DSS generator.
In building DSS, specific generators were designed for certain classes of
problem. In other situations general-purpose software such as
spreadsheets or DBMS packages have been regarded as generators. In
modern DBMS and spreadsheet software, the use of macro and
programming languages facilitates the creation of specific applications.
Different generators had varying strengths and weaknesses in terms of
their provision of the essential components of a DSS: an interface, a
database, and models. In the case of a spreadsheet, modelling is the basic
function of the software; various interface features such as graphs are
provided, but the database organisation is simplistic. DBMS software,
such as Access or Paradox, has good database support, provision for
interface design using forms, report and charts, but almost no modelling
support. In this case, the modelling support has to be added to the specific
DSS built from such a system. In Sprague’s framework, the DSS builder
could make use of tools, which provide some, lower level, data processing.
In software design these might include operations such as sorting or
searching, which although important algorithms in their own right, are
not of direct interest to the decision maker. The design focus is on the
outcomes, the decision, and not on the technological and modelling inputs
to the system.
Sprague’s framework has been widely accepted by DSS researchers and is
frequently used as a basis for discussion of DSS in textbooks (Mallach,
1994; Sauter, 1997; Marakas, 1998; Turban and Aronson, 1998).
Nevertheless not all authors find this framework useful, for example
Holsapple and Whinston (1996) regard the distinction between tools and
a DSS generator as unclear. However, we believe that this framework is
useful for the purposes of this dissertation and so will form the basis of
the discussion on building arc routing DSS in Chapter 9.
Spatial Decision Support Systems for Large Arc Routing Problems
17
1.3.3 ROMC approach
Sprague and Carlson (1982) proposed a framework for identifying the
features needed in a DSS. This is the Representations, Operations,
Memory Aids and Controls (ROMC) approach.
! Representations allow the decision-maker to better visualise
the problem and facilitate understanding of the solution
process.
! Operations are needed to reach a solution, these operations
can be identified at a logical level, in addition to the low level
mathematical and DP techniques employed.
! Memory Aids improve the productivity of the user, without
necessarily providing any conceptual assistance.
! Controls are the levers used by the user to achieve a solution,
the concept of user control of the solution process is inherent in
DSS.
Frameworks such the ROMC approach provide an initial emphasis on the
problem, rather than the techniques used to address that problem. We
believe that this approach is a useful one as it should provide a more
balanced system than many OR/MS systems that start with the model as
the basis for system design. In Chapter 9 we discuss this approach in
relation to our specific problem introduced in Section 7.1
1.3.4 Current DSS technology
The technological environment has changed greatly since Sprague’s
original proposal of this framework (Power and Kaparthi, 1998). Recent
developments include the widespread use of graphical user interfaces, the
growth of network connectivity, the Internet, and the development of
common standards for software. However most of these changes enhance
Spatial Decision Support Systems for Large Arc Routing Problems
18
the relevance of Sprague’s framework as they facilitate system
integration and ease of alteration. New programming tools allow the easy
development of industry standard interfaces that are widely acceptable to
users. The developments include advances in database software and the
use of common standards that allow a wider range of programs
interchange data. There has been considerable development in the area of
client-server systems, where distributed systems, such as DSS, can make
use of centralised databases. Where databases are used, Structured
Query Language (SQL) has become a standard means of sending
database commands. Open database connectivity (ODBC) is an attempt
to provide a standard method of communication between programs and
databases. As most database software and programming tools support
ODBC, it becomes possible to make use of a wide range of data from a
variety of alternative software sources. This approach provides an
effective means of exchanging data between a DSS and other programs.
For example, a DSS might be built by combining separate modelling
programs with access to a database. Given the growth in client-server
systems, these data exchange technologies will have an important role to
play in the development of DSS.
In an assessment of the contribution of these new software developments,
it is important to note the need for a DSS to be a comprehensive system.
A variety of software tools may contribute to decision-making, but may
not constitute a system. A DSS is a system and not a collection of distinct
components. Consequently, the various system features must be closely
integrated. For a comprehensive system to exist, data exchange alone
may not be sufficient. If a DSS generator is to interact with other
software, it will need to interact directly with other program components.
As a DSS generator is part of a larger system, its interaction features will
reflect the trends in software design generally. Therefore, any
developments that facilitate software integration greatly assist in the
building of DSS.
Spatial Decision Support Systems for Large Arc Routing Problems
19
1.4 The nature of this research in DSS
1.4.1 The current state of DSS
Since its introduction, DSS have been an important part of the IS field,
which remains an important one to researchers (Keen, 1998). There is
substantial, but by no means universal, agreement among DSS
researchers as to the definition of DSS. For the purposes of this
dissertation, a widely accepted definition of DSS is used. This defines a
DSS as complete system containing a database, models and a user
interface which allows the user effectively control the system. In a
broader world, where actual applications are being developed, the term
DSS might be used for many systems that do not fully meet this
definition. One important theme of this work is that decision support can
only be maximised by such a comprehensive system. Our objective is to
achieve a balanced examination of the role of the different components of
DSS in the context of a particular problem, that of arc routing DSS.
1.4.2 The structure of this dissertation
This dissertation focuses on a well-established area of DSS application,
that of routing. This area of DSS is discussed in the following chapter. It
goes on to argue for a general enhancement of routing DSS by the
incorporation of technology drawn from Geographic Information Systems
(GIS); the GIS field is reviewed in Chapter 3. Within the general field of
routing DSS, this research concentrates on arc routing problems, a sub
field that has not been extensively researched. The arc routing field is
introduced in Chapter 5 and Chapter 6 reviews the solution procedures
used for these problems.
As decision support is ultimately related to a specific problem, this work
expressly examines arc routing for large sparse networks. Irish rural road
networks are of this type and postal delivery in rural areas represents a
typical problem in this class. Very little work has previously taken place
on this specific sub-problem, which is introduced in Chapter 7. This
Spatial Decision Support Systems for Large Arc Routing Problems
20
chapter develops modified lower bounds for time base arc routing
problems. New heuristics for this specific problem are proposed in
Chapter 8. This provides an examination of the algorithmic requirements
of arc routing for large sparse networks. Chapter 4 proposes a new
framework for the integration of routing techniques and GIS. The final
chapter completes the work by examining the implications of these
general principles on the design of a DSS for routing for large sparse
networks.
Spatial Decision Support Systems for Large Arc Routing Problems
21
Chapter 2 : Decision support for routing
problems
2.1 Routing DSS
2.1.1 The nature of routing problems
The problem of how to collect from or deliver to multiple locations, the so-
called Routing Problem, is one of most important areas of investigation in
OR/MS. Routing problems require a set of locations to be visited in
sequence within some overall objective such as the minimisation of time
or distance. A wide range of constraints may exist which affect this
problem and which make the identification of appropriate routes more
difficult. Work in this area has been taking place for over forty years
since the original formulation of the problem (Dantzig and Ramser,
1959), and a variety of routing problems have been identified. Much
useful work in OR/MS took place in the 1970's, as the computers of the
day became powerful enough to solve an increasing range of problems
(Assad and Golden, 1982). A bibliographic review (Laporte and Osman,
1995) identified five hundred articles representing relatively important
contributions to routing related problems.
Considerable work has taken place in developing solution procedures for
routing problems (see Chapter 5). As computer technology has advanced,
it has become possible to obtain automated solutions to solve ever more
complex problems. As with other OR/MS applications, computerised
routing techniques have increasingly been integrated with other forms of
computer technology, especially databases and graphic interfaces. These
developments have led to modern routing DSS discussed in subsequent
sections.
Spatial Decision Support Systems for Large Arc Routing Problems
22
2.1.2 Suitability of routing problems for DSS
Routing problems are in many ways a typical DSS application. Routing is
of considerable economic value, justifying efforts to support decision-
making in this field. The nature of routing problems is well understood,
so the intelligence phase of the problem is well structured. Many of the
decision variables associated with routing are quantitative in nature and
therefore suitable for modelling by OR/MS techniques. Variables of
interest include; distances travelled, time of journey, volumes carried in
relation to vehicle capacity. Modelling techniques, such as those outlined
in Chapter 5 and Chapter 6, are capable of greatly assisting decision-
making. If the model includes all the variables of interest, then the
solution phase of the problem might be regarded as a structured one.
However, the DSS approach becomes appropriate when other factors are
also relevant. Customer service issues may not be easily incorporated in a
mathematical model, but may be vital to the decision-maker. Some
degree of user intervention is necessary to allow these non-quantitative
factors be considered. While the numerical factors may be assessed using
modelling techniques, the current state of algorithms and technology does
not allow for completely optimal solutions to most practical problems. In
this context, user intervention can contribute to the achievement of
quantitative goals as well as those not included in the modelling process.
The criteria for assessing routing decisions, the choice phase, are largely
straightforward. However, it is difficult to synthesise the various criteria
by which the solution may be measured. How can customer preference be
sacrificed for a shorter route? Are balanced routes strictly required or is
there some scope for small variations in each driver’s workload? These
trade-offs are difficult to incorporate in the models and are usually made
by the user using their judgement and experience.
Consequently, routing problems offer a good example of semi-structured
decision-making, where a DSS may make a superior contribution to a
purely mechanistic algorithmic approach. Such a DSS will contain models
Spatial Decision Support Systems for Large Arc Routing Problems
23
drawn from those identified in later Chapters. The database component
must provide the data to be used by these models but must also provide
the user with the means to direct the solution process towards non-
quantitative goals. The interface must represent both the input data and
the solution outcomes, in a way that allows the decision-maker interact
with the system.
2.1.3 Early routing Software
Typical early routing software was based on the use of location co-
ordinates and straight-line distance was used as a surrogate for actual
travel distance. This simplified the problem and reduced the data
requirements within the limitations of the computers of the period. This
abstraction of the problem assumed that selection of an appropriate path
was a trivial exercise. In practice, of course, the actual route is
constrained by the need to use suitable roads. Generally, the co-ordinate
distance was multiplied by a standard factor, between 1.1 and 1.25, to
allow for the fact that the actual road distance was slightly longer. Some
of these issues are discussed in Stocx and Tilanus (1991). Where there
were major natural features, such as lakes, rivers, mountains or bays, the
road distance might be much longer. This was taken into account by the
use of barriers, which meant that the straight-line distance was
calculated around the edge of the barrier. In the Irish context, these
barrier files would be critical in the long sea inlets found around the
coast. For example, barrier files might be used at Carlingford Lough or
the Shannon Estuary. Barrier files are an example of how the
representation of additional information in the system can make the
derived routes more realistic (Figure 2-1).
While the straight-line distance approach is unsatisfactory in some cases,
it is adequate for many classes of problem. In particular, problems where
widely dispersed points are to be visited can be approached using this
type of technique. Consequently, many successful early applications
Spatial Decision Support Systems for Large Arc Routing Problems
24
concerned supplies delivery to a variety of towns on a regional scale,
where the detail of street networks was not very important.
The true distance approach, using distances calculated from the road
network, reduced these problems. This requires that the road network be
digitised in some way and that the distances between points be
accurately calculated. The use of the true distance approach has become
an almost universal feature of routing DSS design. However, this
approach requires increased data and a consequent increase in the
sophistication of the software used for the organisation of that data. In
practice, routing problems are frequently constrained by time rather than
by distance travelled. Estimates of the time taken for a route may be
derived by using different speeds on different sections of the road
network. The incorporation of this type of additional path data increased
the usefulness of the problem formulation at the cost of making the
software to solve it more complex.
A good example of early routing software was VSPX (Vehicle Scheduling
Program Extended), introduced in 1971. This software employed the
Clarke-Wright savings algorithm (Clarke and Wright, 1964) and ran in
Figure 2-1 : Use of barriers to prevent routes crossing geographic
features
Spatial Decision Support Systems for Large Arc Routing Problems
25
batch mode on IBM mainframe computers. This software was extremely
inflexible and it could not be regarded as a DSS. However, it did allow the
modelling of substantial routing problems and the generation of solutions
to these problems which were relatively efficient in terms of distance
travelled. Early reviews of the field found few users of computerised
techniques, although the potential of these techniques was recognised
(Mole, 1979; Sussams, 1984). A Dutch study examined the suitability of
routing software for milk collection and placed particular emphasis in the
need for flexible, user friendly and interactive and cheap software (Bocxe
and Tilanus, 1985). The study identified limitations with all of the
packages studied, but found that almost all of the packages offered some
improvement over VSPX. This is hardly surprising given the relatively
user unfriendly and inflexible nature of VSPX, and it indicates the
direction in which routing software was developing.
Fisher (1995) suggests that there are no successful examples of this early
generation of routing software. While the inflexibility of the technology
made successful application difficult, this assertion is not entirely
justified. Successful early applications in the Irish context include a
Franz Edelman prize winning application (Harrison, 1979) and a later
paper describing a distribution optimisation project (Harrison and Wills,
1983). What is clear is that this type of early software could only be easily
used by specialists. This meant that it could only be employed at a
remove from the relevant managers in the organisation. This reduced the
effectiveness of routing projects, as managerial concerns might not be
fully communicated to the OR/MS specialists involved in the solution
process. Even if the specialists did have a good understanding of the
problem, the operation of the software did not closely reflect traditional
practice. Consequently, managers had difficulty understanding, and
therefore accepting, the output produced.
Spatial Decision Support Systems for Large Arc Routing Problems
26
2.2 Traditional Vehicle Routing DSS
2.2.1 Requirement for user interaction
For a system to be considered a DSS, and not just a modelling package, it
must be possible for the user to interact with the software (Gorry and
Scott-Morton, 1971). Therefore, the software must reflect the users’ view
of the problem to some extent. When considering routing problems using
non-computerised methods, much reliance is placed on the spatial layout
of the problem. Routes will usually be devised using paper maps,
typically moving from one point to another, which appear to be visually
close. Such manually designed routes will generally form compact blocks
when completed. These compact blocks will generally seem reasonable to
all those involved, including the customers and the vehicle drivers.
Mathematical techniques will obviously tend to produce routes that link
neighbouring points, for example those produced by a travelling salesman
algorithm. However, the routing heuristics commonly used will
sometimes produce routes that look quite bizarre in shape. For a small
number of problems that are tightly constrained, by factors other than
the location of the points to be visited, unusually shaped routes may in
fact be optimal. However, for most problems, skilled manual alteration of
the shape of a route in a DSS can improve the routes generated by the
heuristics that are commonly used. These alterations will improve the
spatial organisation of the route. These changes will also tend to improve
the acceptability of the route to customers and staff. User intervention is
often the only way of ensuring that the routes meet “soft” constraints that
are not part of the mathematical formulation. Experienced schedulers
have considerable knowledge of local conditions and of the relative
importance of the various constraints. Human intervention in the routing
process can substantially improve the quality of the routes produced. For
these reasons there has been increasing recognition that the available
algorithmic techniques can most effectively be used as part of a DSS.
Spatial Decision Support Systems for Large Arc Routing Problems
27
2.2.2 The use of graphics in routing systems
With the greater availability of graphics terminals and PCs, there was
increasing use of graphics to represent routes on screen. By the early
1980's there was increasing recognition of the need to combine OR/MS
algorithms with appropriate graphics (Barbosa and Hirko, 1980). The
latter paper also makes use of the routing related GADS example (Grace,
1977). In a mid-decade review of the vehicle routing field, Bott and Ballou
(1986) argued for the use of graphically based interactive techniques
combined with appropriate algorithms. However, a review of PC based
routing software (Golden, Bodin and Goodwin, 1986), published in the
same year, found that only one-third of the packages reviewed had
extensive graphic facilities. The role of visual interfaces is widely
recognised as a critical issue in the continuing development of OR/MS
generally (Jones, 1994b; Jones, 1999) and routing in particular (Bodin
and Levy, 1994).
One constraining factor on the development of routing software in the
1980's was the need for on-line access to increasing amounts of
information. For instance, in a survey of UK companies, Sussams (1984)
identified a need for order processing systems to be computerised in order
to facilitate computer based routing.
Consequently, routing systems increasingly attempted to incorporate
more of the information that a scheduler might require while evaluating
a route. This trend reflected recognition of the need to include a variety of
information sources relevant to a routing decision. Most traditional
routing problems use information that is particular to the problem.
Examples of such information might be the specific demand volumes of a
firm's customers or the size of an organisation's vehicles. This
information will probably already be available in the Management
Information System (MIS) of the organisation. Effective decision support,
for problems making use of this type of information, requires a suitable
link to the MIS database and an interface that is customised to the
Spatial Decision Support Systems for Large Arc Routing Problems
28
specific problem. A good example of such a routing system was that used
in the Air Products Corporation in Canada (Bell, Dalberto et al., 1983). In
order to organise the larger amounts of information incorporated in
routing systems, database management systems were increasingly
required. An influential review of the field (Bodin, Golden, Assad and
Ball, 1983) noted the over-emphasis placed on algorithmic issues and
indicated that the future lay in a flexible user interface and in better
database support. By the mid-1980's, routing systems combined
algorithms with the increasing use of graphic interfaces and links to
databases. These systems could increasingly be seen as a form of DSS, as
they contain the recognised components of a DSS, i.e. the interface, solver
and database modules. The vendors of these systems increasingly began
to emphasise the flexibility and ease of user interaction of their products.
A good example of the new class of microcomputer based routing software
introduced in the mid-1980’s was the PC version of the ROVER software,
with which the author has had some experience. This was a development
of earlier mainframe software (Fisher, Greenfield, Jaikumar and Lester,
1982). This package could be regarded as a DSS, as it comprised a
database, an interactive interface, and a modelling component, based on
the Fisher-Jaikumar algorithm (Fisher and Jaikumar, 1981). This
software allowed the user select a set of points to be routed from the
database. When solved the user could modify the solution in a number of
ways. While the system could make use of a true distance matrix, in
addition to the co-ordinate approach, no road network detail was
displayed on screen. Consequently, only limited information was
displayed to the user. The system benefited from the use of the most
powerful microcomputers available and required a twin monitor
configuration. Report information was displayed on a text monitor while
the graphical display showed the spatial layout of the routes (Figure 2-2).
Where geographic features affecting the route structure exist, the lack of
information displayed on screen could make decision-making more
difficult. For example, in Figure 2-2 there is obviously some sort of
Spatial Decision Support Systems for Large Arc Routing Problems
29
restriction North of the depot, perhaps a sea inlet. However, from the
information displayed on the interface it is not quite clear what this
might be. The limited visual details provided in this type of interface
meant that the user would probably also need a paper map for additional
information.
How well does this type of system meet the requirements of a DSS as
described in Chapter 1? The ROVER system is a candidate for
consideration as a DSS, as it comprised a database, an interactive
interface and a modelling component, based on the Fisher-Jaikumar
algorithm. This could be characterised as an optimisation model in Alter’s
framework. The model optimises the total distance travelled, subject to
restrictions on vehicle capacity and time window restrictions at the
customer locations. Limitations can be placed on the route structure by
time windows or by specifying the first or last customers to be visited.
When a solution has been derived users can alter the route sequence
manually. By using the interface, a user can perform operations such as
selecting a block of locations to be moved to another route. This type of
Figure 2-2: Graphic display of ROVER software
Spatial Decision Support Systems for Large Arc Routing Problems
30
system represented a dramatic improvement over previous technology,
yet this class of system falls short of the full flexibility demanded by a
DSS. The system does not support a full spectrum of control, as the model
used is a largely a black box. The information provided to the user was
incomplete, owing to the limited nature of the information displayed on
screen. There are few options for the integration of this system with other
types of software.
The arrival of software innovations such as the graphical user interface
(GUI) has had an important effect on DSS design. Firstly, the
incorporation of graphical features in the software provides design
guidelines for systems, making it easier for users to learn to operate the
software. This advantage has existed for Apple Macintosh software for
many years and is now evident in Windows systems. A second important
advantage of modern graphical operating systems is that they allow easy
interchange of information between applications. These features can be
used to pass information to and from routing systems and consequently
can increase system integration. Modern routing DSS has taken
advantage of developments such as graphical user interfaces. Recent
examples include; the Fleetmanager DSS in New Zealand (Basnet,
Foulds and Igbaria, 1996), a system for routing vehicles delivering gas
cylinders in Hungary (Fölsz, Mészáros and Rapcsák, 1995) and an
example dealing with the delivery of dairy products in Spain (Adenso-
Diaz, González and García, 1998).
2.3 Impact of DSS Developments on Routing DSS
2.3.1 DSS developments
In the period since routing systems were first introduced, a variety of
innovations have occurred. New algorithms have been devised and new
forms of computer technology have become common. These developments
meant that computer graphics became relatively inexpensive and that
standard business PCs costing a little over one thousand pounds became
Spatial Decision Support Systems for Large Arc Routing Problems
31
sufficiently powerful to do useful work. Eom et al (1997) note that OR/MS
techniques remain important for DSS, but that these techniques
increasingly are being embedded in systems. These systems incorporate
graphics, visual interactive techniques and a variety of artificial
intelligence techniques including expert systems and neural networks.
Authors reviewing the routing field have noted the potential of these
developments. Both academic and commercial routing software has been
introduced which meets more closely the definition of a DSS. A number of
systems sought to enrich the system dialogue with the user by
incorporating some sort of artificial intelligence. This trend was enhanced
by the exploitation of the improving graphics facilities to provide for
richer user interaction with the system. One theme was the use of
superior graphics and in the provision of enhanced database support to
provide for the display of additional information. Various proposals
related to the provision of a system providing a range of algorithmic
solution techniques, allowing the user select the appropriate technique.
2.3.2 Artificial Intelligence
The DSS field has been influenced by a variety of other technologies and
these have played an important role in routing DSS. One frequently
mentioned area of influence on DSS is that of expert systems and
artificial intelligence (AI). It has been argued that routing DSS can be
enhanced by including a knowledge base component, for instance in
Duchessi, Belardo et al. (1988). This is especially likely to be true for a
DSS designed to support more complex routing problems. In a review of
the vehicle routing field, Fisher (1995) notes that many solution
techniques are designed for specific and limited conditions and suggests
that artificial intelligence techniques may help to identify the appropriate
technique to use. This might arise in the context of model management,
where AI techniques could provide a means to identify the appropriate
models to use (Desrochers, Jones, Lenstra, Savelsbergh and Stoughie,
1999). AI techniques can also have a role in the scheduling phase of
Spatial Decision Support Systems for Large Arc Routing Problems
32
routing decisions, for instance in the assignment of buses to fixed routes
(Chang, Yeh and Cheng, 1998).
One important prototype DSS is the prize winning Tolomeo system
(Angehrn and Lüthi 1990; Angehrn 1991). The design of this system
starts from the belief of the authors that the interface (or language
system) is the most critical element of DSS design. Tolomeo is described
as a DSS generator that can be applied to a wide range of specific
problems. The user can express their needs to the system by modelling by
example and Tolomeo uses artificial intelligence techniques to analyse
the problem representation. A wide range of algorithmic procedures are
provided in the system and the AI component helps the user identify the
appropriate algorithmic representation of a problem, including routing
type problems. In this system, the user can define the problem by using
visual interactive methods and the system then brings the appropriate
modelling techniques to bear on the problem.
2.3.3 User Defined algorithms
Bodin and Salamone (1988) describe a relatively early prototype system
that exploited PC graphics to provide a flexible interactive routing
system. This system, part of a project known as FULCRUM, allowed the
user construct his or her own algorithms for developing an initial solution
and for improving on it. This prototype raised some questions about how
best to represent problems graphically and indicated the need for better
user access to problem data, reflecting the database requirements of a
DSS.
ALTO (Potvin, Lapalme and Rousseau, 1989) and its successor Micro-
ALTO (Potvin, Lapalme and Rousseau, 1994) are good examples of a
more flexible ‘DSS’ approach to routing problems. The designers of Micro-
ALTO contrast the user control over the algorithms used in their system
with the ‘black box approach’ used in commercial routing software. Micro-
ALTO is designed to support node routing problems and provides a
Spatial Decision Support Systems for Large Arc Routing Problems
33
number of general tools and operations to support solving this class of
problem. The system provides tools for the management and display of
transportation networks, including the ability to edit networks and add
or delete nodes and arcs. Micro-ALTO allows user control of the
specification of the problem; the location of the depot, the location of
customers, specification of customer characteristics and definition of the
vehicle fleet characteristics. This information is manipulated in a user
friendly graphical user interface (GUI) environment. Having defined the
network and the problem characteristics, the user can build a solution
procedure from a number of node routing heuristics. These can be easily
applied to the problem, allowing experimentation to identify the best
solution strategy.
2.3.4 Visual Interactive techniques
Visual interactive (VI) modelling is increasingly seen as a relevant
approach for many OR/MS problems (Hurrion, 1986). Bell (1985)
describes two types of VI models. Representational models use graphics to
display the operation of model, this might include the use of bar graphs or
pie charts present numeric results. Iconic graphic models represent a
system rather than a model, the maps used in routing would fall into this
category. Bell uses the example of routing problems to support the use of
VI techniques to facilitate data validation, improve as errors in the
network used can be readily identify from a map representation on
screen. This reflects the major contribution of VI models which is to
communicate with the user and therefore achieve greater user support for
the system. The area of visual modelling continues to be an important
one, see the comprehensive review by Jones (1994b) and the commentary
by Bell (1994) in the same edition of the ORSA Journal on Computing.
It has been argued that the user interface is the most important
component of a DSS and that the interface design may provide a
framework for the entire DSS (Jones, 1991). A similar philosophy
underlies the Tolomeo system described above. By designing the entire
Spatial Decision Support Systems for Large Arc Routing Problems
34
DSS around the interface, the user's view of the problem can be more
accurately captured, thereby providing more effective decision support. In
the case of vehicle routing, the user's view of the problem is a spatial one
and the user can most effectively interact with a system that
accommodates this view. The trend within routing systems has been to
facilitate the spatial representation of routes as part of the interface, in
keeping with the user's perception of the problem.
Jones (1994a) discusses the concept of anchoring, which recognises that
people prefer the problem representation that they are first introduced to.
In Western societies, where a computer is not used, problem solvers
traditionally make use of paper maps for routing problems (this need not
be true in other cultures (Sahay and Walsham, 1996), (Walsham and
Sahay, 1999)). These maps provide a number of geographic reference
points other than just the location of the customers to be routed. In order
for a routing DSS to accommodate this anchored view of the problem, it
may be appropriate to display geographic information other than just the
customer locations. The most obvious need is for a display of the road
network, but other geographic information may also contribute to user
decision-making. Therefore, additional mapping information, such as
street layouts, has become a feature of routing DSS generally.
Progressive systems such as Tolomeo (Angehrn, 1991) or Micro-ALTO
(Potvin, Lapalme et al., 1994) display significant amounts of network and
other geographic information.
2.4 Arc Routing Systems
2.4.1 Characteristics of Arc Routing DSS
As has been the case with algorithmic development, the majority of
routing systems are directed at point-based problems rather than arc
routing problems. However, a number of arc routing systems do exist.
The use of arc routing algorithms imposes requirements on a routing
system that are less important than in the case of a point-based system.
Spatial Decision Support Systems for Large Arc Routing Problems
35
Comprehensive network information is an absolute requirement of an
arc-based system. This may require not only the existence of roads but
restrictions such as no right/left turns. Therefore, arc routing techniques
are incorporated in systems that typically display and store a great deal
more geographic information than is usual in point routing applications.
This reflects the fact that arc routing systems are by definition interested
in the detail of road networks.
However, many arc routing problems also differ in the nature of the data
being used. Many node routing applications are concerned with deliveries
to customers and are quite dynamic in nature, as the delivery volumes for
each customer may change daily. This data is specific to the routing
application. Arc routing problems are often concerned with visiting every
house on a road, for postal delivery or refuse collection, and the data is
therefore relatively static. Data of this sort is frequently derived from
demographic information for the region of service. Where demographic
data is associated with geographic entities then these may become
relevant to an arc routing system.
2.4.2 Specific arc routing applications
Eglese and Murdock (1991) built a DSS to support the routing of road
sweepers travelling in rural areas of Lancashire in the north west of
England. This system was built using Turbo Pascal and incorporated an
onscreen map display. Visual interactive techniques were seen as an
important component of the system. Information on the progress of the
solution was provided by colour changes on the road segments visited.
The system provides for user interaction with the solution procedure by
using a manual override facility. The overall system was seen to be
useful, and brought about a significant improvement in the efficiency of
the routes for street sweepers in this region. Another project in the same
region concerned the routing of winter gritters (Li, 1992; Eglese, 1994; Li
and Eglese, 1996). Unlike the previous case, this problem requires only
one pass per road segment.
Spatial Decision Support Systems for Large Arc Routing Problems
36
An application designed to support the routing of refuse vehicles in
Portugal (Coutinho-Rodrigues, Rodrigues and Climaco, 1993), provided
graphics and heuristic solution techniques on an IBM PC/AT. This class
of microcomputer is some fifty times less powerful than those available at
the end of 1998. This system used the Pascal programming language and
provided user-friendly graphics and data handling. While this system did
not employ a fully-fledged database, it did provide useful support for staff
involved in planning routes for refuse collection in five Portuguese cities.
An unusual routing application is the arc routing system for delivery of
cattle feed in Arizona (Tracey and Dror, 1997). In this case, animal feed
must be delivered to troughs situated in cattle pens along ranch roads
(the arcs). As this type of problem had not been approached using
computer-based techniques previously, it was felt that the display of
information and managerial interaction were important requirements, in
addition to the performance of the solution techniques used. The
requirements for such a system were to have a graphic display, to have a
development environment, which allowed programming of the
application, and to have data query and storage facilities. These
requirements were met by using a GIS as the basis of the system, in this
case PC Arc/Info (ESRI). The developers programmed a number of
routines to customise the system and provide information in a format
suitable for the specific application. This synthesis of arc routing
techniques and a system that could display information of interest to
management provided a successful DSS.
2.4.3 Comprehensive arc routing software packages
The GeoRoute package (Lapalme, Rosseau et al., 1992; Georoute, 1998)
represents the culmination of work undertaken at the Centre de
recherche sur les transports de l’Université de Montréal (CRT). This is
one of the foremost centres of vehicle routing research in the world.
GeoRoute was designed for urban routing problems and featured an
interface with relevant street information being displayed (Figure 2-3).
Spatial Decision Support Systems for Large Arc Routing Problems
37
Network editing features are provided to allow for the alteration of the
comprehensive street database. GeoRoute contains a general framework
within which a variety of node and arc routing algorithms could be used,
inspired by the ALTO software (Potvin, Lapalme et al., 1989) discussed
above, which was previously produced at Montréal. The arc routing
component of GeoRoute was also developed from earlier work (Roy and
Rousseau, 1989). This type of software is very much inspired by the DSS
viewpoint as it offers comprehensive database and model management
facilities (Table 2-1).
GeoRoute has been enhanced and developed for a variety of arc routing
applications. A variant of the program for postal routing is in use in
Canada and number of European countries. Another program variant is
in use for other arc routing applications such as waste collection,
Figure 2-3 : GeoRoute interface
Spatial Decision Support Systems for Large Arc Routing Problems
38
recycling collection, street cleaning, salt spreading and snow removal.
The latter application is especially important in a Canadian context.
Routesmart (Bodin, Fagan and Levy, 1992; Bodin and Levy, 1994) is a
comprehensive street routing system with point and arc routing
components designed to accommodate the complexities of street routing
in urban areas (Bodin, Fagan and Levy, 1997). This software has
provision for one way streets, no right turns and other restrictions. This
software grew out of a number of street routing projects (Bodin and Levy,
1991; Wunderlich, Collette, Levy and Bodin, 1992). Routesmart provides
a set of algorithms that operate with the TransCAD or Arcview GIS
software. These GIS products provide good interface facilities for the
display of maps, which is tailored by the Routesmart routines. The use of
GIS technology to provide an accurate representation of the road network
is seen as a critical component of the software.
Table 2-1: Features of GeoRoute (Georoute)
Enter and display on the map service location attributes such as
quantity, service time, servicing time windows and servicing pattern
(linear, zigzag).
Quickly select and/or colour routes or serviced locations based on the
value of one or more attributes.
Specify the vehicle type, serviced locations, and deadhead travel
paths for a given route.
Choose where and when complementary route activities (ex.: lunch,
unloading/refill, etc.) must take place.
Freeze a subset of routes you do not want to be modified.
Consult up-to-date detailed and summary route statistics.
Spatial Decision Support Systems for Large Arc Routing Problems
39
2.5 The Future of Routing DSS
2.5.1 Trends in Routing DSS
Over the years routing systems have reflected the trend for OR/MS based
DSS generally to be very much model-driven with limited interface and
database components. In earlier years, these systems constituted
modelling software with inflexible data entry and output, the limited
nature of these systems meant that only experts could easily use them. In
recent years, relatively elaborate systems have been built where the
model is just one component of a comprehensive system. Later systems
have employed more sophisticated database techniques, allowing better
integration with other sources of data. Routing DSS is now recognised by
practitioners as an important tool for logistics management (Andel,
1996).
New technologies will influence the development of routing systems.
Network and Internet technology will allow distributed systems that
separate the location of decision from the database and modelling
components. Other developments in user interface design and visual
interactive modelling allow greater ease of use by non-expert users, make
these systems accessible to managers.
These developments will not remove the need for algorithmic
development, but will leverage this development into more useful
systems. Richer data input and more complete interaction with the user
will provide more realistic problem representations, but will require that
the solution procedures used are flexible enough to accommodate these
trends.
2.5.2 Geographic Information Systems
In the arc routing systems described above there has been great emphasis
on the provision of a detailed geographic database and the display of this
information on screen. These features are similar to those provided by a
Spatial Decision Support Systems for Large Arc Routing Problems
40
GIS. Advanced arc routing systems either emulate GIS features in a
customised program (GeoRoute) or exploit GIS technology to build a
routing system (Routesmart). GIS technology is therefore of considerable
relevance to the future of routing DSS. Subsequent chapters of this
dissertation will examine more closely the nature of GIS and its
relationship to routing problems.
Spatial Decision Support Systems for Large Arc Routing Problems
41
Chapter 3 : Geographic Information Systems
3.1 Geographic Information Systems (GIS)
3.1.1 Development of GIS Technology
One area of computer application that has expanded enormously in
recent years is that of Geographical Information Systems (GIS). This
development has largely taken place within a community composed of
traditional users of geographic data, in fields including geography,
geology, planning and forestry. The needs and paradigms of people
working in these fields, rather than the IS or OR/MS communities, have
influenced the development of the technology. The growth of GIS has
been rapid, as numerous potential applications have been recognised. A
recent paper prepared on behalf of the European Commission (IMO,
1995) indicates that the international GIS software market is worth
about one thousand million US dollars (€1100 million) in 1996. The value
of the GIS sector is also determined by the large amounts of computerised
geographic data now available. Indeed the trend in GIS has been for the
data to cost more than the software.
As is the case with DSS there are numerous definitions of GIS; for a
review of these see Maguire (1991). Most call for a system for storing and
displaying spatially or geographically related data. The acronym GIS has
also been used as an abbreviation for Geographical Information Science.
For example, a prominent journal in the area has recently changed its
name from the International Journal of Geographical Information
Systems, to the International Journal of Geographical Information
Science. This trend appears to reflect a view that there exists a core set of
geographical information processing techniques, which form a separate
body of research, distinct from the downstream information system
implementations of those techniques. Given this trend it seems
Spatial Decision Support Systems for Large Arc Routing Problems
42
appropriate to define a GIS as an information system primarily concerned
with the techniques associated with geographical information science.
The development of GIS has been driven by technological innovations
that facilitated the storage and manipulation of large quantities of data.
This trend has its origins in the 1960s, with the fragmented use of
computer technology for automated cartography and the introduction of
address matching software. The development of comprehensive GIS
software required improvements in graphics and database techniques. By
the 1980s, a number of different forms of commercial GIS software
became available, including widely used products such as ARC/INFO
from ESRI. Such software generally used UNIX workstation based
proprietary technology. At the end of the 1980s, GIS software become
available on standard microcomputers, reflecting the increase in PC
performance to levels previously associated with workstations. By the
1990s, many different kinds of commercial GIS software were available.
GIS technology had by this time achieved widespread use in its
traditional areas of application, such as forestry and natural resource
applications. This explosion in the use of computer technology can also be
seen in other areas, where a virtuous circle of declining hardware costs
leads to larger software sales and therefore reduced software costs.
3.1.2 GIS Data
In addition to developments in data handling technology, widespread use
of GIS also required the availability of suitable geographic data. This was
influenced by technical developments in data capture. Initially data
collection from maps was achieved by digitisation using a hand-held
device, an error prone and time consuming process. Later the scanning of
entire maps became feasible and this approach greatly speeded up data
collection. Nevertheless collecting accurate digital data from paper maps
remains a difficult process, as human intervention is usually needed to
identify the objects on the map. Other technical developments such as
Global Positioning Systems (GPS) have greatly facilitated the collection
Spatial Decision Support Systems for Large Arc Routing Problems
43
of geographic data in general. This technology has particular application
in routing as these devices can be placed on vehicles, allowing their
physical location be continuously recorded. Advances in data storage
technology have facilitated the distribution of large amounts of data.
Despite the technical developments, geographic data collection for once
off projects remains an expensive and error-prone procedure. However, as
GIS use becomes more widespread, datasets can be shared by different
users, reducing the cost of the data per project. The increasing use of GIS
was both facilitated by, and responsible for, the increasing volume of
digital spatial data becoming available in developed countries. The wider
availability of appropriately formatted data has also reduced the cost of
data assembly for a given GIS project. The growth of GIS has been driven
by the importance of spatially related data. It is estimated that up to 80%
of data needed for the activities of business and government is spatially
related (Franklin, 1992). There is increasing interest in the role of GIS in
business, with an increasing output of papers and books on the subject
(Grimshaw and Clarke, 1996; Grimshaw, 2000).
The level of GIS development varies from country to country. In the USA,
the basic government mapping data is available free of charge and this
has greatly facilitated the growth of GIS. In European countries, mapping
data tends to be available from government agencies on a cost recovery
basis, with a consequent increase in GIS data cost and a slowdown in its
use by the public. Despite standardisation efforts by the European Union,
concerns remain about the differences in the data formats used in various
European countries (Harding and Wilkinson, 1996). In Ireland, data has
been historically quite expensive and of poor quality because of
inadequate updating of maps. The situation is now improving with the
Ordnance Survey aiming to provide comprehensive digital mapping data
(Anonymous, 2000b). Currently off-the-shelf road networks are available
at various levels of detail, although these are relatively expensive.
Several private sector vendors (Gamma, IRIS) are producing their own
Spatial Decision Support Systems for Large Arc Routing Problems
44
digital maps from satellite photography and global positioning system
(GPS) surveys. Some of these vendors are attempting to collect additional
data on speeds per individual road segment. This type of data collection
will greatly facilitate GIS use in Ireland in the future (Anonymous,
2000a).
3.2 GIS Software
3.2.1 Current developments in GIS software
The rapid expansion in the use of computer technology seen in the GIS
field can also be seen in other areas, where the virtuous circle of declining
hardware costs leads to larger software sales and consequent reduced
software costs. Consequently some mapping software is becoming
available on a mass-market basis, for example the inclusion of simple
mapping facilities in spreadsheets first introduced in Lotus 1-2-3 Release
5. A similar mapping facility is now available in Microsoft Excel. This
mass-market use of mapping and GIS products creates a large demand
for spatial data, causing more to be made available. The decision-makers,
that use such basic mapping products, frequently go on to become
interested in more sophisticated software. Recent improvements in
mainstream PC technologies facilitate this increase in the use of spatial
data. These include inexpensive gigabyte sized hard disks, large high-
resolution colour monitors, graphics accelerators and rewritable CD-ROM
storage.
There now exists a spectrum of software from simple mapping packages
to complex systems for handling spatial data. Many of these systems
operate on widely available PC equipment that already is being used for
decision-making applications. The cost of spatial data for use with this
software is declining and there is greater interest by the vendors of such
software in new (non-traditional) applications. These new applications
include the extension of GIS techniques into fields more usually
associated with OR/MS, for example many GIS products now provide the
Spatial Decision Support Systems for Large Arc Routing Problems
45
facility to find a shortest path in a road network. Transport applications
are increasingly seen as an important application by GIS vendors (Lang,
1999).
3.2.2 Components of GIS
A GIS makes use of geographical and attribute data (Chrisman, 1997).
Attribute data, addresses, populations, etc., are associated with
geographical data. Attribute data can be stored in a conventional
database or flat-file format. Geographical data may be represented as
points, lines or polygons. It is the handling of the geographical data, such
as the existence of rivers, roads or contour lines that requires the use of
the special techniques that characterise the use of GIS. The full
representation of a map requires relatively large volumes of information
and high-powered software and hardware to deal with this volume of
information.
A GIS, as distinct from a simple mapping program, will have a database
of geographic data, allowing linkages between different types of data and
Table 3-1 : Data operations in GIS
GIS Operation Data Operated on Example
Adjacency Point/line
Polygon/polygon
Nearest point to a line
Neighbouring regions
Enclosure Point/Polygon
Line/Polygon
Point in region
Line in region
Buffering Point
Line
Area near point
Area within 500 m of a road
Overlay Polygon/Polygon Area within 2 different types of
region
Spatial Decision Support Systems for Large Arc Routing Problems
46
the ability to query this spatial data. A comprehensive GIS package will
provide a variety of data handling routines (Table 3-1). For example a
GIS database query would allow a group of spatial entities to be
identified, for example all regions close to a region of interest (Figure
3-1). It would allow identification of all roads within a certain distance of
a river by using buffering techniques. In the routing context, this might
allow the identification of road segments close to a route (Figure 3-2).
Therefore, while traditional database approaches can support queries on
the attribute data, GIS is defined by its ability to cater for spatial
queries. Not all applications of mapping data require the full power of a
GIS. A comparison exists with non-spatial data where simple data
manipulation can take place in a spreadsheet program, without the use of
a fully-fledged DBMS.
Figure 3-1 : Use of spatial techniques to identify neighbouring
regions
Spatial Decision Support Systems for Large Arc Routing Problems
47
3.3 GIS and DSS
3.3.1 Relationship between GIS and DSS research
Many areas of DSS application are concerned with geographic data, an
influential early example being the GADS system (Grace, 1977). A more
recent important prototype DSS, Tolomeo (1990), uses a geographical
context for the development of visual interactive techniques. However
mainstream GIS techniques have had limited impact on DSS research.
This situation is beginning to change. Some, but by no means all, recent
DSS textbooks include GIS as a component of management support
systems (Mallach, 1994, Page 428; Turban, 1995). While these texts
stress the usefulness of geographically related information, they do not
provide a complete picture of the relationship between GIS and other
management support systems.
Figure 3-2 : Example of GIS use to identify street segments close
to a route (Keenan, 1998c)
Spatial Decision Support Systems for Large Arc Routing Problems
48
GIS related research is beginning to make an appearance at conferences
associated with DSS. Notable examples include GIS based sessions
organised at the Hawaii International Conference on System Sciences
(HICSS) (Grimshaw, 1996; Murphy, 1996) and the conference of the
Association of Information Systems (AIS) (Keenan, 1997). Some of the
authors presenting at these conferences have argued for better linkages
between GIS and DSS and optimisation (Rolland and Gupta, 1996). There
is increasing evidence of interest in GIS at OR/MS conferences, where
applications integrating GIS and Operations Research techniques are
discussed. Academic journals associated with the DSS field are beginning
to publish GIS related papers. Wilson (1994) explored the relationship
between DSS and GIS in a 1994 paper in an IS journal. Crossland (1995)
presented empirical evidence of the usefulness of a spatial approach to
decision-making, while another study extended cognitive fit theory to
map based presentations (Dennis and Carte, 1998).
Geographical techniques have been identified as being relevant to the
general field of computer graphics, on which much DSS research is based.
Researchers from the IS tradition have noted that computer technology is
especially appropriate for the display of mapping data (Ives, 1982).
Cartography has been seen as being an important source of principles for
the design of business graphics (DeSanctis, 1984). Recent work by
Smelcer (1997) indicated that map use was superior to the use of tables in
some situations. Another study (Swink and Speier, 1999) indicated that
the quality of decision-making using GIS could be influenced by the level
of detail included in the presentation of the data. A recent paper in the
prestigious MIS Quarterly discussed the benefits of SDSS for both
inexperienced and experienced decision makers (Mennecke, Crossland
and Killingsworth, 2000).
GIS techniques are beginning to have an impact on DSS applications.
Surveys of DSS applications, for instance (Eom, Lee et al., 1993), have
identified marketing and routing as important areas of DSS application.
Spatial Decision Support Systems for Large Arc Routing Problems
49
These fields are also recognised as areas of GIS application (Maguire,
1991). Many areas of GIS application would be familiar to the OR/MS
field. One traditional example of the use of modelling is in selecting a
facility location (Harrison and Wills, 1983) and this can be related to GIS
techniques (Ding, Baveja and Batta, 1994). There might be a number of
criteria for such a decision; some of these would be spatial in nature. For
instance, a school might need to be located near to the districts from
which potential pupils would travel. A refuse disposal facility, on the
other hand, might need to be located away from populated areas. GIS
based spatial operations could be used to provide an index of suitability
for sites for such a facility.
These applications often employ demographic data that is widely
available in a suitable format for use in GIS software. Many OR/MS
applications feature a specific problem formulation set in a geographic or
demographic context common to many users. This common set of data is
likely to be available, at reasonable cost, for the GIS. A number of
specialised GIS products exist to exploit this data for a particular class of
applications, an example is the marketing oriented GIS products from
Tactician Corp (Tactician).
3.3.2 Is a GIS a DSS?
Within the GIS field there is increasing interest in the use of GIS
software to provide decision support. This is reflected in a recent GIS
conference entitled “DSS 2000” and in the increasing appearance of
papers referring to Spatial Decision Support Systems (SDSS) at GIS
conferences. While an increasing number of GIS based applications are
described as DSS, these descriptions suffer from a lack of agreement on
what exactly a DSS actually constitutes. This reflects the varying
definitions of DSS in the DSS research community, but also arises from
the separation of GIS research from mainstream IS. As Maguire (1991)
points out, some authors have argued that a GIS is a DSS. In some cases,
GIS applications are described as being DSS without reference to the
Spatial Decision Support Systems for Large Arc Routing Problems
50
DSS literature. Many GIS based systems are described as being DSS on
the basis that the GIS assisted in the collection or organisation of data
used by the decision-maker. This may be a reflection of the trend
identified by Keen (1986) for the use of any computer system, by people
who make decisions, to be defined as a DSS.
However, other authors justify GIS being regarded as DSS in terms of the
definition of DSS. According to this line of argument GIS meets the
requirement of being a DSS, as GIS contains an interface, a database and
spatial modelling components. Mennecke (1997) sees SDSS as an easy to
use subset of GIS, which incorporate facilities for manipulating and
analysing spatial data. These differences of definition reflect the differing
needs of decision-makers that use spatial information. For many of the
current SDSS applications, the main information requirement of the
decision-makers is for relatively structured spatial information. This
group may indeed find that standard GIS software provides for their
decision-making needs.
Many widely accepted definitions of DSS, introduced in Chapter 1,
identify the need for a combination of database, interface and model
components directed at a specific problem. In terms of these definitions, a
GIS would not be regarded as a DSS as it lacks support for the use of
problem specific models. However, the view of GIS as a DSS is has some
support in the existing definitions of DSS. Alter (1980) proposed an
influential framework for DSS that includes data driven DSS that do not
have a substantial model component. As the database component, rather
than models, is central to standard GIS software, it could be regarded as
an Analysis Information System in Alter's framework. Common to all
definitions of DSS is a sense that these systems must support a particular
type of decision. This characteristic distinguishes DSS from general
purpose MIS. While GIS applications may contain the information
relevant to a decision, they are usually general-purpose systems, not
focused on a particular decision.
Spatial Decision Support Systems for Large Arc Routing Problems
51
The view that SDSS is a subset of GIS reflects the need for decision-
makers to focus on their specific problem, and their lack of interest in GIS
features outside this domain. This view suggests that the techniques
needed for SDSS are already within the GIS domain and that a subset of
these techniques can be applied to a particular problem. For some
problems within the traditional area of application of GIS this approach
is sufficient. However, the range of problem areas where SDSS can be
used greatly exceeds the traditional areas of GIS application. It is widely
accepted (Franklin, 1992) that more than 80% of business data has a
spatial component. Therefore, SDSS can potentially be used to assist
decision-makers in problem areas where models drawn from other
disciplines have long been found useful. For these areas at least, a
standard GIS cannot be said to be a DSS as such a system lacks the
support that the use of customised models can provide. For this wider
range of second-order uses of spatial data, additional processing or use of
non-spatial models is required to support fully the decision-maker. The
view of GIS as a DSS would seem to propose that all existing techniques
from OR/MS, accounting, marketing etc., be included in standard GIS
software, a subset of which could then be used to build a SDSS. This all-
inclusive view may reflect a limited perspective on the range of modelling
and solution techniques that can be applied to spatial data.
This dissertation is based on the premise that SDSS can most usefully be
regarded as an extension of GIS, rather than just a subset of it. This view
is contrary to some authors, for instance Mennecke (1997). From this
viewpoint, a SDSS will employ a subset of GIS techniques in combination
with problem specific models not found in GIS. This will extend the
present use of GIS as a DSS, to a situation where a GIS will be used to
build a DSS. This will allow the range of decisions supported to be
extended from those incorporated in the standard GIS software, to other
specialist fields where some customisation of the software is needed. One
such field is routing and specifically the arc routing applications that are
the subject of this dissertation. It is therefore relevant to examine SDSS
Spatial Decision Support Systems for Large Arc Routing Problems
52
in general and to specifically focus on the role of these systems in the
routing domain.
3.4 Spatial Decision Support Systems
3.4.1 SDSS decision makers
SDSS is an important subset of DSS, incorporating GIS techniques with
other modelling approaches, whose potential for rapid growth has been
facilitated by technical developments. The availability of appropriate
inexpensive technology for manipulating spatial data enables SDSS
applications to be created. The benefits of using GIS based systems for
decision-making are increasingly recognised. In a review of GIS, Muller
(1993) identified SDSS as a growth area in the application of GIS
technology. However, the value of SDSS is not determined by its
innovative use of technology. Rather the contribution of these
applications will be determined by how well they support the need for a
spatial component in decision-making.
An important group of SDSS users are those based in the traditional
areas of application of GIS. In these fields GIS was initially used as a
means of speeding up the processing of spatial data and for the
completion of activities that contribute directly to productivity. In this
context, the automated production of maps has a role similar to that of
data processing in business. Decision-making applications will develop
and SDSS become widespread in much the same way that data
processing applications evolved into DSS in traditional business
applications. An example is the DSS for the assessment of geological risk
by Mejia-Navarro (1995). This class of applications is distinguished by a
direct interest of the authors in the spatial operations provided by the
SDSS and by their considerable background knowledge of the spatial
techniques used. For this category of user, the spatial data and the
spatial processing techniques are of direct interest rather than simply
providing the context in which other variables are being manipulated.
Spatial Decision Support Systems for Large Arc Routing Problems
53
The greater complexity of spatial information processing and its greater
demands on IT has lead to the ten to fifteen year time lag in the
development of SDSS (Densham, 1991). With the decline in IT costs,
inexpensive microcomputers can now cope with the demands imposed by
the manipulation of spatial data. The rapid increase in the 1980s in the
use of database managers, led by Dbase II, is being emulated by the
current growth in the use of spatial database tools at present. In the
context of decision support, the increasingly widespread use of PC based
GIS software is reminiscent of the move towards PC based DSS in the
1980s (Table 3-2).
The second group of decision-makers for whom SDSS can make an
important contribution is in fields such as routing and location analysis.
Although the spatial component of such decisions is clear, in the past DSS
design has been driven predominantly by the OR/MS models used. Such
model driven systems often had very limited database or interface
components and the DSS provided little contextual information to the
Table 3-2 : Computerised Support for Decision-making
(adapted from Turban, (1995 Page 19)
Phase Description Traditional Tools Spatial Tools
Early compute, “crunch numbers”,
summarise, organise
early computer
programs, OR/MS
models
computerised
cartography
Intermediate find, organise and display decision
relevant information
database management
systems, MIS
workstation GIS
Current perform decision relevant compu-
tations on decision relevant infor-
mation; organise and display
results. Query based and user
friendly approach. “What if” analysis
financial models,
spreadsheets, trend
exploration, OR/MS
models, decision
support systems.
spatial decision
support systems
Spatial Decision Support Systems for Large Arc Routing Problems
54
user. In future these models will be incorporated into GIS based SDSS,
providing superior interface and database components to work with the
models. The role of the superior GIS data handling facilities will be to
provide a richer context for the use of the specific models and for display
on the user interface.
For the general class of routing problems, the variables of direct interest
might include distance travelled, the number of vehicles used, and the
loads on each vehicle. Early routing DSS were restricted to the use of
data related to directly relevant variables. However, the use of GIS
technology allows the inclusion of other indirectly important information.
For example the inclusion of elevation data would allow more realistic
travel times be used in quantitative modelling of routes. The display of
distinctive natural features such as rivers or mountains on the interface
can make it much easier for the user to understand the representation of
the routes generated. This synthesis of OR/MS and GIS techniques will
provide more effective decision-making.
This dissertation will argue that the use of GIS techniques can extend the
range of decision support to another group of potential SDSS users, those
whose main concern is with vehicle routing problems. GIS will allow
consideration of path constraints that have not been modelled
comprehensively in the past. This group of potential SDSS users has
limited experience of using manual spatial techniques. Such users are not
usually directly interested in the spatial processing techniques provided
by the SDSS but only in the interaction of these techniques with the
OR/MS models. However, the secondary use of spatial data by the models
and the display of spatial information on the interface can greatly enrich
the decision-making process. These could benefit from the geographic
context being fully reflected in the problem representations used.
The third group of decision-makers who will find SDSS important are
those where the importance of both spatial data and modelling is
somewhat neglected at present. In disciplines such as marketing,
Spatial Decision Support Systems for Large Arc Routing Problems
55
additional possibilities for analysis are provided by the availability of
increasing amounts of spatially correlated information, for example
demographic data (Mennecke, 1997). The relevance of GIS to this type of
work is becoming widely recognised (Fung and Remsen, 1997).
Furthermore, the geographic convenience of product supply relative to
customers' locations is an important tool of market driven competition.
The availability of user friendly SDSS to manipulate this type of data will
lead to more use of formal modelling techniques in this field where rather
informal techniques have been used in the past.
3.4.2 GIS as a DSS Generator
Because of the variety of decision-making situations where spatial
information is of importance, clearly SDSS will be an increasingly
important subset of DSS in the future. It is useful to examine the
relationship of GIS software to such systems. Much of the research in the
GIS domain is poorly linked to the traditional DSS literature, even
though the concepts are similar, for instance Djokic (1996) presents a
framework for SDSS similar to that proposed by Sprague for DSS.
Densham (1991) discusses the development of SDSS in the context of the
framework proposed by Sprague (1980) (Table 3-3 below). In building
DSS, specific generators have been designed for certain classes of
problem. In other situations general-purpose software such as
spreadsheets or DBMS packages have been regarded as generators. In
modern DBMS and spreadsheet software, the use of macro languages
facilitates the creation of specific applications. Various generators have
strengths and weaknesses in terms of their provision of the essential
components of a DSS: an interface, a database, and models. In the case of
a spreadsheet, modelling is the basic function of the software; various
interface features such as graphs are provided, but the database
organisation is simplistic. DBMS software, such as Access or Paradox,
has good database support, provision for interface design using forms,
report and charts, but almost no modelling support. In this case, the
Spatial Decision Support Systems for Large Arc Routing Problems
56
modelling support has to be added to the specific DSS built from such a
system (Figure 3-3).
In Sprague’s framework, the SDSS builder may make use of tools that
provide some lower level data processing. In software design these might
include operations such as sorting or searching that are important
algorithms in their own right but which are not of direct interest to the
decision maker. The decision regarding the appropriate mix of DSS tools
and the use of a generator is an important component of the process of
building a DSS. However there is a very real sense in which the types of
DSS design considered for a given class of problem are a function of the
available DSS generators for that class of problem. In practice, a small
DSS project could be built, using an off-the-shelf spreadsheet or DBMS
Table 3-3 : DSS Generator Features
DSSComponent
Spreadsheets DatabaseManagers
GIS
Interface tables, forms,
charts
tables, forms,
reports
multi-layer maps, plots
Database independent cell
entries
linked database
tables
linked spatial and non-
spatial databases
Database
Tools
rudimentary sort
and selection
comprehensive
queries
spatial query
Models built in
mathematical
functions,
statistical and
OR/MS tools
basic mathematical
functions
basic summarisation and
network analysis models
Model
Building Tools
recorded or
programmed
macros
macro and database
query languages
macro languages,
programming interfaces
to other programming
languages
Spatial Decision Support Systems for Large Arc Routing Problems
57
package, in less time than it would take to evaluate fully the range of
alternative methods of constructing the DSS. Therefore, the DSS
solutions actually constructed are strongly influenced by the perceived
availability of suitable generators. Consequently, the effective application
of DSS technology can benefit from additional generator software
becoming available. Awareness of the potential of the use of GIS based
systems as DSS generators will lead to problems, currently being solved
in other ways, being approached by using a SDSS.
3.4.3 Extending GIS to a broader community
The user diversity of potential users of GIS techniques can largely be
catered by a clear focus on the specific problem, rather than on the
SDSS
Spatial Data Non - Spatial Data
Data Access Tools
Models
Interface
GIS ExternalComponents
Figure 3-3 : Building a SDSS by integrating models with GIS
Spatial Decision Support Systems for Large Arc Routing Problems
58
technology used. For ease of system building a DSS generator may be
used, such as a GIS, that has multiple functions. However, for any one
problem or one user many of these facilities may not be needed. Existing
design frameworks such as the ROMC approach (Sprague and Carlson,
1982) introduced in Section 1.3.3, should be used to identify the system
features of interest to the specific user. The general-purpose features of
the generator can then be customised by the system builder to provide the
representations, operations, memory aids and controls appropriate to the
problem. These may differ substantially from user to user. It is an
important characteristic of successful information systems that they
provide information in a format appropriate to the user. Different users of
a given type of information may be accustomed to quite different
presentation formats for the information. This diversity of user
requirement places important demands on the design of the components
of the SDSS, not only the interface but also the database and modelling
components (Grimshaw, Mott and Roberts, 1997).
Another issue that arises when techniques spread to a different class of
application is that distinctive nomenclature may be used in different
disciplines. This poses a problem in the context of SDSS where the
language used by geographers, which underlies the documentation and
interface paradigms for GIS software, may be quite different than that
used by potential users of SDSS. A successful system must provide
system builders with the flexibility to accommodate user preferences. The
DSS components can be configured to provide direct user access to
information of interest, while other features of the DSS provide
contextual information to enrich the decision-making process. Contextual
information can be found in the database, processed by spatial models
and displayed in an appropriate representation on the interface (Table
3-4).
If GIS software is to be seen as a DSS generator, rather than as an end in
itself, then different strategies for interface design might present
Spatial Decision Support Systems for Large Arc Routing Problems
59
themselves. The aim of the system builder must be to cater for the
problem representation of the user, the logical view of the problem, rather
than provide a system too closely related to the physical geographic data.
Different users might have different system representations and
operations. This might operate in a similar way to the concept of sub-
schemas in the context of database management that provide a
distinctive presentation of a database to each user. Not all the data in the
system need be made directly available to every user. Limited access to
information may be provided if required, but the full range of GIS
operations need not be made available. Simplified information
representations, that might be appropriate for users who only indirectly
employ that information, might be inadequate for other users directly
interested in that data. Unlike earlier systems, modern DSS can utilise
richer and more complex information representations, for example GIS
based SDSS is an excellent example of this trend. This implies multiple
features at a level of detail that goes beyond that needed by any one user.
For a given user, some of these representations are directly important,
while others provide only background information (Table 3-4).
A user-based design will not impose unfamiliar control concepts on the
user. For example, a typical operation in GIS might involve selecting a
procedure from several levels of submenu. The spatial data to be used for
Table 3-4 : Contextual Information in SDSS (Keenan, 1998b)
Directly Relevant Contextual Information
Solver customised business models e.g.
marketing, routing, location
General spatial processing tools e.g.
buffering
Database data on planning units, e.g. routes,
administrative regions
Elevation data, base geographic
entities, points, arcs, polygons
Interface decision outcomes, e.g. routes, areas
of influence
Geographic features, e.g. lakes
rivers, mountains
Spatial Decision Support Systems for Large Arc Routing Problems
60
this operation might be then identified by drawing a box on the screen
with the mouse. This approach presents problems for the SDSS user who
is not familiar with many of the operations provided by the GIS. A more
user-centric approach might allow the user draw the box on screen with
the mouse and then a menu would appear offering only that subset of
operations appropriate to that set of data.
Artificial intelligence (AI) techniques might be used to facilitate this
interface simplification. Such an approach might draw on systems such
as Tolomeo (Angehrn and Lüthi, 1990) and ALTO (Potvin, Lapalme et al.,
1994) discussed above (Section 2.3.2). Tolomeo allows the users describe
the problem visually and the interface includes a map that facilitates the
representation of visual features other than those that can be directly
manipulated by the user. These features provide a geographic context
within which the user specifies the problem in terms of the cities to be
visited, etc. This type of intelligent DSS interface could usefully be
incorporated in a fully-fledged SDSS to increase its acceptance to a
broader user community. One such community would be those using
OR/MS in general and routing in particular.
3.5 Current Spatial Decision Support Systems Technology
3.5.1 Suitability of GIS Software for building DSS
There is evidence that GIS software is becoming increasingly suitable for
use as SDSS generator. As GIS designers gain a greater awareness of
decision-making possibilities, their systems will be designed to facilitate
interaction with models. GIS software provides a sophisticated interface
for spatial information. Even limited functionality GIS software will
provide the ability to zoom and to display or highlight different features.
GIS provides database support that is designed to allow for the effective
storage of spatial data. Furthermore, GIS software provides a link
between the interface and database to allow the user easily query spatial
data. However, a GIS is not a complete DSS for the full range of potential
Spatial Decision Support Systems for Large Arc Routing Problems
61
uses of spatial data in decision-making because of the almost complete
absence of problem specific models or support for the organisation of such
models.
The ability to integrate GIS software and models is being facilitated by
modern software techniques that allow a variety of forms of interaction
between the GIS software and the modelling software. The simplest form
of software interaction is the exchange of data. In its most basic form,
this might be achieved by a program preparing a file for subsequent use
by other software. This requires a common file format for use by both
pieces of software. Current trends favour the use of integrated databases
that can be accessed by all types of software. This has led to the use of
common standards, such as SQL and ODBC, which allow a wider range of
programs interchange data. Where databases are used, SQL has become
a standard means of sending database commands. These developments
provide an effective means of exchanging data between GIS and other
programs. For example, separate modelling programs might access a GIS
database. Given the growth in client-server systems, these data exchange
technologies will have an important role to play in the integration of GIS
into DSS.
For a comprehensive system to exist, data exchange alone may not be
sufficient. If a DSS generator is to interact with other software, it will
need to interact directly with other program components. As a DSS
generator is part of a larger system, its interaction features will reflect
the trends in software design generally. The trend in software
engineering has been to make use of self contained modules that interact
in clearly defined and controlled ways. This tendency has led to the
development of structured programming techniques and the use of
modular procedures and functions within programs. In well-designed
software each of these modules should perform a single task and have
well-defined inputs and outputs. Provision exists for these modules to be
included in program libraries that can be used by other programs. In this
Spatial Decision Support Systems for Large Arc Routing Problems
62
context, GIS software can provide an application-programming interface
(API) which allows external programs interact with the GIS software.
GIS software frequently provides some type of API, and this approach has
proved useful for building SDSS. However, this approach requires the
SDSS builder to have a good knowledge of the operation and technical
parameters of the GIS software. The API approach is suitable where ad-
hoc models are being used, but is less suitable where different software
packages need to be integrated. For example, it would be difficult to build
an integrated system using GIS software with an API for the C
programming language and a modelling package with a FORTRAN API.
Using the API approach to add decision support functionality to GIS is
likely to meet the requirement for a comprehensive system, but might not
meet the requirement that a system should be quickly and easily built.
A more recent concept in software engineering is object orientation. This
approach makes use of modules (objects) which bring together both data
and the program elements that operate on that data. The advantages of
the object-oriented approach include information hiding. This means that
the system builder does not need to be familiar with the detailed
operation of the procedures or data structures used, as implicitly the
procedures are designed to work with the data with which they are
connected. This means that the object code can be used without detailed
knowledge of its internal operation. The object-oriented approach
facilitates faster application development while reducing the possibility of
errors being introduced by alteration of the programs. This approach uses
rather larger building blocks than the traditional API approach. The
object-oriented approach may prove somewhat less flexible in meeting the
needs of decision-maker and is less efficient from the point of view of
machine performance. However, for most problems it offers a quick
means of building systems.
Modern programming languages generally include object-oriented
features and GIS software may have an API that can interact with such a
Spatial Decision Support Systems for Large Arc Routing Problems
63
language. The object-oriented concept has been extended to the use of
stand-alone components; these are independent of any particular
programming language. These can pass instructions and data to each
other in accordance with widely used standards. These components may
exist within a single computer or within a network. An example of this
type of technology is the Object Linking and Embedding (OLE)
introduced by Microsoft. Each component has its own data and some
documented operations on that data, commands can be sent to that
component from other programs to perform those operations. If the
industry standards are used a wide range of programs can potentially
make use of these components. Therefore a program can be written in one
of several programming languages and still make use of existing modules.
This provides a degree of flexibility not found with the API approach.
Software applications, for example spreadsheets can make some use of
these standards, allowing communication between these applications and
GIS software. The main advantage of this approach is the flexibility it
allows in the integration of different types of software (Table 3-5).
Present software development trends suggest an object-oriented future, in
which small specialised applications, or applets, will be available for use
Table 3-5 : Software integration techniques for building SDSS
(Keenan, 1998c)
SDSS buildingtechnique
modellingintegration
ease ofsystemintegration
programefficiency
separate programs, data
interchange
low high low
Application programming
interface (API)
high low high
object components e.g. OLE/OCX high high low
Spatial Decision Support Systems for Large Arc Routing Problems
64
as part of a larger package. In the PC environment, the development of
such small applications will be facilitated by the use of visual
development tools. On the PC platform, the dominant standard for
applets is based on Microsoft's OCX standard. An extension of these
trends is the availability of a version of Microsoft's applet technology,
known as ActiveX, developed for use over the Internet.
At present, GIS software is often found on UNIX workstations. However,
in the future, as SDSS spreads to a broader user community, systems are
increasingly likely to be based on widely used industry standards, for
example Microsoft Windows NT. The use of OLE/OCX allows system
builders quickly build systems that concentrate on the detailed modelling
of the specific problem rather than the detail of the lower level processing
of information. Increasing connectivity will mean that these systems will
not stand alone but can connect to more powerful spatial database and
processing facilities by employing client server technology and even by
using the Internet.
3.5.2 Commercial GIS software
Within commonly accepted definitions, GIS software can be combined
with OR/MS models to build a SDSS. Technical developments are moving
in a direction that facilitates this integration. In this context, PC based
GIS products with limited functionality may prove more manageable for
applications design than full workstation based GIS systems. While these
desktop systems lack the power of a full GIS, they may be able to make
effective use of data that has been pre-processed by a full feature GIS. In
order to form the basis of SDSS, however, such systems must offer spatial
database handling with appropriate access tools. It is not sufficient
simply to employ a mapping tool that may provide simple maps, but lacks
spatial query features. The use of this type of technology offers two
possibilities. GIS software may be used for the main interface and
database facilities, using applets for additional modelling or interface
requirements. Therefore marketing or logistics models might be
Spatial Decision Support Systems for Large Arc Routing Problems
65
incorporated in an OCX to be used by PC based GIS software.
Alternatively, the main application might be developed in another
programming language and OCX type applets used to provide some
element of GIS functionality. A number of GIS related tools of this sort
exist, for example Sylvan maps (Sylvan) or MapObjects from ESRI
(ESRI), the market leaders in GIS software. However, these tools have to
be assessed to establish the extent to which they provide true GIS
operations, or whether they are simply mapping tools.
The developments in GIS software since 1990 may allow the use of off-
the-shelf software as the basis for a SDSS. An example of this type of
software is the ArcView package from ESRI. As its name suggests, this
software is primarily designed as to allow the user to view and query
spatial data. ArcView has its own macro language: Avenue, which can
interact with SQL database servers, and use platform specific links with
other software. An optional network analysis package is available for
ArcView allowing its use for a variety of applications that need this
functionality, for example transportation modelling. Another widely used
desktop mapping product is MapInfo (Mapinfo). The MapInfo package
provides the Mapbasic language, which is likely to be developed to
become increasingly similar to other programming tools, such as
Microsoft Visual Basic.
TransCAD (Transcad, 1996) is a PC-based GIS designed specifically for
managing transportation data and to facilitate the use of transportation
models. As such, it is an excellent example of a potential DSS generator
as it provides a number of features that specifically support
transportation modelling. These include provision for a road network
layer, with the ability to store relevant network characteristics such as
turn penalties. The concept of a route layer is supported, allowing
multiple routes to use each road. TransCAD allows extensive tailoring of
the interface around the standard GIS components using the GDK
supplied (Geographic Information System Developers Kit). Applications
Spatial Decision Support Systems for Large Arc Routing Problems
66
developed using this toolbox can communicate with external software
using the widely used DDE, OLE and ODBC standards. SDSS
applications can be built using a combination of these features, for
instance a customised interface together with the use of macros.
3.5.3 Future directions
Given the advances in IT, modern DSS can incorporate a more extensive
range of directly and indirectly relevant information. The designers of
these systems must aim to provide maximum user control over those
aspects of the decision where the user has specific expertise, while
providing the user with maximum support for areas where the user is
less expert. This may require that DSS generator software, such as GIS,
be designed to be flexible so that different types of user can make use of
the intelligence in the system for less critical parts of the decision.
This dissertation suggests, therefore, that much DSS development in the
future will use relatively complex combinations of DSS generators and
tools. A substantial variation in the types of problem and user will exist
within the general group of systems built from such generators. Spatial
systems are a good example of a class of sophisticated DSS. Such systems
have largely been used in the past for problems where the manipulation
of spatial data was the key or only information component of the decision
to be taken. This type of decision required a system that provided users
with full control over the spatial operations in the system. In the future,
SDSS use will be extended to applications where the spatial information
is only an interim stage or a subset of the information required for the
decision. Such information provides the geographic context within which
specific decisions are taken. Users dealing with this broader set of
applications need to be given control over the important variables in the
decision while other processing is performed without the need for
extensive user interaction. With the development of such systems, new
classes of decision and new types of user can be effectively supported.
Spatial Decision Support Systems for Large Arc Routing Problems
67
Chapter 4 : Modelling routing problems in GIS
4.1 Introduction 1
4.1.1 GIS and routing
In Chapter 2 we discussed the development of routing DSS and saw a
number of relatively sophisticated routing systems which incorporate
geographic data and network algorithms, without using conventional GIS
software, for example GeoRoute (Georoute, 1998). Other routing systems
use GIS technology as a platform for building a routing DSS. The use of
mapping data enriches the problem representation found in these
systems. However, many routing systems that use GIS do so only to
provide map display and take little advantage of the spatial data
processing capabilities of modern GIS software. This chapter examines
the way that GIS functionality can be used to support a broader class of
routing problems.
GIS developed in a very different environment from routing DSS. Coming
from a different user base, GIS software frequently did not have any
routing capability. For example, a GIS might be capable of displaying a
road network on screen, but may not have had a road network data
structure that could be easily integrated with OR/MS algorithms. While
GIS has a recognised role in transport applications (Thill, 2000), many of
these applications do not involve network processing in the form required
for vehicle routing.
Because many potential GIS applications require the use of networks,
later products did incorporate network functionality. For example,
Arc/Info and Arcview incorporate a number of network tools. In earlier
versions of these products, the network functionality was not fully
1 A substantial part of this chapter was published in Keenan (1998a)
Spatial Decision Support Systems for Large Arc Routing Problems
68
integrated with the other GIS features. However, in the current version
(Release 8) networks are a fundamental part of the product. In addition,
an increasing variety third party add-ons are becoming available.
Network utilities in most types of GIS software will provide routines for
the calculation of shortest paths. As shortest paths are the basis of
routing calculations, this functionality is of potential use. Shortest path
calculation can be enhanced by a rich representation of the road network,
with full provision for one ways streets etc. Newer software frequently
has a more comprehensive set of routing procedures. A good example of
this trend is the TransCAD GIS software (Transcad, 1996), which offers a
variety of routing tools. However, these routing tools must inevitably be
general-purpose in nature, while effective decision support requires
specific customised techniques for the problems being supported. In
Section 3.4.2 above, we discussed the use of GIS as a DSS generator for a
routing, with the inclusion of appropriate customised algorithms. GIS
products are increasingly available on inexpensive microcomputers.
These products offer significant GIS functionality at much lower cost
than traditional workstation based software. While some of these PC-
based GIS systems are less powerful than their workstation based
equivalents, the subset of functionality offered is often quite suitable for
routing applications.
4.1.2 Spatial Decision Support Systems for routing
Any examination of the role of SDSS in routing starts from the enhanced
potential for spatial data handling inherent in such systems. Traditional
routing DSS incorporates a relatively restricted range of data structures.
On the other hand, GIS allows the use of point, line and polygon
structures. The ability to store and manipulate these structures is
complemented by sophisticated editing tools allowing easy alteration of
spatial data. GIS provides various tools to allow adjacent points and arcs
be identified or to identify the areas of polygon overlay (Chapter 3).
Appropriate spatial algorithms are provided to support these facilities.
Spatial Decision Support Systems for Large Arc Routing Problems
69
However, in the context of routing problems these only provide interim
data that must be utilised by the routing models. Additional routing
functionality must be added to the basic GIS to build a routing SDSS. We
can therefore distinguish between routing DSS, GIS and SDSS that
incorporate GIS and appropriate routing techniques (Table 4-1).
4.2 Information Requirements for Routing
4.2.1 Categories of routing data
In modelling routing problems, three categories of data can be identified.
Routing problems will contain data associated with locations, for instance
data relating to the depots in the problem and the customers to be
Table 4-1 : Main Characteristics of Routing Systems
(Keenan, 1998a)
Vehicle RoutingDSS
GIS SDSS
Data location co-ordinates,
true-distance matrix,
multiple vehicle
parameters
points, arcs and
polygons, complex
network data
points, arcs and polygons,
complex network data and
multiple vehicle parameters
Models customised multi-
vehicle multi-depot
routing models
general purpose
one vehicle shortest
path type models
customised multi-vehicle
multi-depot routing models
Interface
Representation
representation of
points and routes
representation of
multiple layers of
spatial data
customised subset of
available spatial data
Interface
Operations
ability to edit
parameters and alter
route sequence
map editing
facilities
ability to edit maps, routing
parameters and alter route
sequence
Spatial Decision Support Systems for Large Arc Routing Problems
70
serviced. The second category of data in a routing problem relates to the
vehicles that must visit these locations. Vehicle parameters include speed
and capacity. Finally, the problem will contain a set of paths between
locations. From the paths the distances travelled and the travel time can
be derived. The vehicle routing problem can then be defined as a set of
visits by vehicles to locations along a set of paths between these locations.
Constraints in the problem will exist for locations, paths and vehicles.
Many of these constraints are independent of each other. For example, a
change in the stop time at a location may not directly affect the paths
that can be used. Other constraints are interdependent; they are affected
by the interaction between two types of data. For instance, a particular
type of vehicle may not be able to visit a certain location or use a certain
path.
Traditional views of the routing problem have tended to emphasise the
vehicle and location constraints without paying much attention to path
constraints. This chapter recommends a broader definition of vehicle
Table 4-2 : Constraints in Vehicle Routing Problems (adapted
from Bodin and Golden) (1981)
location constraints time to service a location
number of depots
nature of demands - deterministic or stochastic
location of demands - on points or arcs or polygons
operations - pickups or drop-offs
vehicle constraints number of depots
size of fleet
vehicle capacity constraints
maximum vehicle route times
path constraints underlying network - directed or undirected
time to travel a given network segment
vehicle type limitations on network segments
Spatial Decision Support Systems for Large Arc Routing Problems
71
routing problems to accommodate problems where the path taken is an
important component of the problem. In such a broader context,
information factors can be usefully categorised into location, vehicle and
path data. The main constraints of interest in routing problems are
classified in Table 4-2 (also see Table 5-1).
The traditional optimisation techniques used in routing, for example
travelling salesman algorithms, consider a number of points on an XY
plane. As discussed above in Section 2.1.3, early routing software
frequently used location co-ordinates and straight-line distance was used
as a surrogate for actual travel distance. This abstraction of the problem
assumed that selection of an appropriate path was a trivial exercise. In
practice, the actual route is constrained by the need to use suitable roads.
The abstraction used does not require data on the paths used, although
the use of straight-line distance is unsatisfactory in many cases. Using
distances calculated from the road network, the true-distance approach,
reduced these problems; this approach has become an increasingly
important feature of routing DSS design, especially for arc routing
problems (see Chapter 5). Nevertheless, this approach entails greater
data requirements and a consequent increase in the sophistication of the
software used for the organisation of that data. The incorporation of
additional path data increased the usefulness of the problem formulation,
but at the cost of making the software to solve it more complex.
The traditional true-distance approach calculates travel distances in
advance of the actual routing process. This approach reduces a complex
road network into a relatively straightforward matrix of distances
between locations. However simple distance is not the only issue arising
in many types of real world problem. In practice routing problems are
frequently constrained by time rather than by distance travelled. It is
therefore important to associate appropriate speeds with different
sections of the road network. The speeds attached to the road network
may be derived from the road classification, or from measures such as the
Spatial Decision Support Systems for Large Arc Routing Problems
72
number of traffic lanes. Sophisticated routing models may adjust these
associated speeds to take traffic congestion or the existence of steep
gradients. The traditional representation of distance or travel times as a
simple matrix can accommodate only a few of these extra constraints. For
complex problems, a simple matrix approach may be inadequate as it
operates on the basis that the data does not change during the routing
process. This assumption may represent an unreasonable simplification
of the real world situation.
An alternative is to maintain a comprehensive model of the road network
available for real-time calculation of travel times. In urban applications
additional concerns include one way streets, no right or left turns, and
vehicle size restrictions. In rural areas large vehicles may not be able to
negotiate all the roads in the network due to steep gradients or limited
road width. As greater detail is required, more complex decision support
software becomes increasingly justified. GIS software provides the means
to store and manipulate detailed network representation. However, the
GIS software may not have been designed to be easily used with routing
models (Ralston and Zhu, 1991). Successful SDSS implementation
requires both the availability of appropriate data and software that can
provide the relevant network representions. Detailed road network data
is not easily collected by an individual user and the necessary data may
not be available, for instance the detailed layout of grade separated
junctions (Bodin and Levy, 1994). These issues become especially
important in arc routing problems where a variety of traffic restrictions
must be modelled (Eiselt, Gendreau and Laporte, 1995b). Despite these
problems, the value of GIS is recognised as a means of representing real
world road networks and displaying them on screen. However, this
dissertation suggests that there has not yet been full recognition for the
potential of GIS for modelling new and more sophisticated types of
routing problem.
Spatial Decision Support Systems for Large Arc Routing Problems
73
4.2.2 Interdependence of routing parameters
The data requirements for vehicle routing are greatly increased if the
parameters are interdependent or dependent on some other variable such
as time. For example, if traffic congestion is modelled the appropriate
speed may depend on the time of day. Other time related restrictions
might exist such as restrictions on the use of vehicles in pedestrianised
areas, where deliveries might be required to take place early in the
morning. An important class of problem where complex data interactions
exist are dynamic routing problems. In these problems additional data is
generated, in an unpredictable way, during the routing process (Psaraftis,
1995). An example of such a problem, where the location data can change,
is courier parcel collection and delivery where a request to collect a
package from a new location may arise at any time. Similarly real time
information may be obtained on the paths available for routing if
information is received on traffic congestion or road closures due to
accidents, etc. Bertsimas and Simchi-Levi (1996) discuss the modelling
issues for a number of dynamic routing problems. One example is routing
in situations where orders are received in real time.
In a review of dynamic routing, Psaraftis (1995) notes the growth of GIS
systems and GIS related technology. GPS is seen as an important
development in data collection for routing (Imielinski and Navas, 1999).
GPS allows the collection of point and time data, either as single data
points or as a stream of data. A GPS installed in a moving vehicle can
collect continuous data on the vehicle location and the time that it
reached those locations. This is valuable both for data collection in
advance of route optimisation and for validation of the routes produced
(Table 4-3).
The increasing availability of digital geographic data has prompted the
development of a variety of systems for vehicle guidance, and GPS plays
an important role in these. This class of problem, where real time
information is used, requires a routing system that can work in real- time
Spatial Decision Support Systems for Large Arc Routing Problems
74
with all the detail involved, as pre-processing of data cannot readily be
used.
Complex interactions between vehicle parameters and paths may mean
that fully loaded vehicles cannot use all the segments of a network, due to
steep gradients or because of weight limits on bridges, etc. Therefore, a
situation may arise where a vehicle can use a certain route on its return
journey when empty, but not on its fully laden outward journey. In this
situation, the road network cannot be pre-processed easily into a simple
distance matrix; instead the interactions between vehicles and paths
have to be incorporated in the vehicle routing model. This requires the
use of to appropriate routing algorithms and a sophisticated road network
representation. The use of these more sophisticated models places
increasing demands on a DSS designed to incorporate them. These
requirements arise largely from the need to provide additional geographic
Table 4-3 : GPS applications in routing
Data Type Typicalapplication
Point Data Customer Location
Streaming Data Identification of road segment
length
Pre-Optimisation
(data collection)Time related data Identification of speed per road
segment
Point Data Data revision
New customer data
Streaming Data Route length validation
Post-Optimisation
(routevalidation) Time related data Time window validation
Spatial Decision Support Systems for Large Arc Routing Problems
75
information in the database and interface, to support the use of this
geographic data in more realistic models.
The combined location-routing problem is a second form of real world
problem where the geographic parameters are interdependent. Location
problems are complex in themselves and are an existing area of
integration of modelling techniques and GIS (Ding, Baveja et al., 1994).
Many types of problem require both a location phase and a routing phase;
these can usefully be combined into a more complex model. In such
problems, the total demand or supply in the problem needs to be allocated
to a limited number of locations. Vehicle routes are then generated to
visit those locations. Location-routing problems are very relevant to arc
routing, for example in urban postal delivery (park and loop). This
problem arises where the postman must collect his post from a parked
van and service a series of arcs from that point (Bodin and Levy, 1991).
Traditionally these problems had a number of depots with the objective to
associate customers with depots and generate routes to visit those
customers. In addition to these, however, there are some situations where
the initial phase of the problem is an allocation of demand to delivery
locations. In many real world problems, the actual demand is distributed
over a larger area than the actual location where delivery/collection takes
place. For example in public transport routing, a bus stop on a main
street will service people travelling to a segment of that street and to
segments of neighbouring side-streets. Customers will walk from the area
serviced to the bus stop (Figure 4-1). The location of the bus stops is in
itself an important problem that affects the routes generated. The
location to be visited, the bus stop can be represented as a point.
However, the actual demand is distributed on another type of location, for
example an arc or a polygon.
Spatial Decision Support Systems for Large Arc Routing Problems
76
In such a problem there may be a series of potential feasible locations
only one of which will be used in a given area, this may be modelled as a
generalised travelling salesman problem (Laporte, Asef-Vaziri and
Sriskandarajah, 1996). A complete solution of the location-routing
problem will require the allocation of the total travel demand in a town to
a limited number of service locations which are then routed (Figure 4-1).
An example of this type of application is school bus routing which has
long been of interest in the OR/MS field (Chapleau, Ferland and
Rousseau, 1985). School bus routing has also been seen as an application
of GIS techniques (Cortez, Meek and Koger, 1994; Braca, Bramel, Posner
and Simchi-Levi, 1997).
A related problem might arise with convenience shops that provide
customers in their local area with everyday purchases such as
newspapers or bread. These shop locations can be modelled as a point in a
routing delivery problem. However some routing problems might be
concerned with identifying an appropriate subset of shops to visit,
estimating their potential sales from population data. These types of
Figure 4-1 : Area served from bus stop will include
neighbouring streets.
Spatial Decision Support Systems for Large Arc Routing Problems
77
problems can be approached on a two-phase basis, with the allocations
being processed first and the routing completed as a separate stage.
However, this may lead to unrealistic representations of the problem.
Comprehensive modelling of such a problem requires that a model be able
to evaluate trade-offs between the location and routing phases of the
problem. A decision support system for such modelling must be capable of
representing the interactions between these differing types of location
(Figure 4-2).
The scale of the geographic area of interest in a routing problem may also
differ greatly. Routes may service an area comprising several thousands
of square kilometres. In other cases the area to be covered by the route is
much smaller, for example an urban parcel delivery route. In problems
where only a small area is of interest, the detailed geography of the area
becomes a significant issue. In general, modelling urban applications will
require additional attention to geographic parameters. As greater detail
is required, more complex decision support software becomes increasingly
justified.
Figure 4-2 : The total service area is mapped on to a limited
number of locations.
Spatial Decision Support Systems for Large Arc Routing Problems
78
For some classes of routing problem, for example public transport
scheduling, the data required for the problem may be available from
public sources. For instance, the census of population might provide a
means of estimating demand for a variety of routing problems. Much of
this information is associated with spatial units, such as administrative
districts, rather than directly with the network. This requires that
population associated with arcs or polygons is allocated to points, for
example such as rapid transit stations, which are then visited by routes.
For some problems, routing makes use of data on geographic objects other
than the road network. An example might occur in census enumeration,
where an enumerator wishes to complete visits in one census district
before moving on to the next one. In a postal delivery problem, the postal
addresses may be based on street or district names. In a postal problem,
the quantities of letters to be delivered may be derived from population
data based on these districts. This requires a system that can manipulate
polygon data effectively.
4.2.3 Spatial interactions in routing problems
In representing the vehicle routing problem in a DSS, three classes of
data have been identified, data relating to locations, data relating to
paths and data relating to vehicles. The first two of these data types,
locations and paths, are inherently spatial in nature. GIS has the
appropriate database tools to store and manipulate sophisticated
representations of the locations and paths found in complex routing
problems. Therefore, a DSS incorporating GIS techniques can facilitate
problem solving for more complex real world problems.
We would suggest that GIS based systems are needed where data
associated with locations and paths is complex. This is especially true
where elements of the data are interdependent rather than entirely
independent, for instance where the location or path constraints depend
on time or on the parameters of the vehicles in the problem.
Spatial Decision Support Systems for Large Arc Routing Problems
79
The data requirements for the representation of routing problems are
therefore increased if pre-processing of paths into a simple distance
matrix is not possible. This situation occurs where there are multiple
path constraints or where the path constraints are not constant but are
dependent on time or on changing vehicle or location parameters. Some
well-known routing problems fall into this category, primarily the arc
routing problems that are the subject of this dissertation. For these
problems the network itself is the object of the routing process and a
detailed network representation is therefore required.
Table 4-4 : Types of Location Data
Data Types Type of Problem Example
Basic Types point traditional delivery
problems
delivery from warehouses to
shops
arc arc routing postal delivery
polygon generalised TSP post-box collection
Data Inter-
dependencies
predictable
variation by time
differing volumes at
different times
milk collection
dynamic
generation of
locations
real time data
generation
"dial-a-ride" problem
interaction with
vehicle
parameters
continuous demand,
intermittent vehicle
service
refuse collection - volumes
depend on time since last
collection
interaction with
path parameters
location - routing
problem
public transport routing with
siting of transit stops
Spatial Decision Support Systems for Large Arc Routing Problems
80
Table 4-4 indicates some of the types of location data encountered. Those
routing problems with data interdependencies require additional
complexity in system designed to support. Traditional techniques have
neglected the importance of path constraints, where interdependencies
occur between paths and other aspects of the problem (Table 4-5) a more
complex problem results.
One example of a routing problem with potentially complex interactions
between paths and location parameters is hazardous goods routing, for
example the transport of toxic waste or nuclear materials. Routing for the
transport of hazardous goods may wish to avoid certain areas, such as
environmentally sensitive areas, roads with steep gradients, or areas
with bad weather conditions. Hazardous waste routing problems are
documented both in the GIS literature (Freckmann, 1993) and in the
OR/MS literature (Beroggi, 1994). Erkut (1996) notes the potential
contribution of DSS and the importance of GIS in this field.
Routes generated for hazardous goods might wish to avoid populated
areas, while at the same time remaining within a specified distance of
Table 4-5 : Types of Path Data
Data Types Type of Problem Example
Basic Types planar XY co-ordinate distance ship routing at sea
network constrained true-distance approach urban delivery over
street networks
Data Inter-
dependencies
variation by time different speeds in rush
hour congestion
urban courier delivery
interaction with
location parameters
paths avoid objects ship routing in relation
to islands
interaction with
vehicle parameters
vehicle load restrictions weight limits on bridges
Spatial Decision Support Systems for Large Arc Routing Problems
81
emergency facilities (e.g. fire stations). In such situations vehicles may be
required to travel on network constrained paths and to maintain a
certain straight-line distance between the vehicle and point or polygon
locations. Such vehicles might also wish to remain within a certain road
distance of emergency facilities (Figure 4-3). This problem might be
further complicated by the existence of time constraints that limit the
times at which these hazardous products might be transported, for
example avoiding periods when roads are congested. Hazardous materials
routing may need to respond to adverse weather conditions, such as
snowstorms; DSS can help schedulers make these changes (Beroggi and
Wallace, 1994). A GIS based system could model the path of such a storm,
exploiting real time meteorological data.
An example of interaction between paths and locations is security
patrolling around a sensitive installation, such as an airport. This might
require patrols on roads near to the airport. The precise security risk of a
section of road would be influenced by factors such as its distance from
the installation, which might be the irregular boundary of a large airport.
Location to be avoided
Figure 4-3 : Network constrained route avoiding passing
within a certain distance of a point location.
Spatial Decision Support Systems for Large Arc Routing Problems
82
Other factors such as elevation or sight lines to the installation could be
considered. The actual patrols might take place using road vehicles which
are constrained to use the road network, while the region to be patrolled
is derived from the polygon location of the airport (Figure 4-4). The
routines could provide a measure of the need to patrol each road section
and an appropriate arc routing algorithm could be used to design routes.
A system to support such a routing process would need to be capable of
working with multiple interactions between the data structures in the
problem.
Another important class of routing problems, where these techniques are
very relevant, are emergency evacuation applications. These applications
may be concerned with evacuation from an area close to a fixed location.
In many cases potential disaster situations can be simulated for
evacuation planning purposes, for example around a nuclear power plant
(Hobieka, Kim and Beckwith, 1994).
SensitiveInstallation
Figure 4-4 : Patrol area around irregular boundary of sensitive
installation
Spatial Decision Support Systems for Large Arc Routing Problems
83
Other evacuation situations arise in dynamic conditions; for this type of
problem the design of the routes may be strongly influenced by
geographic features. If a flood or earthquake has taken place, many of the
potential routes may pass through areas rendered unsafe by the disaster.
The extent of the danger may be calculated by the GIS by reference to
geographic data and the suitability of particular road segments can be
derived from this (de Silva, Gatrell, Pidd and Eglese, 1993; de Silva and
Eglese, 2000). For example, routes may be required to allow emergency
services visit buildings in the path of a forest fire. The sequencing of such
routes would be largely determined by the need to visit those in most
danger. This situation could be modelled on a GIS based system, using
data on elevation, type of vegetation, etc. Therefore, effective decision
support for this type of problem might benefit from a combination of GIS
and routing techniques. Such a system would allow interactions between
population data, elevation data and location data be modelled. Recent
work in this area (Patel and Horowitz, 1994) utilised GIS software to
derive an approximate measure of risk at different points in a road
network. This network was then used with OR/MS algorithms to evaluate
the minimum risk path through the network. Although this paper does
not address all of the geographic features of the problem, it indicates the
potential usefulness of a combination of GIS and routing techniques.
However, we are not aware of any existing examples of a comprehensive
combination of GIS and OR/MS techniques in a SDSS for this type of
problem.
4.3 The Role of GIS in Supporting Routing Problems
4.3.1 Routing problems supported by traditional DSS
The contribution that GIS techniques can make to a routing DSS will be
small for problems that have no great spatial content. This is especially
true if they are largely concerned with internal data and therefore have
few geographic parameters. Such problems might include many of the
delivery problems addressed in the vehicle routing literature. In such
Spatial Decision Support Systems for Large Arc Routing Problems
84
problems, the schedules may be tightly constrained by non-geographic
elements such as time windows. Some type of GIS software might be used
to provide an attractive on-screen mapping facility, but the database and
spatial query capabilities of the GIS will have little role to play.
One example of such a problem is that of fuel delivery. A variety of
problems are documented in the literature where a multi-compartment
tanker is used to service customer orders of different sizes. The vehicle
may be required to carry a number of different types of fuel such as
different octane grades, unleaded or leaded, heating fuel, or diesel. In
some countries differences exist between fuels for tax reasons, for
example in Ireland diesel for agricultural use is taxed less than that for
road vehicles. A fuel tanker can typically service only a small number of
orders per trip, as few as three or four deliveries per trip. The vehicles
used for fuel delivery will typically have a number of compartments;
orders will have to be allocated to these compartments without mixing
different types of fuel. As different fuel types cannot be mixed, and as
there are only a small number of points to be delivered, the efficient
packing of the vehicle is the most important consideration. For this type
of problem a graphic interface is hardly needed, the quality of solution
being largely determined by the algorithm used. This group of problems
has long been of concern to researchers in OR/MS (Önal, Jaramillo and
Mazzocco, 1996).
For problems that have few geographic parameters, a graphic interface is
clearly useful if there are a large number of delivery/collection points.
However such an interface need not include a great deal of geographic
information and need not be as complex as a GIS. Problems that fall into
this category include the large number of routing problems that are
concerned with delivery of parcels, supplies, etc. to a set of specific
customers.
The spatial complexity of such problems is largely determined by the
number of customers to be delivered, as geographic features other than
Spatial Decision Support Systems for Large Arc Routing Problems
85
the location of customers are not relevant. The complex nature of these
problems means that routing models can be of assistance. However, the
large number of routing possibilities increases the difficulty of the
problems, for example large travelling salesmen problems. Therefore,
suitable algorithms and a facility to modify the routes using a graphic
interface are required. However, the visual component of the interface
and the database component need not include a large number of
geographic features. Commonly used interactive vehicle scheduling
software is well suited to solving this group of problems.
4.3.2 Routing problems requiring GIS support
Routing for standard delivery problems in geographically compact
regions, such as urban areas, requires additional attention to geographic
features. In this case, one-way streets, no left or right turns, traffic
congestion, etc., will have an important part to play. These additional
requirements indicate that a more sophisticated support system is
appropriate. Such a system requires the ability to store and display the
level of detail appropriate to the problem. Therefore, while many routing
problems are straightforward on a regional scale, at a detailed urban
scale more information is needed and therefore GIS based techniques will
be more useful.
For some types of routing problems, additional geographic parameters are
introduced by the fact that the relevant data for the problem is largely
external. Relevant external data will usually include population data,
including socio-economic data for that population. For instance, a
marketing research project may require researchers to visit a number of
locations. The route generated will largely be determined by non-spatial
considerations such as structuring the age or socio-economic groups in the
sample. For such problems a traditional GIS may be useful, but
sophisticated routing techniques are probably not needed. Problems of
this type are frequently addressed by people with GIS expertise, with
little explicit use of operations research techniques. There is an
Spatial Decision Support Systems for Large Arc Routing Problems
86
increasing trend for network analysis tools, for example shortest path
algorithms, to be included in GIS software. These tools allow useful
analysis but are not complete routing algorithms.
For problems that are spatially complex and have a number of geographic
entities involved, this dissertation suggests that the use of a SDSS is
appropriate. Such a SDSS would combine appropriate algorithms with
that subset of GIS data that is pertinent to the problem. Population data
is relevant to most problems for which a SDSS is appropriate. However,
other information found in a GIS may be relevant, including the existence
of geographic features such as mountains, lakes and rivers. Complex
interactions between these features and the road network may have to be
modelled for some types of routing problem. The potential contribution of
GIS software is great because of the facilities for database interactions
between different geographic features found in this software. This
dissertation suggests that a SDSS with both GIS techniques and
sophisticated vehicle routing models is needed for this type of problem.
4.3.3 Spatially complex routing problems
SDSS based systems will provide an important contribution in any
situation where routing, and the road networks used for routing, needs to
be related to other geographic features. Tourist routing may aim to
design routes with the maximum scenic value or with a desired mixture
of sights (van der Knapp, 1993). These routes would require interaction of
a routing algorithm with data on elevation, type of vegetation, location of
rivers and lakes, etc.
An example of routing applications with potentially complex interactions
between spatial parameters might exist in the military field. Military
applications typically make use of off-road vehicles. While such vehicles
are in principle capable of straight-line travel over all terrain, the actual
ground conditions greatly affect the speed at which such vehicles could
travel. Military applications might involve a desire for vehicles to travel
Spatial Decision Support Systems for Large Arc Routing Problems
87
on high ground to gain a commanding position or to travel below the
horizon to facilitate concealment (Rasmussen, 1997). A GIS based system
with the ability to handle elevation data could provide the necessary
routing support environment for such an application.
Agriculture and forestry provide examples where GIS and routing
techniques might usefully be combined; Martell (1998) notes the
relevance of GIS to the use of OR/MS in forestry. Agricultural routing
problems provide an example where complex interaction may take place
between the routes and various geographic features. A variety of
agricultural and forestry routing applications might use off-road vehicles.
These vehicles might be capable of low speed travel across fields, and
higher speed travel on roads. The routes provided would have to optimise
the point at which the off-road vehicle rejoined the road. A GIS could
calculate the quantities of product to be transported by reference to
measures such as crop yield per hectare. A GIS based system could model
travel speeds for such vehicles taking into account gradients, various
types of ground conditions and the existence of obstacles. The ground
conditions or gradient might influence the type of crop planted, which
would determine the quantity of product to be removed from the fields.
The type of crop planted might alter ground conditions sufficiently to
significantly change the speeds of the vehicles, for example if the ground
was ploughed. Therefore routing for this type of problem would involve
many geographic parameters and complex interactions between them.
Such a problem could only be approached by an effective synthesis of
routing algorithms, and a DSS based on GIS techniques. As far as we are
aware, there are no examples of the use of GIS techniques and routing
algorithms to provide comprehensive decision support for a problem of
this complexity.
Spatial Decision Support Systems for Large Arc Routing Problems
88
4.4 Implementing routing SDSS
4.4.1 Data requirements for routing system implementation
The starting point for a system builder considering the relevance of SDSS
techniques to routing DSS is to evaluate three factors. Firstly, the nature
of the input data used in the problem, to what degree is this spatial in
nature? Then consider the spatial content of the processing operations
required and the nature of the output data. Any problem may be
considered in terms of the location, vehicle and path constraints
identified above (Table 4-2). A problem dominated by vehicle and point
location constraints, e.g. loading restrictions, time windows, may not
need spatial techniques. However, the existence of arc or polygon-based
locations, or complex path constraints, indicates a need for GIS
techniques. The routing algorithms used will be largely determined by
the nature of the locations and the vehicle constraints in the problem.
The designer of a SDSS for vehicle routing faces problems reconciling the
different traditional approaches to data between that found in GIS and
that found in the routing software. In a routing SDSS, sections of the
road network will have a traditional role as components of a graph; this
representation will be used for algorithmic purposes. For spatial
processing purposes, the road network must also be seen in relation to
other spatial features. For example, administrative boundaries and
contours might exist independently of the road network. However, the
GIS environment represents the use of road networks from a very
different point of view than does traditional OR/MS software. Therefore,
one obstacle to the development of a SDSS is that GIS software does not
always characterise road networks in a way that facilitates their use by
routing algorithms (Ralston and Zhu, 1991).
As GIS use increases, comprehensive integrated databases of road
networks will become available, for instance the databases for many
countries provided by Navtech (1999). In Ireland road network data is
available from both public sector and private sector sources (Gamma;
Spatial Decision Support Systems for Large Arc Routing Problems
89
IRIS). These can be used in a variety of routing and vehicle navigation
applications (Cova and Goodchild, 1994). These databases will contain
data from different sources and will require some reorganisation to be
entirely suitable for routing programs. The interest in automated car
guidance systems has led to research into the more realistic modelling
road networks (Newcomb and Medan, 1993). Other researchers (Portier,
Berthet and Moreno, 1994) have examined the use of shortest path
algorithms on complex road network representation.
4.4.2 Using GIS data in routing DSS
Decision-making in a DSS is a combination of modelling and user
judgement, requiring the SDSS system builder to identify the role played
by spatial data for both the models and the users. Traditional routing
algorithms do not require much of the data stored in a GIS. Decision
support might be better enhanced by the provision of additional
geographic data for use directly by the user, rather than the introduction
of ever more complex mathematical techniques. The simplest and most
common integration of geographic data and routing is to provide
additional information to the decision-maker, without necessarily
incorporating that data in models used. For example, a map might be
displayed on screen, providing the decision-maker with information about
features that are not explicitly used by the models. This map could
display the routes generated against a background of natural features
such as rivers etc., which help orientate the decision-maker.
This type of system may be largely designed as a traditional routing DSS
with no direct interaction with the GIS database. A routing sequence is
produced as a succession of points in the GIS and these can be displayed
on screen. The only user interface feature required is likely to be the
querying of routes and the insertion and deletion of customer locations.
This functionality can be added by using the graphics features in modern
programming languages (Fölsz, Mészáros and Rapcsák, 1995). However,
it is likely to be more efficient to use existing software for the display of
Spatial Decision Support Systems for Large Arc Routing Problems
90
maps (Tracey and Dror, 1997). For the situation where spatial features
are not used by the models a mapping toolbox, rather than a
comprehensive GIS, is likely to be the basis of such a DSS. Systems of
this level of complexity are likely to become standard in the future, as
they will take advantage of improved computer graphics and the
availability of additional mapping resources. However, such systems
would not constitute a fully-fledged SDSS.
The presence of complex path constraints requires that the models
require input data from the GIS database. This requires a degree of
integration sufficient to allow the GIS database provide the data for the
routing problem, in particular the distance data. Such integration would
be greatly facilitated by the use of GIS software with the software
features to support the use of route structures or distance matrices.
Appropriate GIS software would allow such issues as one way streets and
turn restrictions be modelled. In many cases spatial data is needed at the
input and output stage but no spatial processing is required. This could
be achieved by a three-phase operation. Initially data might be extracted
from the GIS to form a distance matrix. The routing algorithms could
then solve the problem without reference to the GIS, and the completed
routes could be displayed on a map on the screen.
If a static set of points is involved then the operational routing system
might not require a comprehensive GIS at all. Instead, a GIS could be
used to build the distance matrix, which could then be used by a system
not directly connected to the GIS. A mapping program, rather than a
comprehensive GIS, could then display the proposed routes. If the system
is built within a GIS, the modelling programs only require access to the
database, perhaps through technologies such as open database
connectivity (ODBC). In this case, the modelling routines could be run as
external programs from within the GIS, with both the GIS and the
models sharing the database.
Spatial Decision Support Systems for Large Arc Routing Problems
91
Where there is a dynamic component to the path or location data, or there
are complex interactions with vehicle data a true SDSS is needed. This
would require that that the routing models make use of spatial
processing, requiring interaction with the spatial data handling features
of the GIS. This would mean that the routing algorithms would
dynamically seek complex data sets from the database, requiring the use
of GIS operations to build this data. This would require that the
modelling routines dynamically call spatial data handling routines within
the GIS. Therefore, the software techniques used must be capable of
accommodating this form of connection between the programs. Table 4-6
indicates the type of operations that must be implemented in a SDSS for
an agricultural routing problem requiring SDSS techniques.
SDSS applications entail a greater diversity of modelling approaches
than traditional vehicle routing applications. Therefore, SDSS routing
Table 4-6 : Example of an agricultural routing SDSS
Objective Actions Required SDSS technique
identify collection
points
identify convenient locations
on roads for collections
associate polygon volumes with arcs
on road network
build routing problem identify quantities to be
transported
calculate polygon areas in database
and derive volumes from these
build road network generate distance/ travel time
data for use by mathematical
procedures
standard GIS operation called
within macro
establish routing
parameters
enter speed, capacity etc. user intervention using form
interface
route vehicles apply routing algorithmf customised routing model calling
SDSS routines as appropriate
refine solution interactive user modification of
routes
Display routes in GIS interface,
user alterations lead to dynamic
recalculation of volumes
Spatial Decision Support Systems for Large Arc Routing Problems
92
software will employ classic techniques, such as the travelling salesman
and linear programming approaches, in conjunction with techniques
drawn from other fields. These fields might include risk analysis, which is
especially relevant to hazardous materials transportation (List,
Mirchandani, Turnquist and Zografos, 1991). Modelling approaches
drawn from the GIS environment will need to be used in addition to the
traditional techniques used in the routing domain. While dealing with
broadly the same set of data these different approaches may have quite
different problem representations, presenting the system builder with
significant problems. The construction of a SDSS will require an effective
synthesis of these diverse techniques and the data needed to support
them.
4.5 The future of GIS and routing
4.5.1 Types of Routing Software
This chapter has looked at routing problems in terms of the geographic
complexity of the problem and has suggested that a richer set of routing
problems can be modelled by incorporating more geographic data in
routing problem formulation. This diverse range of routing problems will
require different software tools for decision support. Those problems that
have few geographic parameters can be supported using traditional
decision support system software for vehicle routing. These traditional
problems will typically involve only point and arc data, without the use of
polygon data, and will have few path constraints. Those problems that
have complex geographic interactions require a SDSS for maximum
decision support.
Table 4-7 outlines the main features of problems that can be supported
using the different types of software available. In general, traditional
interactive DSS applications will have greatest application where the
number of spatial interactions is high but few different types of
geographic data are involved. A GIS based system will be of greater use
Spatial Decision Support Systems for Large Arc Routing Problems
93
for problems that have a larger number of geographic parameters.
Problems with many geographic parameters may be approached using
conventional GIS tools, if there are not a large number of spatial
interactions taking place. Where problems require the use of vehicle
routing models, then SDSS functionality is needed. The growth of GIS
based decision support for routing will be enhanced by the growing
interest in arc based routing problems that tend to involve more
geographic data.
4.5.2 The use of GIS for routing
GIS is increasingly being seen as a technology relevant to routing. Recent
surveys of routing software (Hall and Partyka, 1997; Partyka and Hall,
2000), in a practitioner orientated publication, included information on
the ability of the software to interact with GIS. Most of the routing
software packages had some ability to interact with GIS, the most
popular GIS tools being Arcview and MapInfo. However, it seems likely
that the degree of interaction between the routing software and the GIS
functionality is currently rather less than the maximum potential
identified above.
Table 4-7 : Support Requirements of Routing Problems
Problem Characteristics Problem Examples Software Required
Model
few geographic parameters,straightforward paths, smallnumber of locations pervehicle
problems mainlyconstrained by non-spatial factors e.g.loading constraints
numerical optimisationsoftware, little need forspatial interface
Intensive few geographic parameters,many potentialcombinations of vehicle andlocation.
delivery problemsover regional areas
interactive routing
Data
many geographicalparameters, complex paths,few combinations of vehicleand location
path finding problems,public transit routing
GIS
Intensive multiple interactionsbetween path, vehicle andlocation constraints
urban routingproblems, arc routing
SDSS
Spatial Decision Support Systems for Large Arc Routing Problems
94
In the academic sphere, in recent years, a number of GIS based
applications have been described at OR/MS conferences and in OR/MS
journals. These applications typically emphasise the value of GIS as a
means of providing a visual map display on screen and as a means of
storing data. For example, in a health care worker routing application
(Begur, Miller and Weaver, 1997), the comprehensive restructuring of a
Proctor and Gamble’s logistics (Camm, Chorman et al., 1997) or the
technician routing application at Sears (Weigel and Cao, 1999). In
general, this class of applications does not appear to exploit the
opportunities for spatial processing and richer path modelling outlined in
this chapter.
This chapter has identified the role of GIS techniques in modelling a
richer set of routing problems. This dissertation has looked at a broad
spectrum of routing problems with respect to three types of constraint;
locations, paths and vehicles. This dissertation has suggested that the
first two of these are inherently spatial in nature, and that path
restrictions have been given less attention in traditional routing
applications. The arc routing problems that are discussed in this
dissertation must, by definition, consider paths. This section has
identified some of the interactions that can take place between these
different types of spatial parameters. The incorporation of routing
techniques into GIS would allow the building of a SDSS. This dissertation
has identified the class of problems where we believe that a SDSS may
contribute. Such a system would incorporate elements of a GIS with
appropriate OR/MS techniques.
Existing work in the GIS and OR/MS fields have concentrated on
different aspects of the routing problem. OR/MS researchers have
developed sophisticated algorithms to deal with various vehicle and
location constraints while paying less attention to path constraints. GIS
researchers have developed techniques to represent different types of
location and networks and to generate appropriate paths through these
Spatial Decision Support Systems for Large Arc Routing Problems
95
networks. This dissertation suggests that a well-integrated combination
of GIS and OR/MS techniques would facilitate decision support for
problems with complex path restrictions and multiple vehicles. This
includes arc routing problems, including the large sparse network
problem, discussed in Chapter 7, that forms the primary focus of this
dissertation.
Spatial Decision Support Systems for Large Arc Routing Problems
96
Chapter 5 : Arc Routing Problems
5.1 Routing Problems
5.1.1 Background
Routing problem definitions include reference to the types of vehicles
used, the types of products carried and the types of road networks on
which the vehicles must travel. In OR/MS the mathematical description
of road networks is drawn from graph theory and routing problems form a
subset of graph theory problems (Evans and Minieka, 1992, Ch. 8, 9 ). An
undirected graph (simply graph) G is an ordered pair (V; E) consisting a
non-empty set of V vertices (nodes) and E edges (arcs) linking them. Arcs
are of the form (i, j) from vertex i to vertex j. The number of arcs from a
vertex defines the degree of a node. We refer to a node as odd or even if its
degree is odd or even. A path exists between two nodes u and v if a
sequence of arcs exists, possibly through other nodes, which would allow
a vehicle pass from u to v.
5.1.2 Types of routing problem
Routing problems come in a number of forms, a widely used classification
is that of Bodin and Golden (1981) (Table 5-1). In its basic form, a routing
problem requires the visiting of a sequence of locations with minimum
distance travelled. Where the locations to be visited are points (nodes),
this is a Node Routing Problem, the best-known example of which is the
Travelling Salesman Problem (TSP). This class of problem has been
intensively researched.
Spatial Decision Support Systems for Large Arc Routing Problems
97
Table 5-1 : Classification in Vehicle Routing and Scheduling
(Bodin and Golden, 1981).
A. time to service a particular node1. time specified and fixed in advance2. time windows3. time unspecified
B. number of domiciles 1. one domicile 2. more than one domicileC. size of vehicle fleet available 1. one vehicle 2. more than one vehicleD. type of fleet available 1. homogeneous case 2. heterogeneous caseE. nature of demand 1. deterministic 2. stochasticF. location of demands 1. at nodes (not necessarily all) 2. on arcs (not necessarily all) 3. mixedG. underlying network 1. undirected 2. directed 3. mixedH. vehicle capacity constraints 1. imposed - all the same 2. imposed - not all the same 3. not imposedI. maximum vehicle route-times 1. imposed - all the same 2. imposed - not all the same 3. not imposedJ. costs 1. variable or routing costs 2. fixed operating or vehicle acquisition costsK. operations 1. pickup only 2. drop offs only 3. mixedL. objective 1. minimise routing costs incurred 2. minimise sum of fixed and variable costs 3. minimise number of vehicles required
Spatial Decision Support Systems for Large Arc Routing Problems
98
The problem of traversing all road segments (arcs or links) in a network
while minimising the total distance travelled is an Arc Routing Problem,
known as the Chinese Postman Problem (CPP). The General Routing
Problem (GRP) is a generalisation in which the TSP and CPP are special
cases (Lenstra and Rinnooy Kan, 1976). In the GRP, we require a
minimum cost cycle that visits a set of required nodes and traverses a set
of required arcs. Research continues in the GRP to identify a unifying
framework for routing problems (Letchford, 1999; Ghiani and Improta,
2000).
In practice, the vehicles used have a limited capacity in terms of the
volume carried or the maximum time of the journey, therefore most
routing problems are said to be capacitated. A capacitated node routing
problem is usually known as the Vehicle Routing Problem (VRP). This
type of problem has been very extensively researched, as can be seen in
Laporte and Osman’s survey review (1995). Where arcs rather than nodes
must be visited, we have the Capacitated Arc Routing Problem (CARP).
This is a much less thoroughly researched field than that of node based
problems. However, many practical problems can be best represented by
CARP rather than VRP formulations. Obvious examples include postal
delivery, refuse collection, and snow removal. Other examples include
meter reading, inspection of electrical cables, and distribution of animal
feed. In 1995, two surveys of the arc routing field appeared which provide
many examples of applications. One a review of postman problems
(Eiselt, Gendreau and Laporte, 1995a; Eiselt, Gendreau et al., 1995b), the
other (Assad and Golden, 1995) a comprehensive review of the arc routing
field. The latter provided a good overview of routing problems generally.
5.1.3 Problem complexity
Combinatorial optimisation problems come in two varieties; those which
can be solved in time bounded by a polynomial in the input length, and
those for which all known algorithms require time which, in the worst
case, is exponential in the input length. Most exponential time
Spatial Decision Support Systems for Large Arc Routing Problems
99
algorithms are merely variations on exhaustive search, whereas
polynomial time algorithms generally exploit some deeper insight into the
structure of the problem. Exponential time algorithms are generally not
regarded as "good" because of the speed with which the computation
times rise as the size of the problem increases. A problem is then said to
be intractable if no polynomial time algorithm can be found for it.
In assessing the solution time of algorithms, those that can be solved in
worst-case polynomial time are known as P-Problems (e.g. an Euler tour).
If the input is of size n, the running time must be O(nk). Note that k can
depend on the problem class, but not the particular instance. The
problems in complexity class P are called tractable. The class of decision
problem that has solutions than can be verified in polynomial time on a
non-deterministic computer is known as NP. These problems form a
subset of the general class of combinatorial problems. It has been shown
that a large number of problems have this property of being the "hardest"
member of NP. These problems are known as NP complete. Garey and
Johnson (1979) list over 300 problems in this class. Routing problems are
generally NP complete (Lenstra and Rinnooy Kan, 1981). This implies
that the number of computations required to solve a problem grows
exponentially with a parameter of the problem, for example the number
of nodes or arcs in a network. This makes it unlikely that an algorithm
can be devised which is guaranteed to give the answer in a time that is
polynomial in the size of the problem.
While a general algorithm cannot be devised to solve the full range of NP
complete problems, special cases may be solved successfully. A variety of
stratagems has been used to devise algorithms that solve special cases of
these problems. These specialised approaches have facilitated the
solution of some particular large TSP problems to optimality. Capacitated
problems are much more difficult to solve than the classical TSP. For
many practical routing problems, only heuristic solutions can currently
be used. These provide good, but not optimal, solutions.
Spatial Decision Support Systems for Large Arc Routing Problems
100
5.2 Arc Routing Problems
5.2.1 The Chinese Postman Problem
The problem of visiting a sequence of arcs has long been of practical
interest and was the subject of much early study. The famous
mathematician Leonhard Euler (1707-1783) first addressed the problem.
Euler was the most prolific mathematical writer ever, finding time (as
well as having thirteen children) to publish over eight hundred papers in
his lifetime. During his travels between Russia and Berlin, Euler visited
the Prussian city of Königsberg (since the Second World War the city is
called Kaliningrad and is part of Russia). The city was divided by the
river Pregel into four separate parts (Figure 5-1) linked by seven bridges.
Euler became interested in the local challenge of crossing each of the
bridges exactly once and returning to the place you started from. Euler
identified that it was only possible to cross a network of bridges if all the
landmasses have an even number of bridges or exactly two landmasses
have an odd number of bridges. As this was not the case in Königsberg,
Euler showed that it was not possible to cross each bridge exactly once.
A
B
C
D
Figure 5-1 : The Königsberg bridge problem (Euler, 1736)
Spatial Decision Support Systems for Large Arc Routing Problems
101
The CPP is so called because of work by the Chinese mathematician
Kwan Mei-Ko, who worked for the postal service during the cultural
revolution. On his return to academia, Kwan wrote a paper on the
problem (Kwan, 1962). The CPP is simply stated as follows: “A postman
has to cover his assigned segments before returning to the post office. The
problem is to find the shortest walking distance for the postman”.
Formally this class of problem is defined on a graph G=(V,A) where V is
the vertex set and A is the arc set. Associated with the graph G is a non
negative cost matrix Cij, giving the weight of the arc from vertex i to
vertex j. If each node in the network has an even number of arcs (nodes
are of even degree), then a postman tour can be directly generated. If this
is not the case, and some of the nodes have an odd number of incident
arcs, then there must be at least one additional traversal of an arc
already visited. These redundant arc traversals serve to make each node
of even degree (Figure 5-2) in the enhanced graph G′. By definition, the
CPP visits all arcs at least once, the object of any solution procedure is to
minimise any additional traversals. Kwan proved that a necessary and
sufficient condition for the optimality of a Eulerian tour on G′, is that not
more than two edges link any vertex pair, and that the length of the
added edges on every cycle does not exceed half the length of the cycle.
A solution to the CPP may be found by adding additional traversals
between the nodes of odd degree. Edmonds (1965) recognised that the
A B
C D
E F
G H
A B
C D
E F
G H
Figure 5-2 : Graph with four odd points (C,D,E,F) and addition of
redundant arcs to make all nodes even (Kwan, 1962)
Spatial Decision Support Systems for Large Arc Routing Problems
102
most efficient set of traversals, as defined by the distances in Cij , may be
identified by solving a Minimum Cost Perfect Matching (MCPM) problem.
Efficient MCPM algorithms exist which can solve the problem in
polynomial time, the best-known approach is the algorithm by Edmonds
and Johnson (1973). Further research has taken place into near optimal
techniques (Avis, 1983) and faster optimal implementations (Cook and
Rohe, 1999). Consequently, a CPP can be solved optimally by using a
matching algorithm to add the minimum set of extra traversals required
making each node of even degree. When this has been achieved, an Euler
tour can be formed on the expanded graph. That provides a tour that
visits each original arc or matching arc exactly once. On a given matched
network there may be many Euler tours, allowing alternative routes that
have a different sequence but all still be optimal.
5.2.2 The Rural Postman Problem
The Rural Postman Problem (RPP) is an extension of the CPP where only
a subset of arcs (edges) from the network are to be traversed (Eiselt,
Gendreau et al., 1995b). This problem is so called because it might arise
in rural areas where not every road is inhabited. The RPP was first
introduced by Orloff (1974) and has been the subject of limited research
since then. The RPP has been shown to be NP-complete if the required
edges are not connected, but instead form a set of disconnected
components. Some work has taken place on optimal approaches
(Corberán and Sanchis, 1994) but given the difficulty of the problem
practical solution techniques have been heuristic based. Many employ a
procedure to connect the network, for instance by using a minimum
spanning tree algorithm. When a connected network has been derived, a
CPP procedure is used to derive a postman (Euler) tour (Pearn and Wu,
1995). Another heuristic approach used a Monte Carlo simulation
approach to generate multiple solutions for the RPP (Fernandez de
Córdoba, Garcia Raffi and Sanchis, 1998), the best of the solutions is then
used. Specialised variations of the RPP exist, for example the RPP with
Spatial Decision Support Systems for Large Arc Routing Problems
103
deadline classes (Letchford and Eglese, 1998) or with turn penalties at
intersections in the network (Benavent and Soler, 1999; Clossey, Laporte
and Soriano, 2001).
5.2.3 Other Uncapacitated Arc Problems
The basic CPP is defined on an undirected network, however a different
situation arises where some or all of the arcs in the network are directed
arcs. A variety of problems are discussed in (Assad and Golden, 1995;
Eiselt, Gendreau et al., 1995a; Dror, 2000). For the directed CPP a
solution can be found if a directed path exists between every pair of nodes
(Edmonds and Johnson, 1973), see Beltrami and Bodin (1974) for a
practical example. The Windy Postman Problem (WPP) arises where the
travel time in one direction on an arc is not the same as the other. An
analysis of this problem is given in Win (1988), other recent work
includes that by Pearn and Li (1994). The Mixed Chinese Postman
Problem is a NP-hard problem, unlike the undirected and directed cases
(Minieka, 1979; Pearn and Chou, 1999). Both heuristic and optimal
branch and cut procedures have been employed to achieve solutions for
this class of problem (Hong and Thompson, 1998). These problems are
comprehensively reviewed in Eiselt, Gendreau and Laporte (1995a). The
Maximum Benefit Postman Problem (Malandraki and Daskin, 1993) aims
to find a tour of maximum net benefit where a benefit is realised each
time an arc is traversed. In such a tour, not every arc will be traversed,
while some arcs may be traversed more than once. An example of this
class of problem is snow ploughing, where multiple passes on a street are
preferable to allowing large amounts of snow to accumulate and then
trying to clear it in one pass.
Spatial Decision Support Systems for Large Arc Routing Problems
104
5.3 The Capacitated Arc Routing Problem (CARP)
5.3.1 Definition of CARP
While an optimal solution exists for the basic CPP, real world problems
require more than one vehicle to visit a given set of arcs. The Capacitated
Chinese Postman (CCPP) is an extension of the CPP to a situation where
a number of vehicles of limited capacity are used to service the arc
network. Each arc in the graph G=(V,A) has a non-negative weight qij and
each vehicle a capacity of W. The restriction on vehicle capacity may be
distance travelled, duration of route or quantity of goods carried.
Therefore the CPP can be regarded as a subset of the CCPP where W >
ΣiΣj qij. The CCPP is a subset of the broader set of CARP formulations.
The CCPP implies that all arcs in the network will be visited. Just as in
the CCP, additional traversals will be needed where nodes of odd degree
exist. However, in the capacitated problem a further set of additional
traversals exists where more than one vehicle travels along the same arc.
If these can be minimised or eliminated then a good solution for the
CCPP will exist. In practical problems some cases arise where more than
one vehicle must use the same arc. In this case, one vehicle will service
the arc while subsequent traversals by that vehicle or other vehicles
merely pass through the arc without servicing it.
The most general formulation of CARP includes the Capacitated Rural
Postman Problem (CRPP), which arises where not all arcs in the problem
are demand arcs. In formulating CARP as a linear program (LP), two
approaches may be used. A complete graph may be assumed or a sparse
graph formulation may be used, the latter approach possibly reducing the
size of the problem. While a number of different approaches have been
examined, the complexity of CARP means that optimal LP solutions are
extremely difficult to achieve.
The presence of capacity constraints in routing problems implies that the
problem solution has two components. Firstly the question of what
Spatial Decision Support Systems for Large Arc Routing Problems
105
vehicle visits each arc or node; this is the allocation phase. The second
issue is the sequence within each route, the sequencing phase. These
phases can be attempted separately or together. In node routing problems
a variety of clustering procedures exist to solve the allocation phase, a
sequencing algorithm like the TSP may be then be used to derive the
complete route. A variety of approaches to finding a CARP solution exist,
and these are discussed below. One approach is to transform the problem
into a node based formulation, this is discussed in the next section.
Linear programming approaches are introduced in Section 5.4 and lower
bounding techniques are discussed in the remainder of this chapter. In
Chapter 6 a variety of heuristic approaches to CARP are discussed.
5.3.2 Representing Arc Routing problems as a TSP
Both node routing problems and arc routing problems can be seen as
subsets of the GRP introduced by Orloff (1974). Therefore, it is possible to
formulate arc routing problems as VRPs and vice-versa. The generalised
TSP is an extension of the TSP where the objective is to visit several
clusters of vertices. Laporte, Asef-Vaziri and Sriskandarajah (1996) show
that arc routing problems, including the CPP and RPP, can be modelled
as generalised TSPs. Jansen (1993) discusses the General Capacitated
Routing Problem, which includes both VRP and CARP.
Pearn, Assad and Golden (1987) examine the relationship between node
routing problems and CARP. They describe a transformation of an arc
routing problem where each arc is replaced by three nodes, called side
and middle nodes. These are spaced equally along the arc (Figure 5-3).
This transformation establishes the equivalence of arc and node routing
problems. However in practical terms the node transformation is not
sij sij mij
Figure 5-3 : Introducing new nodes for each original arc (Pearn,
Assad and Golden, 1987)
Spatial Decision Support Systems for Large Arc Routing Problems
106
likely to be significantly easier to work with than the original CARP
formulation.
The classic CPP is defined on an undirected network, other variations of
arc routing problems are defined on directed networks. The modelling of
real world routing problems utilises a number of modifications of these
basic categories. One way streets can be characterised as directed arcs.
Problems that require both side of the road to be serviced can be
represented as two directed arcs in opposite directions. However, for
problems such as street sweeping a one way street may be serviced by a
vehicle travelling against the flow of traffic. Indeed such an arrangement
may be required if the kerb side sweeping equipment is only fitted on one
side of the vehicle. In this case, service arcs exist in both directions but
extra traversals can take place in one direction only. One example of this
situation occurs in the rural road-sweeping project discussed by Eglese
and Murdock (1991). In this case, where both sides of the road are to be
visited, an Euler tour can be easily formed as all nodes are of even
degree. However, a variety of service level issues may influence the tour
actually produced and significantly complicate the solution procedure.
5.4 Linear programming formulations of CARP
5.4.1 Golden and Wong formulation
A mathematical programming formulation for CARP on a complete graph
was introduced by Golden and Wong (1981)). This provided an integer
programming (IP) formulation of the problem. The objective function (5-1)
seeks to minimise total distance travelled. Constraints (5-2) ensure route
continuity. Constraints (5-3) state that each arc with positive demand is
serviced exactly once. For each arc (i,j) the constraints represented by
(5-4) ensure that it can be serviced by postman p only if he visits arc (i,j).
Constraint (5-5) refers to vehicle capacity and the constraints in (5-6)
prohibit the formation of illegal subtours. Integrality constraints are
given in (5-7).
Spatial Decision Support Systems for Large Arc Routing Problems
107
IP Formulation of CARP (Golden and Wong (1981))
Minimise pij
K
pij
n
j
n
ixc∑∑∑
=== 111
(5-1)
Subject to0
11=− ∑∑
==
n
k
pik
n
k
pki xx for i = 1,….,n
p= 1,….,K
(5-2)
1)(1
=+∑=
pji
K
p
pij ll for (i,j) ∈ E
(5-3)
pij
pij lx ≥
for (i,j) ∈ E
p= 1,….,K
(5-4)
Wql ij
n
j
pij
n
i≤∑∑
== 11
for p = 1,….,K (5-5)
(5-6)
{ }{ }
∈
∈≤+
≥−
−≤−
∑∑
∑∑
∉∈
∈∈
1,0
1,0,;1
1
1~
,
~2~1~2~1
~2~~
~12
~~
pij
pij
pq
pq
pq
pq
pq
Qi
pij
Qi
pq
Qi
pij
Qi
lx
yyyy
yx
Qynx
},....,3,2{of~ofsubset
emptynoneveryand12,...,1~
,...,1for1
nQ
qKp
n −==
−
(5-7)
where n = the number of nodes in the network
K = the number of available postmen or vehicles
W = the postman capacity (W ≥ maximum of qij
j)(i, arc of demand the
j)(i, arc oflength the
=
=
ij
ij
qc
ppostman by serviced is j)(i, arcif 1 otherwise 0
ppostman by traversedis j)(i, arcif 1 otherwise 0
==
==
pij
pij
l
x
E = the set of all edges on the network
Spatial Decision Support Systems for Large Arc Routing Problems
108
Linear programming (LP) formulations of arc routing problems, in
common with LP formations of travelling salesman and vehicle routing
problems, must ensure that illegal subtours are not formed. In the CARP
example a series of cycles could be formed which do not visit the depot
(Figure 5-4). In constraints (5-6) we aim to prevent illegal tours but allow
legal ones.
5.4.2 Belenguer and Benavent formulation
An alternative integer formulation of CARP is available from Belenguer
and Benavent (1998). This uses undirected edges and slightly different
notation than the Golden and Wong formulation. The objective function
minimises the sum of the serviced and deadhead arcs. Constraints (5-9)
(obligatory constraints) and the capacity constraints (5-10) ensure,
respectively, that each required edge will be serviced and that the
capacity of the vehicle is not exceeded. This formulation is difficult to use
directly for optimal solutions. However, it does make an important
contribution as it forms the basis of a lower bounding procedure in
Section 5.5.5 below.
Figure 5-4 : Illegal and legal subtours (Golden and Wong(1981))
Spatial Decision Support Systems for Large Arc Routing Problems
109
5.5 Lower bounds for the CARP
5.5.1 Early graph theory bounds
A number of lower bounds for the CARP have been proposed. The starting
point for a lower bound of a capacitated problem is the basic CPP, which
represents a lower bound for a problem with more than one vehicle.
Superior bounds can result from the identification of areas of the network
where it is known that more than one vehicle will traverse the same arc.
This is especially likely to occur near the depot as all vehicles must enter
Belenguer and Benavent formulation
Let R be the set of required edges and let I ={1,…,K}
.itservicingwithoutedgetraversesvehicletimesofnumber
edgeservesvehicleif1otherwise0
Repy
Repx
ep
ep
∈= ∈=
Minimise epIp Ee
eepIp Re
e ycxc ∑∑∑∑∈ ∈∈ ∈
+ (5-8)
Subject to RexIp
ep ∈=∑∈
allfor1(5-9)
IpQxdRe
epe ∈≤∑∈
allfor(5-10)
{ }1and)(
1allfor2))(())((
∈∈
−⊆≥+
PSEfVSxSxSx
R
fppRp δδ(5-11)
{}1and
1allforeven))(())((
∈
−⊆+
PVSSySx pRp δδ (5-12)
{ } integerand0,1,0 ≥∈ epep yx (5-13)
Spatial Decision Support Systems for Large Arc Routing Problems
110
and leave the depot. Therefore, if there are four vehicles in a problem and
only two arcs linking the depot, it is inevitable that there will be six
additional traversals of one or other of these arcs. Much of the work on
arc routing lower bounds has attempted to identify the distance
inevitably added by these traversals; when added to the CPP solution a
lower bound is derived. An early bound was identified by Christofides
(1973). This bound identified the number of vehicles required M = [ΣiΣj qij
/W] and added to the CPP solution 2M-1 times the length of the shortest
arc incident with the depot. Golden and Wong (1981) point out that the
Christofides bound was invalid in certain cases where the matching
required for the CPP already included additional traversals to the depot.
They proposed a new bound based on the addition of new artificial nodes
to replicate the depot and the inclusion of these and the original network
in the matching process. Further important contributions were made by
bounds developed by Assad, Pearn and Golden (1987) and Pearn (1988).
These bounds require the solution of a matching algorithm on a modified
graph H derived from G; this modified graph incorporates additional
traversals near the depot.
5.5.2 Node Duplication Lower Bound
The Node Duplication Lower Bound (NDLB), (Saruwatari, Hirabayashi
and Nishida, 1992) provides a lower bound for the CARP. This bound
involves the transformation of a given graph G = (V,E) into a graph Gt =
(Vt, Et) which is an augmented graph. This graph comprises the original
demand arcs and a number of artificial arcs. The significance of the
artificial arcs will be explained later in this chapter. This procedure is
more computationally intensive than LB1 but facilitates use of the
branch and bound algorithm.
A new set of nodes, representing copies of the nodes of the original graph,
V, must be created. For each node i ∈ V that is adjacent to a demand arc,
create the set called ‘Family of i’ where Family(i) contains Degree(i)
copies of node i from the original graph. Effectively for each node in V, it
Spatial Decision Support Systems for Large Arc Routing Problems
111
makes a number of copies of the node as is equal to the degree of that
node. Now obtain the set VD is formed by the union of Family(i) for all i ∈
V. Therefore VD = Family(1) ∪ Family(2) ∪ ... ∪ Family(n), where n is
the number of nodes in the original graph.
Then create a set of nodes VS, representing copies of the depot. Letting ‘M’
denote the number of postmen who will service the network, the set VS is
comprised of (2 × M) nodes where each node represents a copy of the
depot node, in the original graph G. The set Vt is now obtained by forming
the union of the two sets VD and VS (i.e. Vt = VS ∪ VD )
One then has to assign each demand arc in G to some arc in Gt. For any
demand arc (i,j) in G, a node is selected in Vt from Family(i), denoted by
‘k’, and another node in Vt from Family(j), denoted by ‘l’. Arc (k,l) is now
set as the demand arc in Gt corresponding to the demand arc (i,j) in the
original graph G. The demand on this arc (k,l) is set equal to that of arc
(i,j) in G. This is repeated for all demand arcs so that no two demand arcs
in Gt have common nodes. This is made possible by the fact that the
number of nodes in Family(i) equals Degree(i) for any node i ∈ VD .
2 3
4 5
6 7
1,2 2,2
3,2
2,2 2,2
1,1
1,3
1,1 1,1
1
Arc Attributes : Cost (Distance) , Volume
Figure 5-5 : The Original Graph
Spatial Decision Support Systems for Large Arc Routing Problems
112
The graph shown in Figure 5-5 above (nine arcs and seven nodes) is used
as an example (Benavent, Campos, Corberan and Mota, 1992). The
transformation of this basic graph into the graph needed for the NDLB is
shown in Figure 5-6. The depot node is adjacent to two demand arcs, and
thus Family(1) contains two copies of node 1. These as nodes are
renumbered as ‘1’ and ‘2’ in the transformed graph. This procedure is
applied to each node in the original graph, simultaneously incrementing
the corresponding node number in the transformed graph accordingly. If
four vehicles are required (M = 4), then there are eight (2M) copies of the
depot node, referred to as ‘depot nodes’, numbered in this case as nodes 19
through to 26 inclusive. The demand arcs are then inserted into Gt , with
arc (1,3) in Gt corresponding to arc (1,2) in G, arc (2,6) in Gt corresponding
to arc (1,3) in G, and similarly for the remaining demand arcs.
The next step is to form a complete graph with a complete cost matrix.
The cost on each demand arc is set to infinity (i.e. they are prohibited
from the MCPM solution). If two nodes are from separate families, the
1
2
3
4
5 7
6
8
9
10
11
12
13
14
15
16
17
18
3
3
1
1
2
2
2
2
2
Volume
Figure 5-6 : The Transformed Graph
Spatial Decision Support Systems for Large Arc Routing Problems
113
cost on the arc between them is set equal to the shortest path between
their original nodes in the original graph G. If one node of an arc (i,j) is
from a family (i.e. i ∈ VD ) and the other node is a depot node (i.e. j ∈ VS),
the cost on the arc between them is set equal to the shortest path from
the node to the depot (node 1) in the original graph G. The cost on an arc
between two nodes from the same family is set to zero. The cost on an arc
between two depot nodes is set to infinity (i.e. these arcs are prohibited).
The lower bound is calculated from the summation of the costs on the
arcs in the optimal matching on the transformed network. This is a lower
bound on the sum of the traversal costs or times for the optimal vehicle
route required for this network. If we take our example in Figure 5-5
above, the initial MCPM of the cost matrix is shown in Figure 5-7. In this
1 2
3
4
5 7
6
8
9
10 11
12 13
14
15
16
17
18
19 20 21 22 23 24 25 26
1,3
1,3 1,1
1,1
2,3 2,3
3,2
1,2
Cost, Volume
1,2
1,2
0,0 0,0 2,2 2,2
2,2
2,2
0,0 0,0
1,2
0,0 0,0
0,0
1,1
Figure 5-7 : Initial NDLB MCPM on the 9-arc example
Spatial Decision Support Systems for Large Arc Routing Problems
114
example, many of the matching arcs are between nodes in the same
family, these have zero cost. Also as there are several vehicles used, there
are additional traversals to the depot. There is also an additional
traversal between node 4 and node 5. This arises because these nodes are
of odd degree. This solution gives a value of 24 for NDLB, (additional
matching 10 + service cost 14).
5.5.3 Bound LB1
Benavent, Campos, Corberán and Mota (1992) provide a comprehensive
review of previous work and introduce new bounds, which they showed to
be superior to previous ones. It is useful to examine these bounds as they
form the basis of the new bounds presented in Chapter 7. The basic
Benavent et al bound is known as LB1. It follows a similar strategy to
earlier bounds, but relies on a more sophisticated analysis of the paths
leading to the depot. Additional nodes and paths are added to a modified
graph H derived from G and a matching algorithm is used to derive a
lower bound from this modified graph. LB1 provides at least as good a
performance as the NDLB discussed in the previous section, but is less
computationally intensive.
LB1 is calculated by reference to the modified graph GR where the
vertices are 1, 2, ..., |Vr|, numbered in non-decreasing order with respect
to their distances to the depot, so s12 ≤ s13 ≤ ... (where s1j is the shortest
path from the depot to node j) and construct a complete graph Ga = (Va ,
Ea) with Va = A ∪ B ∪ S′, where
A = {a1, ...., ar} is a set of copies of the depot.
B contains Degree(i) copies of vertex i, for i = 2, ... , r where r is the
minimum value of p such that degree(2) + ... + degree(k) >= J.
S′ contains a copy of each vertex in S (S contains the nodes of odd degree
in the original graph, G) excluding the depot and those odd vertices
whose copies are already included in set B.
Spatial Decision Support Systems for Large Arc Routing Problems
115
Costs on the edges of Ea are defined as follows: cost infinity between
every pair of vertices in set A, and, for all other pairs, the cost of the
shortest path in G between the corresponding vertices of Ga. The costs of
the edges between copies of the same node will be zero except for the
copies of the depot.
The calculation of LB1 will be explained by the examples below which are
based on the nine-arc problem presented above in Figure 5-5. Suppose
there are four vehicles, each with a capacity of five units. On each
vehicle’s route, it must travel ‘out’ of the depot along an adjacent arc,
using either arc (1,2) or arc (1,3), at the beginning of its route. Similarly,
at the end of the route, each vehicle must travel ‘in’ to the depot along one
of these adjacent arcs. In total, there will be eight traversals of the arcs
adjacent to the depot, four on outgoing journeys and four on incoming
journeys. However, in the original graph there are only two arcs adjacent
to the depot. Consequently, there will be at least six additional traversals
shared between the arcs adjacent to the depot.
The procedure proposed by Benavent et al for the LB1 bound requires the
inclusion of six artificial vertices in Vt which represent copies of the
depot. Call this set A.
A = {11, 12, 13, 14, 15, 16}.
These six copies of the depot are included in Hs so that six traversals of
arcs adjacent to the depot will be included in the lower bound. For each of
these six traversals, a copy of a vertex incident to the depot needs to be
included in Hs . This requires enough copies of vertices 2 and 3 to pair off
with the copies of the depot.
Choose the vertex closest to the depot. In this example, the shortest arc is
(1,2), so a copy of node 2 is included in Vt. When a vehicle arrives at node
2, it must travel along (or have travelled from) another adjacent arc of
that vertex to continue its journey. However, node 2 only has three
adjacent arcs so it cannot match the six traversals needed for the depot.
Spatial Decision Support Systems for Large Arc Routing Problems
116
Therefore, three copies of node 2 are included in Vt (This does not mean
that the construction of Hs eliminates the possibility of more traversals of
arc (1,2) as will be explained later). We proceed by looking at the next
closest vertex to the depot. In this example, this is node 3, which also has
three adjacent arcs. Between vertices 2 and 3, six artificial vertices are
included in Vt. This is enough to offset the traversals ‘to’ and ‘from’ the
depot.
Include three copies of node 2 in Vt and three copies of node 3. Call this
set B.
B = {21, 22, 23, 31, 32, 33}
Consequently, extra traversals are added due to the presence of nodes of
odd degree in G. Of the remaining vertices in the graph, nodes 4, 5, 6, and
7, two of them, nodes 4 and 5, are of odd degree. These vertices must be
matched with other vertices. Therefore, an artificial copy of each of these
vertices is created and added to Vt. Call this set S′. Note that vertices of
odd degree of which copies have been included in set B cannot be included
in set S′ (Figure 5-8).
S′ = {4*, 5*}.
Va = A ∪ B ∪ S′
In set A vehicles cannot travel directly from one copy of the depot to
another. Therefore, between all other pairs of arcs, the cost of the shortest
path ascertained from the original graph will be used. Costs between
Set A
Set B
Set S′
Node 2
4 5
Node 3
Figure 5-8 : Optimal Matching of Hs
Spatial Decision Support Systems for Large Arc Routing Problems
117
copies of the same node will be zero. We then carry out a MCPM on Hs.
LB1 now estimates that arc (1,2) will be traversed (or repeated) at least
three times, as will arc (1,3). In addition, arc (4,5) will be traversed at
least once. This gives a bound of 24, (10+14) which means that for this
example LB1 gives the same bound as NDLB discussed above. It is worth
noting that the inclusion of three copies of nodes 2 or 3 in set B does not
mean that the matching of Hs cannot represent additional traversals of
arc (1,2) or arc (1,3). This possibility is not restricted by the MCPM.
5.5.4 Bounds exploiting cuts away from the depot
Win (1988) studied a number of issues related to arc routing problems
and devised an improved bound which he called ZAW2. This work
recognised that CARP lower bounds could be improved by examining
additional traversals away from the depot, at any point where a
constriction occurred in the graph. Win’s work was extended by
Benavent, Campos, Corberán and Mota (1992) who proposed a superior
bound LB2. This is based on the concepts introduced in LB1, but
improves the bound by also examining cuts in all parts of the graph, and
not just at the depot.
LB2 exploits the properties of cutsets. In looking at LB1 (Section 5.5.2)
we examined the number of times vehicles travelled along the arcs
adjacent to the depot. The set of arcs adjacent to the depot would make
up a valid cutset (as shown in Figure 5-9 below). If these arcs were
removed from the graph, the graph would be split into two components.
Component 1 = {1}
Component 2 = {2, 3, 4, 5, 6, 7}
When dealing with lower bounds, a cutset gives us a lot of information.
Call arcs (1,2) and (1,3) the ‘cutset’ and the arcs (2,4), (2,3), (3,5), (4,5),
(4,6), (5,7), (6,7) the ‘cut sub-graph’. The arcs in the cut sub-graph can be
referred to as arcs ‘beyond the cut’. This means those arcs on the opposite
Spatial Decision Support Systems for Large Arc Routing Problems
118
side of the cut to the depot. We will use the following notation, drawn
from Benevant et al, in the following examples.
ps = number of vehicles to service the cut and the cut sub-graph.
qs = the number of arcs in the cut.
rs = 2ps - qs, the number of extra traversals required of those arcs in q.
cs = the length of the shortest arcs in the cut.
In the example in Figure 5-5 on page 111, assume that there are four
vehicles each with a capacity of five units. Then the number of vehicles
required (ps) to cross the cut is 4 and the number of arcs (qs) crossing the
cut is 2. Consequently, the number of extra arc traversals across the cut
is six. There will be four outgoing crossings, and four incoming crossings.
Of these eight crossings, only two can service arcs, therefore the six other
crossings must be traversals.
rs = (2 ×4) - 2 = 6.
The shortest arc in the cut is arc (1,2). This has a length of 1.
cs = 1.
2 3
4 5
6 7
1,2 2,2
3,2
2,2 2,2
1,1
1,3
1,1 1,1
1
Figure 5-9 : The Cutset from LB1
Spatial Decision Support Systems for Large Arc Routing Problems
119
Benavent et al proposed a new bound, known as LB2, which further
improves on LB1 by considering successive edge cutsets. Consider again
the example above. In calculating LB2 (Figure 5-9) a set U is formed
which initially contains the depot, a second set V′ is formed from V – U.
Sets A, B are calculated in similar way to LB1.
S′ = S ∩ V′s (that is the set of odd degree vertices beyond the cut)
One difference is that we will no longer refer to the elements of set A as
copies of the depot, as we will also be dealing with cuts removed from the
depot. The calculation of LB2 introduces another set of nodes, called set
X. This is not relevant for the cut at the depot, but arises for all cuts away
from the depot.
A = is a set of rs artificial vertices
B contains d(i) copies of vertex ij, where ij ∈ V′s where j = 1,…, h
S″ contains a copy of each vertex in S′ (S contains the nodes of odd degree
beyond the cut) except for those whose copies are already included.
X is a set of max{0, |S′| - rs} artificial vertices
Va = A ∪ B ∪ S′ ∪ X
Costs on the arcs Et are calculated so that the cost on the edges between
set B and set S″ are the shortest paths between the corresponding
vertices in the original graph. Note that the cost on edges between copies
of the same vertex will be zero. For edges between u ∈ B ∪ S″ and v ∈ A
∪ X, the cost on the edge is the minimum distance from node i to any
node in U.
If we perform these calculations for the graph shown in Figure 5-9, then
U = {1}, V′ = {2, 3, 4, 5, 6, 7}
A = {a1, a2, a3, a4, a5, a6}.
B = {21, 22, 23, 31, 32, 33}.
Spatial Decision Support Systems for Large Arc Routing Problems
120
S″ = {4*, 5*}.
X = {} (not relevant at depot)
The MCPM on Hs with a cut at the depot returns a value of 10 (the same
value as LB1). We have now found a lower bound using the information
from the Cut (1-2, 1-3) and the matching of the remaining nodes of odd
degree (nodes 4 and 5). Nevertheless, from the research by Win (1988)
(further developed by Benavent et al) it is clear that more information
can be gained by looking at successive edge cutsets. Moving away from
the depot, out into the graph, in search of ‘good cuts’ generates such
cutsets. A significant cut is likely to arise where the road network is
constricted, for example where there are bridges over a river.
The current lower bound is 10 + total service cost = (10 + 14) = 24. Since
the cost of servicing the graph is constant, henceforth we will use the
term ‘lower bound’ to refer to that time that is additional to the servicing
time. In this case the lower bound is ten. The shortest arc out of the depot
is arc (1,2), with a traversal time of 1 and six additional traversal of arcs
out of node 1 were required. A value L1 is stored to represent the
minimum traversal distance for crossing the cut, this calculation is
similar to that proposed by Zaw.
2 3
4 5
6 7
1,2 2,2
3,2
2,2 2,2
1,1
1,3
1,1 1,1
1
Figure 5-10 : The second cut
Spatial Decision Support Systems for Large Arc Routing Problems
121
L1 = cs × rs
cs = 1
rs = 6
L1 = 6.
Suppose that the dotted line in Figure 5-10 shows a natural cut in the
graph (e.g. a river crossing) the cutset is (2-4, 3-5).
U = {1, 2, 3}.
Suppose that this was the river, the ‘North’ side of the graph has a
demand of 12 (including the bridges across the cut). With a vehicle
capacity of four, three such vehicles are required to cross the river to
service the arcs beyond and including the cut. Each of these three
vehicles will also have to return. In total, the river must be crossed six
times, three times on outgoing traversals and three on incoming
traversals. However, only two demand arcs cross the cut. As a result four
out of the six times that the cut is crossed, a bridge will be traversed
without servicing. We create four artificial vertices that can represent
nodes on the ‘South’ side of the graph.
A = {a1, a2, a3, a4}.
In order to provide for the four artificial edges that will create the
traversals across the cut, we need to match set A with copies of the nodes
on the ‘North’ side of the cut. Following a similar argument as was used
to create set B in LB1, 3 copies of node 4 will be created and 3 copies of
node 5 will created since both nodes have degree 3.
B = {41, 42, 43, 51, 52, 53}.
Also, within the North Side, there may be nodes of odd degree (not
already included in set B) which need to be matched with each other. In
this instance, for nodes 6 and 7, we can see that there are no such
additional nodes.
S″ = {}
Spatial Decision Support Systems for Large Arc Routing Problems
122
Including set S″ in Vt represents the odd nodes not yet included in Vt. Odd
nodes between the depot and the cut are not included. Including only
some of the odd nodes from the original graph might possibly result in a
sub-optimal matching (i.e. a matching which is too high). This might
occur if a MCPM of odd nodes in the original graph involved matchings
across the cut. If in a MCPM on the original graph, a node of odd degree
beyond the cut would be matched with a node of odd degree on the depot
side of the cut, this cannot be catered for by a matching on Hs alone.
L1 = L1+ (cs × rs) = 6 + (4 × 2) = 14.
Therefore, we include artificial vertices in Vt which will prevent this
occurring. The number of artificial nodes created for set X is equal to the
number of nodes in set S″ less rs. If this number is negative, set X is
empty. The costs on At, the arcs in the complete graph Hs should be
calculated same as before.
X = max{0, |S″| - rs}
Va = A ∪ B ∪ S″ ∪ X
The MCPM on this graph returns a value of 8. At first sight, this bound is
inferior to the initial bound of 10. However, if we add the L1 value of 6
calculated earlier, the estimated traversals behind the cut, and add this
to the new bound of 8, we get a total of 14. This gives a better lower
bound than the 10 given earlier, so we save this as our new lower bound.
Create the next cutset.
U = {1, 2, 3, 4, 5}
Using the same principles as before:
A = {a1, a2, a3, a4} (as two vehicles needed to service arcs (4,6), (5,7), (6,7))
B = {61, 62, 71, 72}
S″ = {}
X = {}
Va = A ∪ B ∪ S″ ∪ X
Spatial Decision Support Systems for Large Arc Routing Problems
123
Table 5-2 : Summary of LB2 Algorithm
Set U = { 1 }, L = L1 = L2 = 0;
WHILE U <> V DO
Let V' = V - U and G' be the graph induced by V'
Find the connected components of G'. Suppose G' has ‘k’
components G's = (V's , E's ) and
},:),({)(cutsetedge UjViEjieV ∈′∈∈==′δ
FOR s := 1 TO k DO
)(setandbelow)(seegraphweightedaconstruct then0orIf
Let}2,0{max
min0:)({
q
.
)(
)(e
ss
sss
ss
sss
eves
ess
VEes
HMPmHrS
VSqpr
ccqVeq
Wp
s
ss
=>≠′
′=′−=
=>′∈=
=
′∈
′′∈∑
φ
δ
δ
δ
ENDFOR s.
s
t
Ss
T
t
ss
crLL
LLCLL
mL
∑
∑
=
=
+=
++=
=
1
1
11
}1,2max{2
Set U' = the set of nodes {i ∈ V | i is adjacent to a vertex in
U }.
U := U ∪ U'.
ENDWHILE.
Set LB2 = L2.
ENDALG
Spatial Decision Support Systems for Large Arc Routing Problems
124
Construction of Complete Weighted Graph HS for LB 2
Preliminary
Let Mdist(i) denote the minimum distance between vertex ‘i’ in V'i and any vertex of U .
Suppose that the vertices in V'i are renumbered i1 , i2 ,..... in such a way that Mdist(i1)
Mdist(i2) ........
Let h be the minimum integer such that Degree(i1) + ........+ Degree(ih) ri.
The Vertex Set
Graph HS is a complete weighted graph with vertices are drawn from 4 sets : Set A ∪ Set
B ∪ Set S″ ∪ Set X
Set A is a set of artificial vertices, where = ri
Set B contains Degree(ik) copies of vertex ik V'i for k = 1....h.
Set S″ contains a copy of each vertex in S'i except for those copies who are already in Set
B.
Set X is a set of artificial vertices, where = Max [ 0, (number of nodes in S'i ) - ri ]
The Arc-Cost Matrix
Let wij represent the cost on edge (i,j) in HS . The costs on the edges in HS are infinite
except for the following:
• if i, j ∈ {Set B ∪ Set S″} then wij = the shortest path between the corresponding
vertices in G ( Note that wij = 0 if i and j are copies of the same vertex).
• if i ∈ {Set B ∪ Set S″} and j ∈ {Set A ∪ Set X} then wij = Mdist(ki) where ki
denotes the corresponding vertex to i in G.
• if i, j ∈ Set D then wij = 0.
Again, form the complete graph Hs=(Vt, Et) and carry out a MCPM. The
MCPM on this graph returns a value of 2. Costs on the edges of Et are
calculated as before. Adding the estimated traversals behind the cut, 14,
to 2, we get a new lower bound of 16. Since this is higher than any bound
seen before we store this as our bound. Next, we look at the adjacent
nodes of the cutset. We see that all nodes in E are now members of U, and
so the lower bound algorithm stops. The current, i.e. best, value of LB2 is
16; adding this to the total service time for the graph, we get a lower
bound of 30. The optimal solution for this graph is 32. Therefore, a
solution of 30 appears to be a good lower bound for this problem.
Spatial Decision Support Systems for Large Arc Routing Problems
125
We believe that LB2 provides the best graph theory based CARP bound
published to date. Although LB2 is computationally expensive, owing to
its use of MCPM techniques, it can be feasibly implemented for quite
large graphs. LB2 (Table 5-2) will form the basis of a modified bound for
the Time Capacitated Arc Routing Problem (TCARP) in Chapter 7.
5.5.5 Cutting Plane Bound
While most of the work on lower bounds for CARP has been based on
graph theory rather than linear programming approaches, it seems likely
that, in the future, the latter approach will lead to superior bounds.
Belenguer and Benevent (1997) discuss a cutting plane approach for
CARP that uses a linear programming relaxation to provide a lower
bound. This is partly based on the LP formulation by the same authors
discussed above (Belenguer and Benavent, 1998). The cutting plane
algorithm solves a relaxation of the linear program containing a subset of
valid inequalities for the CARP formulation. A solution is generated and
a set of valid inequalities violated by the optimal solution is identified.
This set of inequalities is added to the optimal solution. This process is
terminated when an upper bound is reached, in which case the optimal
solution has been found, or when no known inequalities are breached.
The latter case provides a lower bound for the solution.
Belenguer and Benevent conducted computational testing on problems
that had previously appeared in the literature, these were relatively
small problems with up to 50 nodes. These results indicated that the
cutting plane algorithm (CPA) produced a superior bound to LB2 and
problems with 50 nodes and 97 vertices were solved to optimality. Their
calculations showed that LB2 had an average gap of 4.21% from the
upper bound. The cutting plane algorithm had a much-reduced gap of
0.74%. The latter reached the optimal value in over half the test
problems; some three times the proportion of optimal solutions reached
using LB2.
Spatial Decision Support Systems for Large Arc Routing Problems
126
Chapter 6 : Solutions for the CARP
6.1 Heuristic Solutions for the Capacitated Arc Routing Problem
6.1.1 Single Pass heuristics
One approach to CARP is to use an algorithm that considers vehicle
allocation and route sequence simultaneously. A variety of single pass
(parallel) heuristic algorithms have been proposed to derive feasible, but
not necessarily optimal, solutions for arc routing problems. These use
heuristic methods to build directly a postman tour. Christofides (1973)
introduced the first of these generic algorithms, which is known as the
Construct-Strike Algorithm. In this algorithm feasible cycles (with total
arc demand less than or equal to the vehicle capacity) are constructed
which, when removed, do not cause the remaining graph to become
disconnected. Then the arcs serviced in the cycles created in the first are
deleted from the current graph and this process is repeated until no more
cycles are found. Then a matching problem is solved for the nodes of odd
degree and two copies of the depot, the additional paths are added to the
graph and the initial step is repeated. This approach was developed by
Pearn (1989) who developed a Modified Construct-Strike Algorithm with
an improved arc selection procedure. Pearn indicated that the original
algorithm was O(mn3) and the revised algorithm O(mn4), indicating that
solutions can be found reasonably effectively. Later work (Coutinho-
Rodrigues, Rodrigues et al., 1993) tested the algorithms using a heuristic
matching of the odd degree nodes (odd nodes were matched with the
lowest cost path between them). They found that this approach achieved
similar results, but avoided the excessive computation time entailed in
the use of an optimal MCPM algorithm.
Spatial Decision Support Systems for Large Arc Routing Problems
127
Golden and Wong introduced an algorithm inspired by the Clarke-Wright
heuristic for vehicle routing (Clarke and Wright, 1964). This begins with
each demand arc being serviced by a separate cycle and then attempts to
combine the cycles. This became known as the Augment-Merge Algorithm.
Pearn (1991) improved on this algorithm, making it suitable for operation
on relatively sparse graphs with large edge demands, relative to the
vehicle size. The Parallel Insertion Algorithm (Chapleau, Ferland,
Lapalme and Rousseau, 1984) was inspired by the insertion procedures
used for node routing problems. This requires assessment of two issues;
which arc to insert and where to insert it. For each arc we must
determine the existing route into which it should be inserted in order to
minimise the detour incurred. When this route has been identified, we
must determine which remaining arc should be inserted next.
The Path-Scanning Algorithm (Golden, DeArmon and Baker, 1983) forms
a cycle by adding an edge according to one of five edge selection criteria.
Each of the criteria is used to form a solution and the one with the lowest
cost is used. Pearn (1989) proposed that the edge selection criteria be
chosen randomly and found that this improved the solution. Coutinho-
Rodrigues, Rodrigues, and Climaco (1993) proposed the use of three
additional criteria.
Single pass heuristics are appropriate for many classes for problems,
where substantial volumes are to be collected or delivered. Coutinho-
Rodrigues et al (1993) review a number of heuristics in the context of a
refuse collection application in Portugal. A review of these algorithms is
given by Assad and Golden (1995), who compare the computational
performance of the various algorithms, this is reproduced in Table 6-1. It
can be seen from this table, that the data sets used were much smaller
than most practical real-world arc routing problems. Furthermore, the
networks used to test these algorithms were largely complete, which is
untypical of real road networks.
Spatial Decision Support Systems for Large Arc Routing Problems
128
6.1.2 Route-first cluster-second heuristics
The Route-First Cluster-Second Approach (RFCS) is based on the creation
of a giant tour through all the arcs. This tour is then decomposed into a
set of sub-tours, each of which is feasible with regard to the capacity of
Table 6-1 : Heuristic performance (adapted from Assad and
Golden(1995))
GDB P89A P89B P91
Test Beds
Number of problems 23 20 15 30
Number of nodes (n) 7-27 11-17 11-17 13-27
Number of edges (m) 11-55 55-136 45-116 23-51
Density (%) 13-100 100 70-90 15-30
Algorithms (% over
Lower bound)
Construct-strike CS 17.91 2.29 3.85 71.95
Path-scanning PS 11.03 3.13 3.59 57.97
Augment-merge AM 9.16 56.78
Random path scan RPS 8.05 2.76 3.16 51.30
Augment-Insert I AI1S 12.39 44.17
Augment-Insert II AI2 13.95 44.97
GDB (Golden, DeArmon et al., 1983)
P89A (Pearn, 1989) dataset A
P89B (Pearn, 1989) dataset B
P91 (Pearn, 1991)
Spatial Decision Support Systems for Large Arc Routing Problems
129
the vehicle. This approach can be used for VRP applications (Beasley,
1983). In the arc routing context, Beltrami and Bodin (1974) describe a
RFCS solution approach for a refuse collection problem. In this
application, a giant Euler tour is constructed, using a simplified matching
procedure, and this route is subsequently partitioned into feasible route
segments. A similar problem was faced by Stern and Dror (1979), who
describe a meter reading application in Israel. They initially solve a RPP
using a heuristic matching technique. The large Euler tour is
subsequently partitioned by removing edges from the tour until the time
based capacity constraint for each route is met. In this application, the
subdivision of the tour was made easier by the fact that the required
routes were open ended. This arose as the meter reader was not required
to return on foot to the depot, but could travel by public transport
instead. The RFCS approach can take advantage of the fact that the CPP
can be solved optimally. Therefore, if a good partition can be achieved,
good solutions can be derived by this technique. However, in practical
problems it is not always straightforward to subdivide the uncapacitated
solution.
6.1.3 Cluster-first route-second heuristics
An alternative approach to RFCS is to derive an initial cluster and then
sequence the points or arcs within that cluster, this is the Cluster-First
Route-Second Approach (CFRS). This approach has been successfully
used in node routing problems, a well-known example of an algorithm of
this type being the sweep algorithm of Gillett and Millar (1974). This
algorithm assumes a problem with a set of n customer locations (in terms
of rectangular co-ordinates with the depot at the origin) and demands,
and a set of vehicles and capacities. The customers are re-numbered in
terms of increasing polar co-ordinate angle. Then starting with the
customer with the smallest co-ordinate angle, the locations are
partitioned into groups. The first route consists of locations 1,2,...,J
(remembering that the depot is location 0), where J is the last location
Spatial Decision Support Systems for Large Arc Routing Problems
130
that can be added without exceeding the vehicle capacity constraint. The
second route contains locations J+1, J+2,...,L, where L is the last location
that can be added to the second route without exceeding the vehicle
capacity constraint. The remaining routes are formed in the same
manner. The second stage of the algorithm involves evaluating the
impact of swapping customers between routes, in an effort to improve the
solution. A variety of stratagems are used to identify other possibilities,
including shifting the X and Y axes so that the first location becomes the
last, the second the first and so forth.
In the arc routing field, CFRS solutions are less common. One recent
example of the use of a CFRS approach in arc routing is in winter gritting
in Germany (Amberg, Domschke and Voß, 2000). Another example of
interest is the Arc-Partitioning Problem (Levy and Bodin, 1989; Bodin
and Levy, 1991) which arises in postal delivery in the US. This is a
location routing problem, where a series of walking tours are generated
from a parked vehicle. The algorithm must derive both the location where
the vehicle is parked and the sequence of arcs to be visited from that
location. The objective is to break a network into partitions where the
workload in each partition is approximately the same. Seed points are
used as the basis of the partitions, these can be input by the user or
generated automatically by the algorithm. This automatic procedure aims
to maximise the minimum distance between all pairs of seed points.
When seed points have been generated, partitions are derived and a
balancing routine is used to swap arcs between partitions to better
balance the solutions. The balancing step may lead to a revision of the
seed points.
6.1.4 New approaches to CARP
Considerable recent work has taken place in OR/MS on new general local
search heuristic techniques, for example genetic algorithms and
simulated annealing (Eglese, 1990). These could be used to improve an
arc routing solution; Eglese (1994) discusses the use of simulated
Spatial Decision Support Systems for Large Arc Routing Problems
131
annealing for a winter gritting problem on rural road networks.
Greistorfer (1994) discusses the use of tabu search for arc routing. A
recent paper (Hertz, Laporte and Mittaz, 2000) conducts extensive testing
of the tabu search approach on the CARP datasets used by some of the
earlier researchers (Section 6.1.1). The tabu search approach was shown
to reach the optimal solution in eighteen of the twenty-three cases tested
and was shown to be superior to earlier heuristics. On larger randomly
generated examples, the tabu search approach was compared to the best
lower bound derived from three approaches; the CPA bound (Section
5.5.5) the LB2 bound (Section 5.5.4) and the NDLB bound (Section 5.5.2).
The tabu search approach obtained solutions within 5% of the best lower
bound, which was probably not optimal. These results suggest that this
area of research is likely to provide further useful approaches for arc
routing problems.
6.1.5 Real world network representation issues
Heuristic approaches may have an advantage over optimal techniques in
that they can be more easily modified to accommodate a variety of issues
that arise in practical problems. In arc routing, several potential
complications arise at junctions. Left or right turns may be prohibited or
be undesirable for slow moving vehicles, depending on which side of the
road that vehicles use. In the case of some applications, such as snow
clearance (Figure 6-1), a right turn (in a country where vehicles travel on
the right) may be preferable to travelling through an intersection
(Gendreau, Laporte and Yelle, 1997). Similar issues arise in street
sweeping applications, where it may be more desirable for the vehicle to
continue along the street or turn right rather than turn left or to make a
U turn. Bodin and Kursh (1979) describe the use of penalties to reduce
the number of undesirable pairings of arcs (Table 6-2). Similar issues
arise in the paper on refuse collection by McBride (1982).
Spatial Decision Support Systems for Large Arc Routing Problems
132
Where both sides of the street are to be visited a number of issues arise at
junctions. In pedestrian applications, for instance urban mail delivery,
crossing the road may impose a significant delay. This type of constraint
can be modelled by the use of penalties. In an urban postal delivery
application Roy and Rousseau (1989) discuss the use of penalties at
junctions to reflect the additional time take to cross the street (Figure
6-2).
Table 6-2 : Turn Penalties (Bodin and Kursh, 1979)
Movements Points
U-turns 8
Deadhead to sweep 5
Sweep to deadhead 5
Left hand turn 4
Right hand turn 1
Straight ahead (no turn) 0
Figure 6-1 : Block Design for snow ploughing
(Gendreau, Laporte et al., 1997)
Spatial Decision Support Systems for Large Arc Routing Problems
133
6.2 Branching approaches to CARP
6.2.1 Branching techniques
The heuristic procedures described above provide a method of reaching a
good solution to arc routing problems. Each heuristic is likely, but not
guaranteed, to reach a solution close to the optimal. Linear programming
approaches, discussed in Chapter 5, could be used to identify the optimal
solution to the problem. Unfortunately, optimal approaches are currently
impractical for many real world problems. In OR/MS, a synthesis of
optimal and heuristic approaches exists in the class of techniques known
as branching algorithms. These identify a lower bound for the problem
and an upper bound (i.e. a feasible, but non-optimal solution). The
algorithm then uses a systematic enumerative procedure to find a
feasible solution closer to the lower bound, which may be recalculated
using information generated from the procedure. When the lower and
upper bounds are equal, an optimal solution has been found.
Figure 6-2 : Network enhancement by addition of penalties at
junctions (Roy and Rousseau, 1989)
Spatial Decision Support Systems for Large Arc Routing Problems
134
The discussion above and in Chapter 5 indicated that lower and upper
bounds for the CARP could be identified. However reaching an optimal
solution from this information remains extremely difficult. While lower
bounds can be calculated, these calculations do not provide feasible
solutions. A feasible solution may not exist at a lower bound, and it is
extremely difficult to establish whether such a solution exists or not.
Branching algorithms will perform better than total enumeration if some
possibilities can be discarded, when further calculation would lead to
answers above the upper bound (not optimal) or below the lower bound
(infeasible). Branching techniques will therefore perform better if the gap
between the upper and lower bounds is extremely small, allowing many
possibilities be discarded. Where this gap is large, it will be
computationally impractical to find the optimal solution.
A branching approach might be based on an LP based formulation of the
problem, for example the cutting-plane (CPA) lower bound (Belenguer
and Benavent, 1997) discussed in Section 5.5.5. These authors state that
they are working on a fully automated Branch and Cut procedure for
CARP. This will allow larger examples be solved to optimality, although
heuristic procedures will still be needed for many classes of practical
problem. Other branching procedures can be based on customised
approaches, such as the only known example of an optimal branch and
bound CARP procedure by Hirabayashi, Saruwatari, and Nishida (1992).
We will examine this algorithm in further detail in the following section
and will use it for some computational comparisons will lower bounds in
Section 7.5.
6.2.2 The Tour Construction Algorithm
The linear programming formulations discussed in Section 5.4 could
potentially form the basis of an optimal solution procedure for CARP.
Problems of a practical size are intractable with these formulations.
Therefore, in most real world problems other approaches must be used.
An optimal solution procedure for CARP was introduced by Hirabayashi,
Spatial Decision Support Systems for Large Arc Routing Problems
135
Saruwatari, and Nishida (1992). This used a customised, graph theory
based, branch and bound procedure based on a lower bound by the same
authors, the Node Duplication Lower Bound (NDLB) (Saruwatari,
Hirabayashi et al., 1992) (see Section 5.5.2). The Tour Construction (T.C.)
algorithm, as devised by Hirabayashi et al (1992), is an algorithm based
upon the Branch and Bound method, where an original problem is split
into two sub-problems. Then each sub-problem is dealt with in turn and
split into two further sub-problems, and so on. We must determine how to
create two sub-problems from each sub-problem, and identify what
characteristic differentiates each of the two subsequent sub-problems.
In this algorithm, the left sub-problem is characterised by an arc that is
prohibited from the optimal solution. The right sub-problem and all child
sub-problems will definitely contain that arc in the solution. The T.C.
algorithm uses a MCPM function, which is passed a matrix of the costs on
the arcs in the specific graph. The MCPM function returns the optimal
matching and a value for the lower bound. However, the algorithm may
require the creation of a very large number of sub-problems before a
feasible solution is found. This requires multiple use of the MCPM
algorithm. This is both processor and memory intensive, making it
difficult to solve large, or even medium sized, problems. An improved
version of the algorithm (Kiuchi, Shinano, Hirabayashi and Saruwatari,
1995) exploits the use of multiple computers in parallel to reach a
solution more rapidly.
As we have seen above in Section 5.5.2, the initial lower bound for the
NDLB bound is obtained by summing up the costs on the arcs in the
initial optimal matching. This lower bound is a lower bound on the sum of
the costs or traversal times for the optimal vehicle route given the matrix
that is sent down to the MCPM function. The Node Duplication
Transformation now provides significant advantages for the operation of
the branch and bound procedure.
Spatial Decision Support Systems for Large Arc Routing Problems
136
The presence (or absence) of a feasible solution is based on the concept of
an alternating path. Because of the characteristics of the node-duplicated
graph, we can identify paths whose arcs are alternately an MCPM
solution arc (dotted lines) and a demand arc (solid lines). An alternating
path which starts and ends at the depot node and which does not violate
either the time capacity constraint or the volume capacity constraint of
the vehicle is called a postman path. A collection of mutually disjoint
postman paths where all demand arcs are contained in some postman
path is called a postman tour.
1 2
3
4 5
7
6
8
9
10 11
12 13
14
15
16
17
18
19 20 21 22 23 24 25 26
1,3
1,3 1,1
1,1
2,3 2,3
3,2
1,2
Cost, Volume
1,2
1,2
0,0 0,0 2,2 2,2
2,2
2,2
0,0 0,0
1,2
0,0
1,3
1,3
1,1
Figure 6-3 : Postman paths on the 9-arc example
Spatial Decision Support Systems for Large Arc Routing Problems
137
In the example in Figure 6-3, there are four alternating paths. Dotted
lines are used to identify artificial arcs found by the MCPM on the
complete transformed graph. When calculating the total time of the
alternating paths, we use the traversal times for MCPM solution arcs (or
artificial arcs) and the service time for demand arcs. The total costs for
the four alternating paths above are 2, 4, 6, and 14. This would be a
postman tour if postman capacity were 14 or more. If capacity was less
than 14, then the alternating path of length 14 is not a feasible postman
path, so a postman tour would not exist.
A branch and bound algorithm proceeds by bringing the upper bound (the
feasible solution) closer to the lower bound (a lower limit on the optimal
solution). Initially the algorithm sets the upper bound to the maximum
total length of time possible to service the network, which is the product
of the number of vehicles by the time capacity per vehicle. One unit of
time is added to this and the total service time subtracted in order to give
the upper bound. In many cases, this does not provide a very tight bound,
as one of vehicles may be largely empty in the optimal solution.
The tour construction algorithm splits each sub-problem into two further
sub-problems (Figure 6-4). The left sub-problem has the characteristic
that the branching arc can never be in the MCPM solution, while the
right sub-problem has the characteristic that the branching arc must be
in the MCPM solution. For each arc (i,j) in the MCPM solution, the lower
bound of the sub-problem with (i,j) prohibited from being in the MCPM
solution is found. The branching arc (k,l) is the arc that returns the
highest lower bound, when prohibited.
Starting from the original problem, at each iteration work is carried out
on the right sub-problems, while saving the left sub-problems. Movement
down the right hand side is continued until the sub-problem becomes
infeasible or the sub-problem’s lower bound is greater than or equal to
the upper bound. As the algorithm proceeds, additional restrictions are
placed on the solution, by the addition of definite arcs. Therefore, the
Spatial Decision Support Systems for Large Arc Routing Problems
138
lower bound of each sub-problem is always greater than or equal to lower
bound of its parent.
The other occasion at which the algorithm stops at a sub-problem is when
we have a postman tour. When this occurs, the upper bound is set to the
lower bound of the current sub-problem, and this sub-problem is stored as
the banker. The banker is defined as being the best solution found so far.
As the upper bound has been decreased, all stored left sub-problems
whose lower bound is greater than or equal to the upper bound are
deleted (We ignore sub-problems whose lower bounds are equal to the
upper bound, as we are looking for one optimal solution).
A
B C
D
F
E
G
Figure 6-4 : Branch and Bound Sub-Problems
A
B C
D
F
E
G
Optimal Solution
Figure 6-5 : Fathomed Sub-Problems and Optimal Solution, Sub-
Problem G
Spatial Decision Support Systems for Large Arc Routing Problems
139
When all of the remaining sub-problems have been deleted, all branches
have been fathomed (Figure 6-5). At this point, the current feasible
solution (banker) is thus the optimal solution (Sub-Problem G).
Alternatively, if a sub-problem existed with a lower bound less than the
upper bound, then that sub-problem would be treated in the same way as
sub-problem A.
A definite arc is a set of arcs that form a particular artificial arc. This
artificial arc is a path of two or more arcs, which are guaranteed to form
part or all of a postman path in the final solution. Initially, at the
beginning of the algorithm, the definite arcs are all the demand arcs.
When we branch from a sub-problem to its right sub-problem, the
branching arc is forced to be in the solution. This has the effect of making
two definite arcs into one definite arc. For example if arc (9, 12) is the
first branching arc in the nine-arc example, we now have a definite arc
from Node 5 to Node 8. Arcs (5, 9) and (8, 12) are demand arcs and arc (9,
12) must be in the solution. We also notice that the length of the definite
arc has become 7 hours. An advantage of this is that Nodes 9 and 12 are
excluded when finding the MCPM solutions to sub-problem C (when
finding the branching arc for sub-problem C) as they are automatically in
the MCPM solution. As we progress further down the right hand side, we
need to look at fewer arcs when finding an MCPM solution. We also have
less definite arcs but the definite arcs are getting longer. Therefore, at
each right sub-problem we look at the new definite arc formed and see if
we can prohibit more arcs from being in the MCPM solution.
When we examine the transition from sub-problem C to sub-problem E, if
the new definite arc causes an arc in the MCPM solution of C (and
therefore A) to be prohibited, it becomes necessary to find a new lower
bound and MCPM for sub-problem E. It is clear that until a new lower
bound is calculated, the right sub-problems’ lower bounds will remain the
same, and less than the upper bound. A summary of the procedure for the
TC algorithm can be found in Table 6-3.
Spatial Decision Support Systems for Large Arc Routing Problems
140
6.2.3 Practical feasibility of branch and bound algorithm
The computational experiments by Hirabayashi, Saruwatari, and Nishida
were on relatively small well-connected graphs with large demands per
arc. This type of network suited the techniques used in this algorithm.
These results indicated that the solve time increased with the number of
demand arcs |R|. However, the computational demands of the algorithm
Table 6-3 : Outline Code of the Tour Construction Algorithm:
Set U := ∅ ;
Transform network using Node Duplication;
Find Lower Bound and MCPM of the original problem;
Add original problem (Lower Bound and MCPM) to U;
Set Upper Bound := maximum possible lower bound + 1;
WHILE ( U is NOT Empty ) DO
Choose Sub-Problem with smallest Lower Bound from U.
WHILE ( LB < UB ) AND (sub-problem is feasible) DO
IF Postman Tour exists THEN
Upper Bound := Lower Bound;
Banker := Postman Tour;
Delete any Sub-Problems whose Lower Bound ≥ Upper Bound;
ELSE REPEAT
Select Branching arc (x, y);
Add Sub-Problem to U where arc (x, y) is prohibited;
Make definite arc and do prohibit arcs;
UNTIL (Sub-Problem is feasible) OR (Arc in MCPM is prohibited);
IF Arc in MCPM was Prohibited THEN
Find new Lower Bound and MCPM.
ENDIF;
ENDWHILE;
ENDWHILE;
Optimal Solution is in Banker;
ENDALG;
Spatial Decision Support Systems for Large Arc Routing Problems
141
are also reflected in the number of sub-problems generated and this tends
to increase greatly as the size of the problem increases. The
representation of each sub-problem requires additional machine memory.
Two MMangtSc students in UCD (Montwill and Naughton, 1994),
working with the author, implemented the branch and bound algorithm
and tested a number of minor modifications. Their experiments identified
excessive use of computer memory as an important limitation on the
ability of this algorithm to solve larger problems. As the algorithm uses a
complete representation of the network and as a large number of sub-
problems are generated, large amounts of memory are used in the
solution procedure
Montwill and Naughton identified a number of strategies that could be
used to improve the performance of the branch and bound algorithm. An
estimate of the matching could be used, by not updating the matching for
every small change in the solution. This would tend to improve solution
time without addressing the problem of excessive memory use. Another
approach used generates more lower bounds, increasing the computation
time, but possibly reducing the number of sub-problems. The use of a
heuristic approach to identify a good initial upper bound is also a means
by which the number of sub-problems could be reduced.
The experiments on a parallel branch and bound algorithm for CARP
(Kiuchi, Shinano et al., 1995) provides a method of approaching
somewhat larger problems, although in this paper only problems
previously solved were attempted. The use of multiple machines in this
example means that greater computing power was used. This makes it
difficult to assess fully the contribution of parallelisation to solution of
CARP. Further improvement in performance is probable if superior
algorithmic and computer science techniques are employed. However, the
branch and bound approach seems unlikely to be easily extended to solve
large practical problems.
Spatial Decision Support Systems for Large Arc Routing Problems
142
Chapter 7 : CARP on Irish Rural networks
7.1 Postal delivery of rural networks
7.1.1 Irish Rural Road Networks
Earlier chapters of this dissertation have reviewed the DSS field,
identifying routing DSS as an important application. Within the routing
field, this research has placed a particular emphasis on arc routing
problems. This chapter discusses the focus of this dissertation, arc
routing for large sparse networks, a problem has not been discussed
substantially in the literature. The particular application is rural postal
delivery, which is an interesting TCARP.
The problems discussed in this dissertation are characterised the use of
large sparse networks, such as the Irish rural road network. By European
standards, Ireland has a substantial rural population that is dispersed
over an extensive network of rural roads. For each one thousand
population, Ireland has roughly twice as many kilometres of road as in
Belgium, Denmark, and France, and over three times as many as in Italy,
the Netherlands, and Spain (Dept of the Environment, 1999). It is
estimated that there is 87,149 km of non-national roads in the Republic of
Ireland, almost all in rural areas.
The extensive rural road network partly reflects the fact that, by western
European standards, Ireland still has an above average proportion of the
population engaged in agriculture. The current situation also reflects the
historical situation in Ireland, where in the nineteenth century the rural
population was several times its present size. This large population was
served by an extensive network of rural roads. Subsequent population
change has reduced the population size, but has left the comprehensive
road network in place. This has created a large network of rural roads
with scattered small populations living on each road segment.
Spatial Decision Support Systems for Large Arc Routing Problems
143
The road network for CARP can be examined by reference to the ratio of
arcs to nodes (sparsity) and the proportion of odd degree vertices. In
theory a network could have an arc-node ratio of 2, as in Figure 7-1. This
network also has only odd degree vertices. At the other end of the
spectrum, extremely high arc-node ratios could exist (Figure 7-2), many of
the artificially generated networks used in testing routing algorithms are
of this type. In real-world networks, the arc/node ratio lies between 2
(Figure 7-1) and around 4, the latter case would imply a uniform grid
with four-way crossroad junctions.
Rural roads give rise to a mathematical graph that differs from one
derived from urban road networks. Eglese and Li (1992) note the high
frequency of T-junctions (nodes of degree 3) in rural road networks, and
discuss the need for consideration of the network characteristics in
assessing the quality of any routing solution. The tractability of arc
routing problems is largely a function of the number of odd degree
Figure 7-1 : Network with arc-node ratio of 2
Figure 7-2 : Network with high arc/node ratio
Spatial Decision Support Systems for Large Arc Routing Problems
144
vertices. As there are few junctions with more than four incident roads,
odd degree vertices typically arise from T-junctions (nodes of degree 3)
and cul-de-sacs (nodes of degree 1). Both of these features are very
common in rural road networks.
Consequently, the Irish rural road network gives rise to a graph that is
extremely sparse compared to the networks usually encountered in the
routing literature (Figure 7-3). This arises because roads leading to
individual houses are frequently part of the public road network. If we
assume replacement of bi-directional arcs with unidirectional arcs in each
direction, the arc/node (arc/vertex) ratio is between 2.3 and 2.8. This
Figure 7-3 : Extract from Irish Rural road network
Spatial Decision Support Systems for Large Arc Routing Problems
145
implies that on average each node is connected to two or three other
nodes. As there are hundreds of nodes in a typical rural network , this
implies that only a small proportion of nodes are directly connected by a
single arc, less than 1% for a real world problem.
This sparse nature of rural road networks is in contrast to most of the
examples in the literature, which use artificially created networks that
are almost completely connected (see Section 6.1.1). Harrison and Wills
(1983) refer to Irish road networks at a regional scale; these networks are
significantly less sparse than the rural networks. Mtenzi (2000) examined
the operation of TSP algorithms on similar Irish road networks. In
Ireland the national and regional road network gives rise to a graph with
an arc/node ratio of 3.1. Zhan and Noon (1998) tested shortest path
algorithms on real road networks from the USA and found a similar
arc/node ratio of 2.68 to 3.28.
Figure 7-4 : Extract from Dublin City main road network
Spatial Decision Support Systems for Large Arc Routing Problems
146
The sparsity of urban road networks varies, depending on the road layout
in the city. Suburban areas are frequently deliberately designed with
many cul-de-sacs, which would give rise to a low arc/node ratio.
Nevertheless, the main urban road network is typically well connected,
the main road network in Dublin city has an arc/node ratio of 3.24
(Figure 7-4). North American cities laid out in a block structure could
have an arc/node ratio of close to 4. The Manhattan network in Figure 7-5
has an arc/node ratio of 3.68, reflecting the comparative lack of T-
junctions and cul-de-sacs.
7.1.2 Rural Postal Delivery
The main example used in this dissertation is that of postal delivery on
these rural networks. The author has had considerable experience, over a
ten-year period, working with An Post (Irish Post Office) on routing
problems. This work in turn benefited from experience gained as a
summer relief postman on rural routes, while the author was an
undergraduate student. While the research undertaken for this
Figure 7-5 : Extract from New York City road network
Spatial Decision Support Systems for Large Arc Routing Problems
147
dissertation was not directly conducted with An Post, the proposals for
decision support presented are directed at the type of problems faced by
that organisation. The relatively large rural population in Ireland means
that servicing this population is expensive for Irish organisations. This
differs from many other developed countries, where the rural population
is such a small proportion of the overall customer base of an organisation
as to be of little importance in the overall cost structure.
In Ireland, unlike many other countries, almost all rural households
receive a postal delivery to the doorstep. This is true even if such houses
are located at the end of cul-de-sac laneways. Rural delivery takes place
from delivery offices situated in a village or small town. Such offices
service deliveries to a small urban area (in the village/town) and a larger
rural area. A delivery office may have only a single postman, but most
have between three and six delivery routes. If the urban area is large
enough, one or more postmen may deliver on foot in that area. This
situation is not of direct interest here. However, in a practical solution
the presence of foot deliveries may mean that a van route represents only
part of the postman’s working day. Postmen are required to sort their
deliveries before setting out to deliver the day's post. Consequently, a
typical working day of eight hours would have approximately six or six
and a half hours available for travel in the delivery van. This time would
include any travel time from the post office to the first and last deliveries.
7.2 CARP on large sparse networks
7.2.1 The Large Sparse Capacitated Arc Routing Problem (LSCARP)
This dissertation uses the example of the Large Sparse Capacitated Arc
Routing Problem (LSCARP) on Irish rural road networks. The LSCARP
represents an economically important CARP on a sparse network. The
problem of visiting all houses in a given area will be a CCPP, as the roads
to be serviced will generally form a connected network. However, in some
sparsely populated districts, the arcs to be visited might not be initially
Spatial Decision Support Systems for Large Arc Routing Problems
148
connected and a CRPP is encountered. In the practical solutions
discusssed later in this dissertation we assume that all arcs are to be
visited.
Because of Ireland’s dispersed population distribution, Irish rural postal
delivery is such that few houses are serviced on each road segment (arc).
In the LSCARP the route is serviced by van, this differentiates the
problem from many others found the literature. Many arc routing
examples discuss pedestrian routes, for instance postal delivery (Bodin
and Levy, 1991) or meter reading (Stern and Dror, 1979). Other examples
in the literature discuss routing problems where the vehicles must travel
slowly to operate machinery, for instance road sweeping (Eglese and
Murdock, 1991), refuse collection (McBride, 1982) or snow clearance
(Gendreau, Laporte et al., 1997). Most of these examples are capacitated
by volume restrictions, for instance the amount of refuse that can be
carried on a vehicle.
Unlike these applications, rural postal deliveries are undertaken by vans
that can travel relatively quickly between houses. These vehicles have no
effective volume restriction, as they can carry much more mail than could
be feasibly delivered in one day. The only effective capacity restriction is
the total time for the route, including travel to the first delivery and from
the last delivery to the post office. The service time for each arc is related
to the length of the arc and the number of houses to be serviced on it. The
traversal time for an arc is related only to its length (assuming a constant
speed of travel). In general, arcs can be traversed or serviced relatively
quickly, as vehicles travel comparatively fast and there are few houses.
Consequently, the service time for each arc is typically a small proportion
of the capacity of a vehicle. A postman working a six-hour day in a van
could service more than two hundred road segments in some cases. A
typical rural delivery area might have approximately one thousand arcs
to be serviced by three or four vehicles. This means that each vehicle
Spatial Decision Support Systems for Large Arc Routing Problems
149
visits some two or three hundred arcs, which is a much larger network
than is typical in arc routing problems.
Rural postal delivery also differs from many other arc routing problems,
in that each road segment need only be serviced once, as it is possible to
service customers on both sides of the road. Many of the complications
found in urban areas do not arise on rural roads. These include problems
with crossing the road, no right or left turns, traffic restrictions, etc.
Therefore, from an algorithmic point of view, arc routing in a rural Irish
context presents a problem that is relatively free of constraints in
comparison to other real world problems. Consequently, this problem is
differentiated from most of the examples in the literature by the much
larger problems encountered and the sparsity of the networks used.
7.2.2 Time Capacitated Arc Routing Problems
TCARP represents a problem where a time rather than volume
limitation provides capacity restraints for the problem. This typically
arises in problems where volume constraints are not relevant, for
instance meter reading. Postal delivery problems may be volume
constrained if delivery takes place on foot, as the postman cannot easily
carry all the mail. This dissertation is concerned with rural postal
delivery by van, and so is a time-capacitated problem. In this class of
problem, the length of the working day is the main factor in the
clustering of distinct routes.
A simple approach to a TCARP problem is to substitute units of time for
units of volume and apply techniques devised for volume constrained
problems. However, this approach does not recognise the specific
characteristics of TCARP. In a volume-based problem the capacity of the
vehicle is used up by servicing arcs, mere transit of an arc does not add
any volume to the route. In TCARP, either traversing or servicing an arc
will exhaust the time capacity of a route. Volume based solution
procedures are likely to give good solutions to TCARP, as any technique
Spatial Decision Support Systems for Large Arc Routing Problems
150
designed to minimise distance will obviously tend to reduce arc traversals
and consequently will reduce the overall time spent in making these
traversals. Nevertheless, superior solutions can be identified by
procedures that exploit the distinct characteristics of TCARP.
7.3 Modified bounds for LSCARP
7.3.1 Lower bounds for the LSCARP
The complexity of CARP precludes the use of optimal techniques for large
problems, such as those arising in rural postal delivery. The absence of
optimal solutions makes it difficult to assess the efficiency of heuristic
approaches. Lower bounds provide such a basis for comparison. In this
section we propose a lower bound procedure for CARP that is based on
those in Chapter 5, in particular on the LB1 and LB2 bounds (Benavent,
Campos et al., 1992) in Sections 5.5.3 and 5.5.4. Our revised bounds
incorporate some improvements on LB1 and LB2. The following sections
propose some modifications specifically directed at time based problems.
Section 7.3.4 identifies a modification to the way LB2 generates
successive cutsets, this change should improve the bound generated for
both volume and time based problems. Section 7.4 provides
computational examples of the revised lower bounds. These bounds are
used subsequently as the basis of comparison for the heuristic techniques
for the LSCARP discussed in this dissertation.
7.3.2 Depot based lower Bounds for TCARP
In this section, we propose modified versions of the NDLB (Saruwatari,
Hirabayashi et al., 1992) and the LB1 and LB2 bounds (Benavent,
Campos et al., 1992) discussed in Chapter 5. These modifications reflect
the differences between time and volume constrained problems. These
bounds, which we will call TNDLB, TLB1 and TLB2 respectively, provide
bounds appropriate to time constrained problems.
Spatial Decision Support Systems for Large Arc Routing Problems
151
The original NDLB and LB1 approaches calculate bounds by adding
additional traversals at the depot to reflect the fact that multiple vehicles
must travel along the arcs leading to the depot. The number of traversals
required is twice the number of vehicles required to service the network.
The number of vehicles required reflects vehicle capacity and the demand
in the network. In a time based problem this relationship is slightly more
complex as the vehicle capacity is consumed both by traversing and by
servicing an arc. A straightforward modification of a volume-based lower
bound could include traversal time in service time and calculate a bound
on this basis, using time units instead of volume units. A time-based
approach needs to recognise that vehicle capacity is also consumed by arc
traversals generated by the matching procedure.
The modified version of NDLB, known as TNDLB, calculates the number
of vehicles required by reference to the time taken to traverse all arcs, to
service all arcs and to make any additional traversals needed to generate
an Euler tour. The time based bound TLB1 contains similar
modifications. In some cases, the calculation of the lower bound might
indicate that additional traversals are required, and this might entail a
higher number of vehicles. In this case, the bound is recalculated based
on the increased number of vehicles.
The time modified bounds, TLB1 and TNDLB, can be expected to reflect
the relationship between their original counterparts, where LB1 was
superior to NDLB (see Section 5.5.3). This implies that TNDLB is not
especially useful for comparison purposes. However, as we discuss below,
the development of TNDLB allows the use of a modified version of the
optimal branch and bound procedure based on that bound (see Section
6.2.2). Li and Eglese (1992) proposed a bound similar to TLB1, which
they called Time Constrained Lower Bound (TCLB). In addition to
modifications for time constraints, this bound improves on LB1 by
incorporating some changes in the method of calculating the paths to the
depot. However, TCLB only examines cuts at the depot and we believe
Spatial Decision Support Systems for Large Arc Routing Problems
152
that in large sparse networks it is important to use a bound that
examines cuts away from the depot. Consequently, in this research this
dissertation has concentrated on developing a time-based bound based on
those bounds that examine parts of the graph away from the depot.
Further examination of the implications for cuts away from the depot of
the ideas introduced by Li and Eglese would constitute a useful extension
of the work presented in this dissertation.
7.3.3 Bounds for TCARP based on the entire graph
In Section 5.5.4 we saw that the LB2 bound improves on LB1 by looking
at any additional traversals needed at all points in the graph, and not
just at the depot. The additional traversals needed are assessed at a
number of cuts in the graph distant from the depot. For each cut in the
graph, the algorithm identifies the number of vehicles required to service
the area beyond the cut, and LB2 then calculates any additional
traversals required for this number of vehicles crossing the cut. LB2
represents an effective bound, which can be calculated in reasonable
time. As LB2 was originally designed with volume constrained problems
in mind, the calculation of the number of vehicles crossing a cut is made
by comparing the volume to be serviced beyond the cut and the capacity
of each vehicle. The calculation of the bound for volume constrained
problems assumes that the vehicle travels to the cut without servicing
any arcs along the way. Therefore for volume constrained problems the
entire capacity of the vehicle is available to service arcs beyond the cut.
In deriving a modification of LB2 for TCARP, alterations are made to
both the calculation of the vehicle capacity and the calculation of the
capacity required to service beyond the cut. In TCARP when a vehicle
reaches a distant part of the graph, its potential capacity will have been
reduced by the time taken to travel to and from the depot. Therefore, the
available capacity of a vehicle depends on where in the graph the cut is
located. The capacity needed to service beyond the cut must include both
the service time and any unavoidable traversal time beyond the cut. Both
Spatial Decision Support Systems for Large Arc Routing Problems
153
of these amendments to the basic LB2 tend to increase the number of
vehicles required. The increased number of vehicles leads to an increased
number of traversals of the cut and therefore provides a tighter lower
bound for TCARP than the simple substitution of time for volume.
The original LB2 calculations assume that the number of vehicles
crossing the cut is known. For the time-based problem, traversal times
beyond the cut must be calculated exactly in advance to establish the
effective vehicle capacity. Identifying the traversals requires a
computationally intensive matching algorithm, a process that could make
TLB2 much slower than LB2. An exact approach would mean that a
matching procedure would be used to calculate the number of vehicles
and then a second MCPM procedure used to calculate the bound itself. In
order to avoid calling the MCPM procedure twice, an estimate of the
traversal time beyond the cut is used. This provides an estimate of the
number of vehicles required, which can be revised when the full MCPM
calculations take place as part of the lower bound calculations. If the
estimate proves to be incorrect the number of vehicles can be revised and
the lower bound calculated a second time. If this occurs infrequently, the
use of an estimate is less computationally intensive than using a MCPM
at each iteration to calculate an accurate number of vehicles.
In TLB2 a MCPM is performed initially on the entire network, this
provides a measure of the additional traversals required in the
uncapacitated problem. The uncapacitated problem provides a lower
bound for the traversal time on the network as a whole. A ratio between
the time taken to visit all network arcs and this additional traversal time
is calculated. This ratio is then used to estimate the additional traversal
time for any cutset in the network. This ratio will differ somewhat from
network to network depending on the proportion of odd degree vertices.
At each cut, the vehicle capacity for that cutset is defined as the original
vehicle capacity (in time units), less the time taken to travel from the
cutset to the depot and back again. For TLB2, this reduced vehicle
Spatial Decision Support Systems for Large Arc Routing Problems
154
capacity is compared with the estimated traversal time and total service
time for arcs beyond the cut. From this calculation, the number of
vehicles is derived and this is used in the lower bound calculations, which
are otherwise similar to those for LB2.
7.3.4 Modification to LB2 cutset strategy
As described in Section 5.5.4 above, LB2 generates a cutset by taking all
nodes adjacent to nodes in the existing U to form the set U′. These nodes
are then added to those in U to form a new set U at each iteration. The
resulting cut is formed by including all arcs from the existing graph
which connect the new U set and the component, V′, formed by removing
the nodes in U from the original graph. This adds a complete set of
adjacent nodes to U at each iteration, thereby ignoring a large number of
possible cutsets. Our proposed bounding techniques incorporate a
modification to LB2 that adds nodes individually to the set U,
consequently generating a larger number of cutsets. This larger number
of cutsets may contain a superior solution to that obtained with the more
computationally efficient LB2. This strategy was first tested by two
MMangtSc students working with the author (Breslin and Keane, 1997)
and was found to improve the LB2 bound in some circumstances.
As the set U′ generally contains more than one node, the revised selection
procedure must choose between these (Figure 7-6). The revised procedure
selects nodes from U′ in increasing order of degree and adds them to U.
The selected node is marked to ensure that it will not be considered again
and an iteration of LB2 is carried out as before. The unmarked node in U′
U ′′′′
U
Figure 7-6 : Multiple nodes in U′′′′ at each iteration
Spatial Decision Support Systems for Large Arc Routing Problems
155
with the next lowest degree is selected and included in U. This procedure
is repeated until all nodes from U′ are included in U. Once all the nodes
from U′ are included in U, then a new U′ is identified in the same manner
as used in the original LB2, where U′ is the set of nodes adjacent to nodes
in U. The calculation of L1 remains similar to the approach used in LB2,
L1 is recalculated only when all of the nodes in U′ have been added to U.
7.4 Computational Results for TLB2
7.4.1 Example of new cutset strategy
The modifications to LB2 can be illustrated by the example shown in
Figure 7-7. This example is used as the degree of the nodes in each
successive set U′ are not all the same. The first cut in Figure 7-7 is Cut
(1-2, 1-3). The basic LB2 calculations would use this as the basis of
calculating a lower bound and would then proceed by adding all adjacent
nodes to the set V at each iteration (Table 7-1). Our alternative strategy
can be seen by looking at the graph in Figure 7-8. At the initial step we
select Node 3 as having the lowest degree and include it in U.
2
3
5
4
6
7
9
8
1 2, 7
2, 4
3, 7
5, 6
5, 9
4, 5
4, 5
7, 9
3, 5
2, 5
2, 5
3, 5
3, 7
Cut 1
Cut 2
Cut 3
Cut 4
Traversal Time , De mand
Figure 7-7 : Original LB2 cutsets
Spatial Decision Support Systems for Large Arc Routing Problems
156
Therefore U′ = {i ∈ V | i is adjacent to a vertex in U}. U = U ∪ Node b,
such that Degree[b] = Min(Degree[i]) where i ∈ U′ and i is not already
included in U.
For the example in Figure 7-7, the iterations will add nodes one at a time
in increasing order of connectivity and will produce sets U, U′ and V′ as
in Table 7-2 and Figure 7-8.
Table 7-1 : The sets at each iteration using the original LB2
procedure
IterationNumber
U U′′′′ V′′′′
1 {1} {2, 3} {2, 3, 4, 5, 6, 7, 8, 9}
2 {1, 3, 2} {4, 5} {4, 5, 6, 7, 8, 9}
3 {1, 3, 2, 4, 5} {6, 7,9} {6, 7, 8, 9}
4 {1, 3, 2, 4, 5, 7, 6, 9} {8} {8}
2 3
5 4
6 7
9 8
1 2, 7 2, 4
3, 7 5, 6 5, 9
4, 5 4, 5 7, 9
3, 5
2, 5 2, 5 3, 5
3, 7
Cut 1
Cut 2
Cut 4
Cut 1a
Cut 2a
Cut 3a
Traversal Time , Demand Cut 4a
Cut 1b
Cut 2b
Figure 7-8 : Selecting Nodes one at a time, in increasing order of
connectivity
Spatial Decision Support Systems for Large Arc Routing Problems
157
7.4.2 Simple Computational Example for TCARP
This example is based on the graph shown in Figure 7-8 and uses the
cutset selection modification discussed in Section 7.2.2 above. The first
iteration adds node 1 to set U. The bound TLB1 can then be calculated,
based on the number of vehicles required to enter and leave the depot.
Subsequent iterations examine cuts away from the depot, providing our
bound TLB2. In this example, we use a vehicle capacity of 50 minutes.
The overall network has a total of 9 nodes, 13 bi-directional arcs, a total
travel time of 45 minutes and a service time of 79 minutes (including the
travel time for each arc).
TLB1 starts with an estimate of the number of vehicles, found by adding
the total time required to traverse the network arcs and the additional
traversal time required for the arcs added by the MCPM. In this case,
only nodes 2 and 4 are of odd degree, so there is an additional traversal of
Table 7-2 : The sets at each iteration using the modified cutset
selection
IterationNumber
U U′′′′ V′′′′
1 {1} {2, 3} {2, 3, 4, 5, 6, 7, 8, 9}
2 {1, 3} {2} {2, 4, 5, 6, 7, 8, 9}
3 {1, 3, 2} {4, 5} {4, 5, 6, 7, 8, 9}
4 {1, 3, 2, 4} {5} {5, 6, 7, 8, 9}
5 {1, 3, 2, 4, 5} {6, 7,9} {6, 7, 8, 9}
6 {1, 3, 2, 4, 5, 7} {6,9} {7, 8, 9}
7 {1, 3, 2, 4, 5, 7, 6} {9} {8, 9}
8 {1, 3, 2, 4, 5, 7, 6, 8} {8} {8}
Spatial Decision Support Systems for Large Arc Routing Problems
158
the arc between them, giving an additional time of 3 minutes. This figure
is used to calculate a matching ratio for use later in the calculation of
TLB2, in this case the matching ratio is 3/(45+3) = 0.063 (matching
/(traversal + matching)). The estimated number of vehicles is therefore
(79+3)/50 = 2 ((service+matching)/capacity). The remaining calculations
required for the bound are similar to those for LB1 (see Section 5.5.3).
This gives a lower bound of 7, reflecting two additional traversals to the
depot. For this vehicle capacity, TLB1 is the same as LB1, as the
additional traversal time is not sufficient to warrant the use of another
vehicle.
We tested two versions of the TLB2 bound, TLB2a and TLB2b. TLB2a
uses the same cutset procedure as the original LB2 bound. For this
example TLB2a gives a lower bound of 7, which is identical to that found
by TLB1. This outcome is not unexpected given the small size of the
Table 7-3 : TLB2 example
Iteration
Number
(Table
7-2)
Service
time
beyond
cut
Available
Vehicle
Capacity
Estimated
matchingEstimatedVehiclescrossing
cut
Actual
Vehicles
TLB2
1 79 50 2.56 2 2 7
2 75 50 2.25 2 2 7
3 68 46 1.75 2 2 7
4 55 46 1.50 2 2 7
5 46 40 0.63 2 2 8
6 36 36 0.31 2 1 8
7 27 36 0.19 1 1 8
8 10 32 0.00 1 1 8
Spatial Decision Support Systems for Large Arc Routing Problems
159
network. TLB2b proceeds by using the revised cutset selection procedure
discussed in Section 7.3.4 and selects nodes in the sequence shown above
in Table 7-2.
The operation of the algorithm is shown in Table 7-3. An estimate of the
additional traversals is obtained by multiplying the matching ratio
(0.063) by the cost of travelling the arcs beyond the cut. This is shown in
Table 7-3 in the column headed estimated matching. The estimated
matching is added to the cost of servicing the arcs to obtain an estimate of
the total time needed for the region beyond the cut. This estimated total
time is used to determine the number of vehicles needed, and to calculate
a bound, based on the LB2 calculations. The number of vehicles needed is
verified when this bound has been calculated. The operation of this
feature can be seen from iteration 6 in Table 7-3. At this point nodes
{7,8,9} are beyond the cut. Vehicles crossing this cut have a maximum
capacity of 36(50-14), this reflects the fact that vehicles must travel for a
minimum of 7 minutes (to node 5) to reach and return from the cut. There
are 36 minutes of service time beyond the cut and the estimate for the
matching required is 0.31 (0.063 * 5). This is added to the service time,
implying that two vehicles are required. In this case, each of these nodes
is of even degree, so there is no actual matching required. Consequently,
only the service time of 36 is relevant and the actual number of vehicles
required is one. The bound is then recalculated with one vehicle.
On this small example, the TLB2 bound gives a bound of 8. This offers
some improvement over the standard LB2 bound of 7 as a result of the
use of modified cutset strategy. This is a result of the modified cutset
strategy rather than the time-based alterations. In iteration four the
revised cutset procedure leads to a situation where two arcs cross a cut
and two vehicles are required to service the arcs beyond the cut. This
means that two extra traversals are added, giving a higher bound. The
original cutset strategy used in LB2 would not have examined this cut, so
the best bound would have been 7.
Spatial Decision Support Systems for Large Arc Routing Problems
160
Table 7-4 : Summary of Time based bound
units) time (in cut at vehicle ofcapacity adjustedunits) time (in vehicle of capacity
distance (over node to node from travel totime
(deadleg) servicingwithout arc traverse to time
time traversal including arc service to time
==
==′=
s
ijij
e
e
WW
djitet
et
)
Set U = { 1 }, L = L1 = L2 = 0;
∑∑
∈
∈
+=
Gee
Gee
estt
tMR
MCPM(G) (calculation of matching ratio)
WHILE U <> V DO
Let V' = V - U and G' be the graph induced by V'. Find the connectedcomponents of G'. Suppose G' has ‘k’ components G's = (V's , E's ) and
},:),({)(cutsetedge UjViEjieV ∈′∈∈==′δSet V'' = V' – U' and G'' be the graph induced by V''
Set Uc = U
WHILE U'' <> U' DO
U'':= U'' ∪ i {i ∈ V | i is adjacent to a vertex in U, i ∉ U'' }
FOR s := 1 TO k DO
)(set and LB2 tosimilar
Hgraphweightedaconstruct thenorIf
Let
},{max
min
:)({
)(
capacity) vehicle adjusted( }min{*
s
.
)(
)(e
ss
ss
ss
sests
eves
ess
ests
VEeest
is
HMPm
rS
VS
tpr
tt
tVet
MRW
t
p
tWW
s
ss
=>≠′
′=′−=
=>′∈=
+×
=
−=
′∈
′′∈∑
0
20
0
1
2 1
φ
δ
δ
δ
ENDFOR s.
U := Uc ∪ U''
},max{ 1221
LLttLL
mL
Gee
Gee
t
ss
++′+=
=
∑∑∑
∈∈
=
ENDWHILE
s
t
SstrLL ∑
=
+=1
11
Set U' = the set of nodes {i ∈ V | i is adjacent to a vertex in U }.
ENDWHILE.
Set LB2 = L2.
Spatial Decision Support Systems for Large Arc Routing Problems
161
In this example, most of the cuts identified by LB2 have three arcs
crossing the cut, implying that few additional traversals are needed. In
this extremely small network, the time modification does not provide a
superior lower bound. However, the greater number of cuts examined by
TLB2 leads to a significant improvement on the large sparse networks
discussed subsequently. The operation of the time-based bounds is
summarised in Table 7-4.
7.5 Comparison of lower bound and optimal results
7.5.1 Time based branch and bound procedure
In order to assess the lower bound procedures introduced above we used a
time-based version of the branch and bound procedure introduced in
Section 6.2.2. Our algorithm uses 32-bit Windows 95 code in Borland
Delphi and is derived from the Pascal code built in UCD by Montwill and
Naughton (1994). This code embodies a matching procedure derived from
the FORTRAN code used by Derigs (1981). Our algorithm is similar to
the original procedure (Hirabayashi, Saruwatari et al., 1992) with a
modification for time. Previously only relatively small networks have
been solved with this procedure, the largest network solved in the
original paper contained fifty demand arcs. Montwill and Naughton were
also able to solve a fifty-arc problem. However, this problem could be
solved only for the two-vehicle situation; using a larger number of
vehicles increased the number of subproblems to the point where it
became computationally impossible to solve the problem. This suggests
that time-based problems are more difficult to solve on sparse rural
networks, in comparison to the complete networks examined by the
Japanese designers of the optimal branch and bound procedure.
The difficulty of solving an arc routing problem optimally was confirmed
by our experiments, despite our use of a much more powerful computer
than that used by Montwill and Naughton. Intuitively it might seem that
the use of modern computers with more RAM might allow larger
Spatial Decision Support Systems for Large Arc Routing Problems
162
problems be solved. In fact, the exponential increase in the number of
subproblems meant that the use of more memory had little practical
effect on the ability of this procedure to solve problems of practical size to
optimality.
The branch and bound procedure starts with a complete representation of
the original graph and generates artificial nodes so that the number of
nodes in the completed graph is twice the number of undirected arcs in
the original. Each subproblem generated contains a cost matrix between
the artificial nodes, therefore a 50 arc problem would have 100 nodes and
would have a 100 x 100 cost matrix. If floating-point numbers were used
this would mean that each subproblem would typically occupy more than
60KB of RAM. This would make such a problem very difficult to solve.
For the work recorded in this dissertation, we tested this algorithm on a
Pentium III 500Mhz PC with 256MB of RAM running the Windows NT
4.0 operating system. By storing only the changes from one sub problem
to another and by using single byte storage structures we were able to
reduce the marginal storage required for each subproblem to around 100
bytes. This configuration typically allowed 750,000 subproblems be
generated before the machine ran out of RAM. This larger number of sub-
problems did not greatly improve the ability of the procedure to reach an
optimal solution, indicating that memory use remains an important
constraint on the operation of this algorithm.
The Tour Construction algorithm uses the NDLB lower bound, which is
inferior to the LB2 bound (see Section 5.5.2). This is potentially
important, as the algorithm is generally unable to find a solution when
the solution is above the lower bound. It would be quite difficult to devise
a branch and bound procedure based on the LB2 approach and beyond
the scope of this dissertation. The systematic approach of the branch and
bound technique is likely to provide a good feasible solution and therefore
to provide a useful basis for comparison for the lower bounds, even if the
optimal solution has not been identified.
Spatial Decision Support Systems for Large Arc Routing Problems
163
number of nodes 27
number of demand arcs 35
total demand (inc service) 139 minutes
total service time 47 minutes
Arc Weights (demand, cost) in minutes (rounded off to nearest integer)
Figure 7-9 : 35 arc network
Spatial Decision Support Systems for Large Arc Routing Problems
164
7.5.2 Computational experiments
In order to assess the TLB2 procedure we used a 35-arc network derived
from the Irish rural road network (see Figure 7-9). This network is still
much smaller than a real world problem, but is around the largest size
that allows the calculation of an optimal solution for some vehicle
combinations. All of the arcs in this network are demand arcs. We tested
this network with various vehicle capacities, so that from two to five
vehicles would be needed (Table 7-5). We tested TNDLB, TLB1 and TLB2
using the original LB2 cut procedure (TLB2a) and our proposed
individual arc cut procedure (TLB2b) (see Section 7.2.2). These bounds
were programmed in Borland Delphi and employ a matching procedure
derived from the FORTRAN code used by Derigs (1981).
In this 35-arc example, the results for TLB2a and TLB2b were the same.
TNDLB and TLB1 gave the same results, as would be expected from the
results reported in Section 5.5. The branch and bound procedure
identified the TNDLB solution as the optimal one where two or three
vehicles were used. Where more vehicles were used, the depot-based
bound was not optimal and the branch and bound procedure failed to
converge on a solution.
Table 7-5 : Computational results for 35 arc network
VehicleCapacity
Numberof
VehiclesTNDLB TLB1 TLB2a TLB2b
BranchAnd
Bound
80 2 16 16 16 16 16
55 3 18 18 18 18 18
45 4 20 20 22 22 24#
35 5 22 22 26 26 28#
TLB2a : arc cut strategy of original LB2TLB2b : individual arc cut strategy# best solution when machine ran out of memory
Spatial Decision Support Systems for Large Arc Routing Problems
165
The difference between TLB1 and TLB2 can best be seen if a larger
number of vehicles is used. For instance if five vehicles are used, each
with a vehicle capacity of 35. TNDLB and TLB1 give a value of 22 for the
cut at the depot alone. This reflects the fact that five vehicles (ten trips)
must pass along three arcs (1-10), (1-13) and (1-9). Consequently, seven
additional traversals are needed (Table 7-5).
TLB2 gives a higher bound for the cut shown in Figure 7-10, across arcs
(10-11), (10-14), (9-5), (9-2), (13-25) and (13-27). At this cut the vehicle
capacity is reduced to 33, after deduction of the time taken for vehicles to
travel to and from the cut. The path taken to reach the cut is
unimportant in this network as arcs 1-9, 1-10, and 1-13 each require one
minute to transit. In this case, node 11 and node 5 form two separate
components and the algorithm identifies the additional traversals of arcs
(10-11) and (9-5). The rest of the network is reached along arcs (10-14),
(9-2), (13-25) and (13-27). Five vehicles are still needed to service the arcs
beyond the cut, so six additional traversals are needed along these four
arcs. The algorithm calculates these extra traversals as 17 minutes,
which is added to the two extra minutes incurred traversing arc (10-11)
and arc (9-5). The TLB2 procedure then adds an estimate of the
additional traversals between the cut and the depot. In this example, this
is 7 minutes, reflecting the fact that 5 vehicles make 10 trips (inward and
outward) along only 3 arcs. The TLB2 bound is therefore 26, which
provides a much superior bound to TLB1 in this case.
1
95
10
11
13
27
25
14
2
Figure 7-10 : Maximum TLB2 cut for vehicle capacity of 35
Spatial Decision Support Systems for Large Arc Routing Problems
166
7.5.3 Implications of computational results
The above examples indicate that the TLB2 bound is much superior to
the TLB1 bound and the both outperform the TNDLB approach. Later in
this dissertation, we look at larger real world network examples. These
larger networks cannot be feasibly solved to optimality. In this situation
lower bounds provide the only means of assessing the quality of a
heuristic solution. The following chapter introduces heuristic approaches
to TCARP. In Chapter 9 these approaches are applied to real world
networks and their performance is compared to the TLB2 bound
discussed in this Chapter.
Spatial Decision Support Systems for Large Arc Routing Problems
167
Chapter 8 : Solution procedures for TCARP
8.1 Introduction
8.1.1 Solutions for TCARP
In the previous chapter, we introduced a time-based formulation and
proposed modified lower bounding procedures to make them suitable for
time capacitated routes. The objective of this dissertation is to examine
large networks. Currently available optimal procedures, such as those
discussed in Section 6.2, do not solve these problems to optimality; only
heuristic approaches are likely to be successful. This chapter discusses
these heuristic techniques.
Solution procedures for TCARP can be assessed by comparison with the
lower bounding methods introduced in the previous chapter. However, in
the context of the focus on decision support in this dissertation, any
algorithmic techniques must also be evaluated with respect to the real
world requirements of the problem. In a DSS context, the output from
routing models is subject to further manipulation by the decision-maker
(see Section 2.2.1). From a DSS point of view the mathematical objective
function used in the model is only one dimension of the quality of the
solution.
8.1.2 Algorithm requirements for the LSCARP
A number of specific decision support issues arise in the context of
routing postal deliveries on Irish rural networks. Customer service is
improved if deliveries are completed early in the day. This may be
especially important for particular businesses or other premises on the
route. However, by definition, some of the customers must be visited at
the end of the route. In the context of an eight-hour day, this means that
deliveries take place in the afternoon. While customers are unlikely to be
Spatial Decision Support Systems for Large Arc Routing Problems
168
happy about this late delivery, they are more likely to accept it if it
appears to be an inevitable consequence of geography.
One component of user acceptability is the perception of the overall shape
of the route and whether it appears to be ‘logical’. Human perception of
route design is largely a spatial one; if a route appears to be a logical
shape, then it will be acceptable to all parties involved, including drivers
and customers. Practical routing projects often note the loss of goodwill
that results from the fact that routes do not appear to be logical (Bocxe
and Tilanus, 1985). Some solutions generated by automated procedures
meet this requirement better than others. A second concern for
acceptability is that deliveries are completed as early as possible. Service
considerations dictate that if a route passes along a road more than once,
then the deliveries should occur on the first pass and that deadleg
traversals of the road should occur later in the day. Customers do not
wish to see a postal van passing their door early in the day, if they have
to wait until afternoon for their deliveries.
In the context of the algorithms discussed in this chapter, the overall
spatial distribution of the route is of interest. A spatially compact route is
more likely to meet service needs and is easier for the decision-maker to
manipulate in the DSS. Routing techniques need to be assessed with
respect to the ease of user intervention in the process. A facility for user
intervention will allow any specific circumstances be taken into account.
8.2 Heuristic approaches
8.2.1 CARP solution techniques
As previously discussed in Chapter 6, three basic approaches to solving
CARP exist. Many arc routing procedures have been based on Single-Pass
heuristics where a single iteration provides both a clustering of arcs and
a sequence for each route (Section 6.1.1). An alternative is to separate the
clustering and routing phases, using either RFCS approach (Section
6.1.2) or the CFRS approach (Section 6.1.3).
Spatial Decision Support Systems for Large Arc Routing Problems
169
8.2.2 Route-First Cluster-Second approach
In the context of LSCARP the RFCS approach generates a large Euler
tour. This requires that odd degree vertices are initially matched. In a
multi-vehicle problem this large route is infeasible, as it is longer than a
single vehicle can service. This large infeasible route is then split up into
routes within the maximum service time for the vehicle. In this approach,
a matching algorithm is initially used to make the network Eulerian.
Within this matched network, a large number of possible Euler tours
exist. Consequently, the large infeasible tour needs to be suitable for
splitting into smaller tours. Therefore, a number of heuristics are used
when building the large tour to ensure that suitable routes result from
the splitting of this large tour. The large Euler tour may be decomposed
into sub-clusters or cycles, and are then combined into new feasible
routes (Eglese, 1994).
8.2.3 Cluster-First Route-Second Approaches
CFRS approaches follow an intuitively appealing sequence of first
choosing the grouping of the routes and then dealing with the sequence to
be followed within each route. This seems especially appropriate for arc
routing where the Euler tour provides an optimal sequence within a
route. Cluster-first techniques are widely used in point based vehicle
routing, where the technique is seen to provide compact routes. Initial
clustering would seem to address the requirement in the LSCARP for
compact routes that are easily understood. We examined this approach
for LSCARP as an alternative to the less successful RFCS discussed
below in Section 8.3.
The objective of clustering techniques in routing is to achieve a compact
and easily routed grouping of arcs or nodes. Vehicle routing problems
often employ a coordinate-based approach, for example that of Gillett and
Millar (1974) discussed in Section 6.1.3. In the context of arc routing on
sparse networks, the coordinate based approach would appear to be
Spatial Decision Support Systems for Large Arc Routing Problems
170
inappropriate. In this context, we sought to retain some of the motivation
underlying other clustering techniques, but to employ a method that took
full account of the actual road layout.
Our investigations revealed the CFRS method to be the most appropriate
approach and we discuss in detail two techniques based on this approach.
We also discuss our less successful experiments with the RFCS approach.
8.3 Route first, cluster second algorithm
8.3.1 Tour construction
In our RFCS algorithm, phase one uses a matching algorithm to allow a
single vehicle Euler Tour to be generated. The algorithm proceeds by
adding arcs to the Euler tour, on the matched network, giving preference
to demand arcs. The route is built in a direction generally moving away
from the depot. This phase continues until approximately half of the
route time of a vehicle has been allocated. In the second phase of the
algorithm, we attempt to keep the algorithm in the region of the most
distant point routed in phase one. This attempts to ensure that routes are
reasonably compact. A third phase begins when the route has completed
more than three-quarters of the route time for a single vehicle. In this
phase, the large route makes its way back towards the depot. When the
route time for a single vehicle is exhausted, the algorithm goes back to
phase one (Table 8-1).
At the end of the routing process, when the large route has returned to
the depot, checks are made to ensure that all sections of the network have
been routed. If the large route returns prematurely to the depot, the
addition of these unallocated arcs may affect the quality of the routes
generated. When the large route is completed, it is subdivided into routes
within the vehicle capacity. While the multi-vehicle solution will
inevitably introduce additional traversals, the design of the large tour
was such that these should not add unnecessarily to the solution.
Spatial Decision Support Systems for Large Arc Routing Problems
171
8.3.2 Computational example for the RFCS algorithm
The operation of the RFCS algorithm can be demonstrated using the
(extremely simplified) network in Figure 8-1. Assume that the vehicle has
Table 8-1 : Pseudo code for RFCS approach
Matchnodes(network)
Currpoint := 0;
While all arcs not routed do
If currpoint < vehcapacity*.45 then
AddEulerOutwards
If currpoint > vehcapacity*0.45 and < vehcapacity*0.75 then
AddEulerRegion;
If currpoint < vehcapacity*.75 and vehcapacity <= 1 then
AddEulerToDepot;
If currpoint > vehiclecapacity then currpoint := 0;
EndWhile;
Splitroutes(routenet);
For i := 1 to numroutes do
Matchnodes(routenet[i];
Eulertour(routenet[i]
Endfor
Table 8-2 : Infeasible route for RFCS procedure
Sequence of nodes visited
1-S-2-S-4-S-5-S-2-T-1-S-3-S-6-S-8-S-9-S-6-S-10-S-12-T-10-T-6–T-3-S-7-S-11-T-7-T-3-T-1
T = traversal of arcS= service of arc
Spatial Decision Support Systems for Large Arc Routing Problems
172
a capacity of 24 time units. The first step is to perform a MCPM on the
network, adding additional arcs between the nodes of odd degree. This
adds additional traversals for arcs (1-2), (1-3), (3-6) (3-7) (7-11), (6-10) and
(10-12), giving an uncapacitated solution time of 62. A large route is
devised on this matched network, arbitrarily starting at arc 1-2. The
infeasible route generated is shown in Table 8-2. This route is then
subdivided into feasible routes. Route 1 in this solution is the same as in
the other approaches. Route 2 is then built from the remaining arcs on
the large infeasible route until a point is reached where the vehicle must
return to the depot. This means that route 2 can visit arc (6-10) and then
return to the depot, giving a route of exactly the vehicle capacity of 24
units. However arc (10-12) is not visited, leaving this isolated arc to be
visited by another route. The final route visits this arc and the remaining
arcs, giving a route of 23 units (Table 8-3).
1,3
2,4
1,3
1,2
2,4
1,2
2,3
2,3
2,5
2,4 2,5
1,3
4,6
9
12
8
6
10
11 7
3
1
5
4
2
(cost, demand)
Figure 8-1 : Simplified network for heuristic examples
Spatial Decision Support Systems for Large Arc Routing Problems
173
8.3.3 Evaluation of the RFCS algorithm
This strategy should ensure that the routes are reasonably compact and
that the split points on the large infeasible route occur close to the depot.
In the above example the heuristic worked fairly well, but inefficiencies
can result from particular network configurations. Any amendments to
overcome this deficiency would need to recognise the structure of the
network.
8.4 Tree based Approach to clustering for the LSCARP
8.4.1 Clustering on rural road networks
Our first CFRS is based on observation of the characteristics of the
problem. A typical LSCARP features deliveries from a small town or
village, this is typically a central point on the local road network.
Generally, five or six roads will converge at a village, providing a well-
connected node on the road network. Deliveries radiate out along these
roads to the more sparsely connected rural road network, where
deliveries take place. Any algorithm designed to cluster routes needs to
recognise these characteristics in the road network (see Figure 8-2). A
network constrained equivalent to the coordinate-based approaches is to
regard delivery routes as a tree routed at the depot node (in the village).
Recent work by New Zealand based researchers (Basnet, Foulds and
Table 8-3 : Routes generated by RFCS algorithm
Route Sequence of nodes Volume
1 1-S-2-S-4-S-5-S-2-T-1 222 1-S-2-S-3-S-6-S-8-S-9-S-6-S-10-T-6–T-3-T-1 243 1-T-3-T-6-T-10-S-12-T-10-T-6-T-3-S-7-S-11-T-7-T-3-T-1 23
T = traversal of arcS= service of arc
Spatial Decision Support Systems for Large Arc Routing Problems
174
Wilson, 1999) has also recognised the importance of the tree-like
structure of rural roads in designing clusters for milk collection routes.
8.4.2 The shortest path tree clustering algorithm
This first phase of our algorithm starts by identifying the shortest path to
every node in the delivery area. We identify nodes near the depot where
some of the shortest paths diverge. These are known as split nodes. The
next phase identifies the set of nodes whose shortest path passes through
a split node. A matching is performed on the set of arcs linking these
nodes. This provides a measure of the service time required to visit this
set of nodes. If this is less than the capacity of a route then this set of arcs
forms the basis of a cluster for one route. If the service time required
exceeds the capacity of a single vehicle, then split nodes further away
from the depot are examined. For more distant split nodes the aim is to
identify clusters of approximately one hour. When such clusters have
been defined, adjacent clusters are combined to give a cluster close to the
size of the vehicle. An abbreviated pseudocode version of the algorithm
can be seen in Table 8-4.
Figure 8-2 : Treelike structure of rural road network
Spatial Decision Support Systems for Large Arc Routing Problems
175
8.4.3 Computational example of tree clustering algorithm
In this example, we again use the simple network in Figure 8-1, with a
vehicle capacity of 24 time units. In the first phase, shortest paths are
calculated to each node. The nodes nearest the depot are examined. The
shortest paths from nodes 8, 9, 12, 10, 6, 7, 11 pass through node 3. The
shortest paths from nodes 5, 4 pass through node 2. Beyond node 2 there
are only even degree nodes, so no matching arcs are required.
Consequently, to service beyond node 2 requires a service time of 3+4+5 =
12. A vehicle travelling to node 2 could service arc (1-3), with a service
time of 6 and must travel along arc (1-3) on its return journey, adding an
additional traversal of 4. This indicates a total time of 22, which is close
to the vehicle capacity.
Table 8-4 : Pseudo code of tree clustering algorithm
While Not all arcs routed do
Next splitnode ∈ U
Calculate_cluster_routetime
While cluster-routetime > minclustersize do
SubCluster
EndWhile
While cluster-routetime < vehicle_capacity do
Amalgamate Subclusters
Endwhile
Buildroute
Endwhile
Spatial Decision Support Systems for Large Arc Routing Problems
176
We now look at the other split node. Beyond node 3 there is a total service
time of 25 (3+3+2+3+4+2+5+3). There are also additional matching
traversals of arcs (3-7) (7-11) (6-10) and (10-12). Therefore the total time
required is 32 (25 + 7), which exceeds vehicle capacity. Therefore we move
to node 6, which now becomes a split node. Beyond this node, there is
total service time of 14 and traversal time of 3. If we look at the arcs
linking node 6 to the depot, we see that these arcs cannot be serviced
within the vehicle capacity. Even if we do not service these arcs, vehicles
must at least travel from the depot to and from node 6, adding an
additional traversal of 6. This is within the vehicle capacity, so the region
beyond node 6 can form a cluster, with the arcs between node 6 and the
depot being traversed but not serviced by this route.
At this point, a third route is generated to service these arcs and arcs (3-
7) and (7-11). This vehicle has a total service time of 15 and additional
traversals of 6. This gives a total route time of 21. Consequently, in this
example the shortest path tree clustering approach provides a solution
with three routes of 21, 23 and 22 units respectively (see Table 8-5 and
Figure 8-3 below). This is in fact the optimal solution for this problem.
Table 8-5 : Routes from Tree Clustering approach
Route Sequence of nodes visited Volume
1 1-S-2-S-4-S-5-S-2-T-1 222 1-T-3-T-6-S-10-S-12-T-10-T-6-S-8-S-9-S-6-T-3-T-1 233 1-S-3-S-6-T-3-S-7-S-11-T-7-T-3-T-1 22
T = traversal of arcS= service of arc
Spatial Decision Support Systems for Large Arc Routing Problems
177
8.4.4 Evaluation of shortest path tree clustering approach
This approach meets many of the requirements for achieving a practical
solution for LSCARP. Routes are compact in keeping with user
expectations, this also simplifies sorting of the mail before the delivery.
The clustering reflects the actual configuration of the road network.
Unlike a coordinate-based approach, a tree-based technique is likely to
perform adequately even in unusual situations where a river or sea inlet
unbalances the road network. As far as possible different vehicles serve
roads leading out of the village where the depot is located. This means
that there are few unnecessary traversals on the approach to the depot,
and this is likely to provide a near optimal solution.
1,2
1,2 9
12
8
6
10
11 7
3
1
5
4
2 Route 1
Route 2
Route 3
Figure 8-3 : Routes derived from clusters
Spatial Decision Support Systems for Large Arc Routing Problems
178
Inevitably, this approach also has some limitations. As it combines
relatively large clusters together, it does not fill vehicles very precisely.
This may not be entirely inappropriate in a DSS context, as the user can
refine the solution interactively. In some cases, routes can more closely
match vehicle capacity, but only if non-adjacent clusters are joined. This
compromises the compact nature of the routes. In addition, while the user
can easily adjust the route at the edges, it could be argued that this
clustering approach does not facilitate radical re-sequencing of the route.
This may be needed to meet non-geographic constraints, and this type of
flexibility is inherent in the DSS approach.
8.5 Insertion heuristic for clustering for the LSCARP
8.5.1 Justification for using insertion procedure
The clustering techniques discussed above are influenced by those found
in vehicle routing problems, an alternative approach might be derived
from techniques specifically developed for arc routing problems. The arc
routing techniques discussed in Section 6.1.1 have generally been
designed for small and well-connected networks (see Table 6-1). Pearn
(1991) tested somewhat larger networks and found the best performing
heuristic to be the augment-insert algorithm. The networks on which
these results were achieved are much smaller than those discussed here,
although they have more in common with those discussed in this
dissertation than most of the other examples in the literature. A common
approach of arc routing algorithms is to attempt to identify cycles in the
network and to form routes from those cycles including the depot. We feel
that this approach is extremely difficult to implement for the LSCARP, as
routes may have hundreds of arcs and travel quite far from the depot.
Instead we have tested a simpler arc insertion algorithm that identifies
arcs in a region of arcs surrounding a seed arc, this was originally tested
by two MMangtSc students working with the author (Brady and Murphy,
1998).
Spatial Decision Support Systems for Large Arc Routing Problems
179
8.5.2 Operation of the algorithm
In the spirit of the augment-insert heuristic, our modified heuristic
begins by identifying a seed arc (Table 8-6). In this case, the seed arc is
the unserviced demand arc with the greatest total time for service and
travel to and from the depot. This approach constructs routes around
those arcs that are most difficult to service, as any inefficiency in
servicing these arcs will have a significant impact on the overall solution
efficiency. When a seed arc has been identified a demand arc is inserted
into the cluster if its distance from the demand arc is less than a certain
amount. This criterion is selectively altered to ensure that the arcs added
Table 8-6 : Pseudo code for insertion algorithm
While Not all arcs routed do
Selectseedarc
While Not converged do
converged := (abs(a-b)<tolerance)
If Not converged then
mid := a+(b-a)/2
GetCluster
If (clustersize > capacity) then
b:= mid
Else begin
a:= mid
EndIf;
EndIf
EndWhile
Topuproute
EndWhile
Spatial Decision Support Systems for Large Arc Routing Problems
180
provide a route that is close to the overall route time. If the resulting
cluster does not include the depot, then copies of the depot are added to
the route. A matching algorithm is then used to identify the appropriate
paths to the depot (this approach is similar to that used in developing
lower bounds in Section 5.5).
8.5.3 Computational example
The operation of this procedure can be seen by reference to the problem
discussed above (Figure 8-1) with a vehicle capacity of 24 (as before). We
start the process by choosing the arcs most distant from the depot, (Max
cij+d1i+d1j) in this case arc (4-5) (Figure 8-4). We then accumulate arcs
near the seed arc until we reach a figure close to vehicle capacity. In this
case, we add arcs (2-4) and (2-5) and (1-2). This gives a route time of 22
1,3
2,4
1,3
1,2
2,4
1,2
2,3
2,3
2,5
2,4 2,5
1,3
4,6
12
8
6
10
11 7
3
1
5
4
2
9
Figure 8-4 : Seed Arcs for Insertion heuristic
Spatial Decision Support Systems for Large Arc Routing Problems
181
units, the next possible arc (1-3) has a service time of 4 and cannot be
added to the route. Therefore, the first route is the same as route 1
generated previously by the tree based procedure.
We then proceed by looking at the remaining arcs to identify the next
seed arc, this gives a tie between arcs (8-9) and (10-12) (see Figure 8-4). If
we select arc (10-12) then we next add arc (6-10) and subsequently arcs
(6-8) (6-9) and (6-3). This gives a route with a total time of 24, which is
exactly the vehicle capacity (Figure 8-5).
However, this route is slightly problematic, as arc (8-9) is traversed but
not serviced by this route. This inefficiency is a consequence of the
relatively myopic criteria used to add arcs to the cluster. Arc (8-9) now
becomes the seed for third route. Demand arcs (1-3) and (3-7) are now
1,3
2,4
1,3
1,2
2,4
1,2
2,3
2,3
2,5
2,4 2,5
1,3
4,6
9
12
8
6
10
11 7
3
1
5
4
2
Figure 8-5 : Routes 1&2 using insertion heuristic
Spatial Decision Support Systems for Large Arc Routing Problems
182
added, and arcs (3-6) (6-8) and (6-9) are traversed without service. Arc (7-
11) cannot be serviced within the vehicle capacity by this route and so a
fourth route is needed (Table 8-7).
8.5.4 Evaluation of Insertion heuristic
In some ways, this approach has the opposite characteristics from the
tree-based approach discussed in Section 8.3. This approach proceeds arc
by arc, this allows vehicles be filled accurately. However this approach is
myopic and is not based on the actual network structure. Consequently,
this technique may not perform well on actual networks, as in the
example above. In its current form, this approach does not provide a
particular shape of route and this may be a problem. Nevertheless, it
could be argued that the arc selection procedure could be refined to
provide a more flexible route structure more in keeping with the DSS
approach. For instance, the user might be allowed to select the seed arc
and thereby direct the routing process.
8.6 Refinements to heuristic algorithms
8.6.1 Network simplification
The rural networks used in this dissertation are not only extremely
sparse but, for routing purposes, there is also some redundancy in the
Table 8-7 : Insertion heuristic route
Route Sequence of nodes visited Volume
1 1-S-2-S-4-S-5-S-2-T-1 222 1-T-3-T-6-S-10-S-12-T-10-T-6-S-8-T-9-S-6-T-3-T-1 223 1-S-3-T-6-T-8-S-9-T-6-T-3-S-7-T-3-T-1 214 1-T-3-T-7-S-11-T-7-T-3-T-1 12
T = traversal of arcS= service of arc
Spatial Decision Support Systems for Large Arc Routing Problems
183
road network. This leads to a situation where a number of nodes of degree
two divide a single road section into multiple arcs (see Figure 8-6). For
routing purposes, these arcs can be combined into a single arc by
combining the service and demand values for the individual arcs. This
reduces the number of arcs in the problem and might reduce the solution
time for the problem.
Many forms of CPP contain regions connected to the rest of the network
by a bridge and this property can be exploited in a problem solution
(Hamers, Borm, van de Leensel and Tijs, 1999). In a sparse rural
network, there are many cul-de-sac road sections. A route visiting these
cul-de-sacs must inevitably travel back the way it came, leading to
predictable additional traversals. These additional traversals are
foreseeable in advance if a single route visits the cul-de-sac.
Figure 8-6 : Multiple arcs in a single road section
1,2
2,3
1,2
1,2
2,4
2,4
2,3
2,4 1,2 2,3
(cost, demand)
Figure 8-7 : Complex Cul-de-sac road section
Spatial Decision Support Systems for Large Arc Routing Problems
184
Consequently, the cul-de-sac could be routed in advance, reducing the
complexity of the ultimate problem to be solved. The demand and service
time required for the cul-de-sac can be added to the arcs at the base of
cul-de-sac. This may also allow a single road section be combined into
one. For instance in the network in Figure 8-7 the arcs could be combined
into a single arc with an equivalent “demand” of 40 and a service time of
5. This can greatly simplify the network and reduce the number of the
arcs, but it effectively leads to clustering of arcs and may reduce the
flexibility of algorithms dealing with the network.
8.6.2 Route refinement
Clustering algorithms of the type described above provide routes that
approximately meet the requirements of the problem. These techniques
may provide close to optimal solutions. However, even better solutions
can be obtained by refining the solutions obtained. When all routes have
been generated, a number of instances will occur where different routes
pass along the same road section. One of these routes will service the arc
while the others will merely traverse it. Where this occurs, any of the
routes may service the arc without an increase in the overall time for the
problem. In a similar way, the sequence of an Euler tour can be altered
within a set of matched arcs without affecting the overall solution time.
Other refinements to the overall route may improve the solution time
(Hertz, Laporte and Hugo, 1999). The LSCARP solutions are
characterised by a small number of routes that typically do not greatly
overlap, so it seems probable that these techniques would only lead to
modest improvements. Some of the new general local search heuristic
techniques (Section 6.1.4) could be used to improve an arc routing
solution. A combination of clustering and refinement algorithms is likely
to prove the best strategy for practical problems.
Spatial Decision Support Systems for Large Arc Routing Problems
185
Chapter 9 : Decision Support for the LSCARP
on Irish road networks.
9.1 Specific DSS Characteristics
9.1.1 Network characteristics
The specific program of rural postal routing in Ireland has a number of
distinct characteristics (Harrison and Deegan, 1992). As outlined in
Chapter 7, the mathematical graph used in the LSCARP on rural roads
in Ireland is quite different from the networks generated from urban road
networks. This difference in network structure has implications for the
algorithms used, and requires an appropriate representation in the DSS.
This chapter discusses the design of a comprehensive SDSS for arc
routing and tests the heuristic procedure introduced in Chapter 8 on
realistic networks.
9.1.2 Management Requirements
The specific DSS discussed in this dissertation is directed at the
operations of the Irish post office rural delivery service (although without
their direct involvement). For the purposes of postal delivery, rural
customers are serviced by a large number of delivery offices (DO) situated
throughout the country. Consequently, each DO is responsible for quite a
small area. A smaller DO may be responsible for one or two routes
serving four or five hundred houses. A larger DO might serve a region
with two thousand houses and eight or nine routes.
9.1.3 Address structure
However, we must also consider an important difference in the address
structure of the postal routes in rural areas in Ireland. Urban routes in
Ireland, as is the case in most countries with a few exceptions such as
Spatial Decision Support Systems for Large Arc Routing Problems
186
Japan, are based on numbered houses along streets. The address
structure is therefore congruent with the mathematical representation of
the routing problem. In rural Ireland this straightforward relationship
between network arcs and addresses does not hold true. Addresses are
based on small spatial regions, known as townlands. A typical townland
will have between three and twenty-five houses, with an average of about
seven. This distinctive arrangement is slightly at odds with the
structures expected by computerised databases and routing software. In
Northern Ireland, the Post Office has attempted to redefine addresses
based on road names, but this has met with considerable public
resistance. As the Post Office in the Republic is more responsive to local
concerns, there is a recognition that any technology must adapt to the
address structure, rather than the other way around.
The separation between the address structure and the network
representation used for arc routing has several implications for an arc
routing DSS. Population data is collected on spatial units, rather than
any given collection of streets. This could affect any attempt to use
demographic data as the basis of an arc routing DSS. Delivery areas are
defined as a block of townlands, rather than being directly defined as a
set of arcs to be visited. The lack of easy correlation between the
addresses and the route structure could lead to a situation where two
sections of road in the same townland are located on different routes. The
division of the demand in the same townland across different delivery
routes would make postal sorting difficult. Therefore, one objective of a
routing system would be to devise routes where this did not occur. In
Figure 9-1 there are several road sections in Townland A, yet there are no
intersections there. Each of these road sections is in more than one
townland, which may have implications for how the populations
(demands) are distributed on the arcs. GIS techniques, discussed in
Chapter 4, can greatly facilitate this interaction between line and polygon
structures (Table 9-1).
Spatial Decision Support Systems for Large Arc Routing Problems
187
Townland B
Townland A
Figure 9-1 : Overlay of townland boundaries and road network
Table 9-1 : Examples of GIS operations to facilitate use of
townland structure
GIS Operation Application in rural postal routing
Data editing Splitting arcs at townland boundaries
Polygon-Line Identifying roads within townland
Distributing population over road segments
Line-Polygon Identifying townlands visited by route
Spatial Decision Support Systems for Large Arc Routing Problems
188
9.2 ROMC approach
9.2.1 Principles of the ROMC approach
The ROMC approach (Sprague and Carlson, 1982) introduced in Section
1.3.3 is one well-established approach that can provide useful guidelines
for the design of a DSS for the LSCARP. This approach is largely based
on the user interaction with the system and can be contrasted with the
more traditional OR/MS approach of describing the system in terms of
the modelling component. Our approach to designing the DSS reflects our
belief in the importance of visual interaction, reflecting some of the issues
discussed in Section 2.2.2.
We also propose a system built around the use of a GIS as a DSS
generator (Section 3.4.2). GIS provides the rich interface necessary to
implement our DSS, and contributes useful database functionality for
such a system. While our proposed system does not fully exploit the
modelling possibilities of SDSS, outlined in Chapter 4, it does indicate
how a GIS based system can enrich support for arc routing problems.
9.2.2 Representations
Figure 9-2 : Removal of irrelevant detail in interface and
algorithmic representations
Spatial Decision Support Systems for Large Arc Routing Problems
189
In a routing DSS the problem representation must allow the user access
to the problem definition and input data, in addition to the problem
solution. A variety of visualisation techniques can convey the nature of
the input data. The display of the road network is an obvious
representation for a routing DSS. For an arc routing problem, the DSS
might offer the option of highlighting cul-de-sacs or of displaying the
fundamental network with cul-de-sac sections removed (Figure 9-2
above). The values associated with each arc could be represented visually.
Population density or demand is an important factor and this can be
represented on the interface by different colours or line weights (Figure
9-3).
The solution of the routing problem is a series of routes and these can be
displayed visually. For an arc routing problem, the number of extra
traversals largely determines the quality of the solution. A customised
Figure 9-3 : Representation of arc attributes in different colours
and line weights
Spatial Decision Support Systems for Large Arc Routing Problems
190
arc routing DSS could have mechanisms for the display of deadleg
traversals on arcs (Figure 9-4) and could highlight where more than one
route passed along a section of road. Customer service considerations
could be facilitated by using different colours on screen to represent
approximate delivery times. These techniques allow easy user
understanding of the solution presented by the system and this facilitates
user intervention in the solution process.
9.2.3 Operations
In our DSS a number of fundamental operations are required. We need to
be able to assign arcs to vehicles, the clustering phase of the process. We
need to sequence the arcs being serviced on any particular route, the
sequencing phase. Finally, we need to be able easily to alter the routes,
moving arcs from one route to another. These logical operations may
require quite complex mathematical processing; a well-designed DSS will
conceal complexity this from the user, while still allowing the user
maximum control over the process.
In the case of the LSCARP, clustering procedures such as those discussed
in Section 8.2.3 can be used. For arc routing problems, an optimal Euler
tour procedure can be used for the sequencing phase. In principle, the
system might facilitate considerable user intervention in these
procedures.
Figure 9-4 : Only sections of road visited more than once are
shown
Spatial Decision Support Systems for Large Arc Routing Problems
191
9.2.4 Memory Aids
Memory aids provide additional information to the user without intro-
ducing any fundamental new problem representations. This would in-
clude the ability to query the database and examine the characteristics of
particular arcs or routes. The variety of text based information boxes
found in modern DSS software would be an example of this. Another
desirable feature in this category is context sensitive help, allowing the
user to receive relevant assistance in operating the system. Modern DSS
generators should provide customised help for all system operations,
including those added by the user. In a SDSS, modern GIS software
allows a variety of labels and text boxes to be displayed on the maps.
9.2.5 Controls
A successful DSS must be convenient to use and this can be achieved by
the use of modern interfaces. Current software features the user-friendly
features such as the use of menus, dialog boxes and toolbars. In an SDSS,
visual techniques can be used to select regions of the map that can then
be processed by the algorithmic techniques. These methods typically
include the ability to select individual road segments or junctions and
sections of the map within a square or circular area. Combined with the
other elements of the DSS the ROMC design framework allows a
comprehensive system to be developed (Table 9-2).
Table 9-2 : Summary of ROMC features
ROMC feature Feature in DSS for LSCARP
Representations Visual displays of data values
Elimination of unnecessary detail
Operations Route resequencing
Memory Aids Context sensitive access to data values
Controls Command Menus
Spatial Decision Support Systems for Large Arc Routing Problems
192
9.3 TransCAD GIS
9.3.1 TransCAD Software characteristics
The construction of a comprehensive DSS to fully implement the ideas
discussed above would necessarily require many man years of
programming and is therefore beyond the scope the this dissertation.
However some of the ideas discussed in this Chapter have been
implemented in the form of a SDSS built from a synthesis of TransCAD
GIS software and the heuristic based solution techniques discussed in
Chapter 8. The TransCAD software is well suited to this purpose as it is
designed with transportation applications in mind. TransCAD is
developed by the Caliper Corporation, and comes from the same software
family as the Maptitude and GIS+ products. These are designed to
provide substantial GIS functionality on the PC platform and have a
user-friendly graphic user interface.
While the other Caliper products are aimed at general business
applications of GIS, the TransCAD software adds features for the display
of routes and their interactive manipulation by the user. TransCAD
comes in two basic varieties, the basic version provides standard GIS
functionality and the ability to display routes, although without an
algorithms to generate such routes. The full version additionally contains
a variety of pre-programmed routing algorithms to generate routes. The
full version can be used as a very capable routing DSS for many classes of
problem.
Built-in features of the GIS allow the user to easily query the attributes
associated with each arc or to display arc attributes as different colours or
line weights (Figure 9-5). This allows implementation of some of the
mechanisms discussed in the ROMC approach in Section 1.3.3 above. The
academic version of TransCAD (version 3.0), which was employed for this
work, has essentially the same functionality as the full version (although
it is much cheaper!).
Spatial Decision Support Systems for Large Arc Routing Problems
193
9.3.2 GDK
The full and academic versions of TransCAD contain an arc routing
procedure. However the allocation approach used is based on urban street
networks and is not suitable for the rural applications discussed in this
dissertation. TransCAD provides a Geographic Developers Kit (GDK)
which allows access to almost eight hundred GIS and routing related
functions through a macro language known as Caliper Script . The GDK
allows construction of customised applications and GDK macros were
combined with Delphi programs for our work. This provides a simple
prototype of a DSS for the LSCARP, although more work is needed to
make it a useable DSS. A particular limitation is that only simple line
data was available for building the system. We did not have access to
datasets with polygon data (e.g. townlands) which would have allowed
Figure 9-5 : TransCAD interface with TransCAD arc routing data
Spatial Decision Support Systems for Large Arc Routing Problems
194
the full capability of SDSS to be demonstrated (Chapter 4). Therefore,
this dissertation has not implemented the interaction between the road
network and the townland data discussed in Section 9.1.3.
Our prototype arc routing DSS is one where the GIS software provides
the system interface and operations within the GIS call the customised
programs. This structure means that the GIS provides the interface,
rather than the alternative approach of using the programs to provide the
interface and call the GIS routines where appropriate (Keenan and
Brodie, 2000). Our customisation of the GIS was largely concerned with
interaction with the routing algorithms. A macro is used to generate data
files for use by the routing algorithm and the routing program builds the
data tables needed for the display of an onscreen route. The routing
program then calls a GIS macro using a DDE link to display the results
onscreen.
9.3.3 User Interface Features
TransCAD provides the ability to build a customised interface and the
menu structure provided is just one example. Interface customisation s
allows the users be presented with an interface that is appropriate for
their needs, concealing functionality that is no of interest for the specific
application. The main interface components available for customisation
in TransCAD are
! Dialog Boxes
! Toolboxes
! Toolbars
! Menus
In our prototype SDSS, we did not consider it necessary to build a
sophisticated interface, and therefore did not fully exploit these features
of TransCAD.
Spatial Decision Support Systems for Large Arc Routing Problems
195
9.4 Initial Computational Experiments with routing algorithms
9.4.1 Early testing of RFCS and Tree based approaches
The initial phase of our investigations sought to devise heuristic solutions
and compare their results with optimal solutions obtained using the
branch and bound procedure introduced in Section 6.2.2. It quickly
became clear that even medium size problems (around 100 arcs) could not
be solved optimally. For larger problems (more than 100 arcs), we
therefore found it more useful to compare the heuristic solution to a lower
bound. Early testing compared the algorithms to the NDLB lower bound
(Section 5.5.2) for different sizes of networks (Keenan and Naughton,
1996), some results are presented in Table 9-3. These results are
expressed in distance (KM) rather than the time units used in the
presentation of later results.
9.4.2 Assessment of initial testing
Our early work on the RFCS approach (Section 8.3) indicated that it is
inferior to the tree clustering approach (Section 8.4) in routing efficiency
and that it also provided less compact routes from a DSS viewpoint.
While Table 9-3 indicates that the RFCS approach can be marginally
better than the tree clustering approach in some circumstances, this
difference is only of the order of one percent. On those routes where the
Table 9-3 : Initial computational results
Area No of arcs Vehicles NDLB LB RFCS TreeClustering
1 649 4 269.7 288.4 292.8
2 600 3 307.1 330.2 322.5
3 483 3 168.7 183.7 184.26
4 362 3 203.8 224.8 213.8
Spatial Decision Support Systems for Large Arc Routing Problems
196
tree clustering approach is superior, the solution can be up to five percent
better, as in Area Four.
As optimal solutions proved impossible to calculate, we implemented the
superior bounds to the NDLB (Section 7.3). We also noted at this stage
that the clustering algorithms could not be easily adjusted to make full
use of vehicle capacity. Consequently, we went on to develop the Insertion
clustering algorithm (Section 8.5). This provided the possibility of testing
improved heuristics against improved lower bounds. Unfortunately, much
of this early work reported in Table 9-3 was lost owing to a computer
failure. Given the problems with the RFCS approach it was decided not to
recreate this algorithm, but to proceed instead with the CFRS procedures
and these approaches were incorporated in the DSS. The solution
procedures incorporated in our prototype DSS apply the tree clustering
algorithm and the Insertion clustering algorithm
9.5 Final Computational Experiments with routing algorithms
9.5.1 Data sets used.
In order to test the algorithms, we have derived three cases from the
rural road network. These do not correspond exactly to particular
administrative areas or postal delivery regions, but are broadly similar in
characteristics. While the road networks are drawn from real data, the
population distribution on these roads is randomly distributed to
approximate the actual population distribution. Every arc has a
population of at least one, so that every arc must be visited. If real data
were available, the algorithmic techniques could be readily applied to it.
Each region comprises a sparse road network of hundreds of nodes and
arcs. Some of the road sections in the digital map are divided into several
arcs, giving rise to nodes of degree two. Each arc has a length in metres,
and as we are working with a time based problem we calculate a travel
time based on an approximate speed of 35 KM/H. For convenience, travel
times are rounded up to the nearest integer value (minutes), so each arc
Spatial Decision Support Systems for Large Arc Routing Problems
197
has a minimum travel time of one minute. This rounding up slightly
overstates the time take to travel the network. In a realistic situation, a
more sophisticated calculation of time per arc might be used. The basic
data for each region is given in Table 9-4.
9.5.2 Case One
Case One is a region with a road network of 221 nodes and 245 arcs. Its
population of 820 implies an average population of 3.34 per arc (Figure
9-6). The depot in this region is located centrally, close to the only bridge
on a river. Manual route clustering would probably use the existence of
the river as the starting point for logical clusters. This situation often
arises in practical routing problems as delivery offices are often situated
in towns on villages located at a river crossing. This network has a
distinct structure and an efficient route is likely to exploit this structure.
The existence of the bridge is likely to lead to a higher lower bound.
The road network has many nodes of degree two and three and a network
reduction can reduce the network (Figure 9-7). Of the population of the
region, 341 live on the West Side of the river, with the remaining 479 on
the East Side of the river. This has implications for the routing heuristics
used, as an efficient route will probably stay on one or other side of the
river.
Table 9-4 : Region data
Case No of
Nodes
No of
Arcs
Population ServiceTime
(mins)
One 221 245 820 1073
Two 382 424 1211 1655
Three 508 550 1224 1828
Spatial Decision Support Systems for Large Arc Routing Problems
198
Figure 9-6 : Case One - road network with population density
Depot
Figure 9-7 : Case One - reduced network
Spatial Decision Support Systems for Large Arc Routing Problems
199
9.5.3 Case One solutions
We ran computational tests for this region using the Tree clustering
approach (Section 8.4) and the Insertion clustering approach (Section
8.5). On this network, these heuristics provided a solution with
approximately one minute of computation time (on a 500 MHz Pentium
III PC). We compare the heuristic results with lower bounds generated
using the procedures discussed in Section 7.3.
Our experiments indicated that the heuristics provided solutions that
were comparable with the lower bound. In all cases, the additional
traversals required by the heuristic solution are less than 45% above the
Table 9-5 : Case One - total time summary
Capacity (mins)
Algorithm 360 420 480
Uncapacitated 1192 1192 1192
TLB1 1196 1194 1194
LB2 Volume 1198 1198 1194
TLB2a 1202 1200 1198
TLB2b 1202 1200 1198
Tree 1227 1211 1207
Insertion 1247 1253 1254
Insertion with top-up 1248 1243 1258
Uncapacitated – lower bound with single vehicleTLB1 – depot based lower boundLB2 volume – original LB2TLB2a : time based bound with arc cut strategy of original LB2TLB2b : time based bound with individual arc cut strategyTree – Tree clustering heuristicInsertion – Basic insertion clustering heuristicInsertion –Insertion clustering heuristic with top-up to fill capacity
Spatial Decision Support Systems for Large Arc Routing Problems
200
traversal distance of the lower bound solution. When total route time is
compared, including service time, all solutions were less than 6% above
the lower bound. The tree based clustering approach provided solutions
with a total time only 1% greater than the lower bound (which itself
might not be optimal). It is clear that these procedures have the potential
to provide routes close with close to optimal efficiency.
If we examine the detail of the routes (see Appendix A), the tree heuristic
provided more efficient routes than the insertion approach. This reflects
the fact that it exploits the inherent shape of the network. However, the
insertion approach, especially with the use of the top-up procedure, made
better use of the vehicle capacity. This was especially noticeable at a
vehicle capacity of 420, when the insertion procedure with top-up was
able to solve the problem in three vehicles rather than the four needed by
the tree heuristic. Consequently, the relative efficiencies of the two
procedures appears to be partly dependent on the relationship between
overall demand and vehicle capacity.
9.5.4 Case Two
The second case is region with 382 nodes, 424 arcs and a population of
1211, giving an average population density of 3.85 per arc. The depot in
this region is situated slightly off-centre in a well-connected part of the
network (Figure 9-8).
The solutions for Case Two follow a similar pattern to the first example
(Table 9-6 and Appendix B). The tree based heuristic generated routes
with fewer additional traversals but did not always make optimal use of
the vehicle capacity. Where a capacity of 480 minutes was used, the
insertion procedure with top-up was able to fill four vehicles to 98%
capacity and have a small route of 79 minutes remaining. The tree
procedure in this case generated routes that were only 80%-90% full and
the remaining route was 242 minutes.
Spatial Decision Support Systems for Large Arc Routing Problems
201
However, for a capacity of 420 minutes (7 hours) the tree approach was
able to derive very effective routes and cover the region with only five
vehicles in circumstances where the basic insertion heuristic required six
vehicles. This indicates that the tree approach is capable of identifying
excellent clusters if they exist in the network. The insertion heuristic
with top-up also achieved a five-vehicle solution, indicating the value of
this post-processing. The gap between the heuristic solutions and the
lower bound is a little higher in this case, reflecting the larger number of
Depot
Figure 9-8 : Case Two - road network and reduced network
Spatial Decision Support Systems for Large Arc Routing Problems
202
vehicles used and that the lower bound is not as tight on a better
connected network. On this network, the TLB2b bound gives an identical
performance to the TLB2a bound. As the TLB2b bound examines more
cuts, it takes much longer to calculate, in this case thirteen minutes as
against one minute for the TLB2a bound.
9.5.5 Case Three
The third example has the largest road network with 508 nodes and 550
arcs. This region has a population of 1224, implying an average of 2.2
people per arc. This region requires a basic service time of 1828 minutes
and therefore around six routes are required. The depot is situated in a
relatively central and well-connected location, with about seventeen
minutes driving time needed to reach the outer arcs on the network
(Figure 9-9).
Table 9-6 : Case Two - total time summary
Capacity (mins)
Algorithm 360 420 480
Uncapacitated 1832 1832 1832
TLB1 1838 1836 1836
LB2 Volume 1844 1838 1838
TLB2a 1859 1848 1844
TLB2b 1859 1848 1844
Tree 1926 1931 1908
Insertion 2055 2008 2001
Insertion with top-up 2048 2001 1986
Spatial Decision Support Systems for Large Arc Routing Problems
203
Testing of this Case resulted in a similar pattern of results as the other
examples (Table 9-7 below and Appendix C). The time based lower bound
(TLB2a) provided a significantly higher bound than the volume-based
approach. The use of the modified cutset strategy in bound TLB2b, which
took one hour to calculate, did not lead to a higher bound than TLB2a.
This result is similar to the other networks and seems to suggest that the
TLB2b approach does not offer a higher bound on this type of real
network.
Depot
Figure 9-9 : Case Three - road network and reduced network
Spatial Decision Support Systems for Large Arc Routing Problems
204
The tree based routing heuristic performed well on this network, with few
extra traversals needed. However, this efficiency was achieved at the
expense of inefficient use of vehicles. In the solution for a vehicle capacity
of 360 minutes, the tree-based approach required eight vehicles. This
meant that each vehicle on average used only 80% of the time available to
it. The solutions for the tree and insertion approaches for a capacity of
420 minutes required seven vehicles, although the insertion with top-up
approach reached a feasible solution with six vehicles. These differences
reflect the existence of particular patterns in the network. One approach
to achieving a feasible solution with the minimum number of vehicles
would be to relax slightly the capacity constraints and then to move arcs
between the resulting routes to bring them within capacity. This
emphasises the need for an automated procedure to refine the clusters
generated.
Table 9-7 : Case Three - total time summary
Capacity (mins)
Algorithm 360 420 480
Uncapacitated 2139 2139 2139
TLB1 2149 2149 2144
LB2 Volume 2160 2156 2147
TLB2a 2178 2166 2158
TLB2b 2178 2166 2158
Tree 2294 2228 2234
Insertion 2363 2459 2449
Insertion with top-up 2389 2419 2466
Spatial Decision Support Systems for Large Arc Routing Problems
205
9.6 Conclusion
9.6.1 Decision Support for Routing
This dissertation is concerned with decision support for routing problems.
This represents a relatively popular type of DSS application and an
important area of OR/MS modelling research. This dissertation has
indicated that the field continues to develop, with increasing emphasis on
integrated systems that associate the algorithmic techniques with
sophisticated database and user interface features. This dissertation has
suggested that routing DSS needs to continue to take advantage of
continuing developments in IT. We identified GIS and the growth for
spatial data as one such important trend. This dissertation has suggested
that existing routing applications do not fully exploit the potential of GIS
and has argued for the better integration of OR/MS modelling techniques
and GIS.
9.6.2 Modelling issues
This dissertation has examined large sparse arc routing problems, which
is presently an under-researched area of OR/MS. Modified lower bound
procedures are proposed for this type of problem and these have been
validated for small networks against our implementation of an optimal
procedure. As it proved impossible in this dissertation to extend the
optimal approach to larger networks, we have developed three heuristic
solution procedures. One of these was initially tested, and discarded as
not being a promising line of research. We extensively tested the other
two heuristics in comparison to our revised lower bonds. This research
demonstrates that heuristic solutions can provide close to optimal
solutions for this class of problem. These heuristic solutions require a
minimal amount of computation time on a modern machine and are
therefore quite suitable as the basis for a DSS where real time solutions
are required.
Spatial Decision Support Systems for Large Arc Routing Problems
206
The proposed solutions were also examined from the point of view of their
probable user-acceptability in the context of a DSS. The two heuristics
tested had different strengths; the tree based clustering procedure offered
efficient and acceptable solutions, but did not use up vehicle capacity very
well. In some cases an extra vehicle was required for solutions generated
by this heuristic, because each vehicle was not completely filled. The
insertion heuristic did allow all of the vehicle capacity to be used, but did
not provide very efficient or compact routes. The routes generated by this
approach did not take full advantage of the geographic layout of the
network. This could lead to routes that were quite different from those
that a human scheduler would choose, such routes were not likely to be
very acceptable to users.
9.6.3 Further Research on Modelling Techniques
The work presented in this dissertation provides a solid foundation for a
successful implementation of a DSS for large sparse arc routing problems.
Further research in the field might usefully combine the tree based
clustering approach and the insertion approach. A superior synthesis of
these approaches might use an initial clustering that reflected the
network structure and that used other heuristics to extend the basic
routes to the full capacity of the vehicle.
Another approach to improving our work might be a modification in the
way that the tree-based heuristic generates routes from the depot. The
current strategy grows a tree along a single path; this neglects the
possibility of the vehicle returning by a different route from its outward
journey. There is scope for further testing of the parameters of the tree
heuristic, alternative strategies for combining neighbouring clusters
might lead to an improved performance. Greater attention to these
details might lead to a heuristic that could generate very close to optimal
solutions in a few seconds of computer time. Such an algorithm would be
ideally suited to practical applications, where strict optimality is not
essential.
Spatial Decision Support Systems for Large Arc Routing Problems
207
This dissertation presents potentially important time based modifications
to the existing volume based lower bounds. The examination of more cuts
discussed in Section 7.3.4 doesn’t seem in practice to represent an
advance in sparse networks. The graph theory based lower bounding
approaches discussed here might be further improved by reference to
other work; notably the TCLB discussed in Section 7.3.2. The enhanced
lower bounds and the optimal branch and bound procedure could be used
to provide guidelines for the design of a close to optimal heuristic. If
critical cuts in the network can be identified and if routes are designed to
pass through these critical points in an optimal way, then the end-result
will be close to the lower bound. This would provide good solutions, and
might make the use of a branch and bound approach feasible for larger
problems.
One obvious area of further development is the post feasibility
improvement of the routes. A systematic automated swapping procedure
is needed to move arcs to and from routes where two routes pass close to
each other. Such a procedure might have an important role in balancing
the routes produced and in ensuring the vehicle capacity was fully used.
A relatively straightforward post-processing approach might greatly
improve the performance of the heuristics. While conceptually simple,
such a procedure would require significant programming effort and this
work is left to future researchers.
A further requirement for a comprehensive DSS solution to this problem
is the implementation of modelling support for manual intervention in
the solution. Such intervention would include the ability to identify
manually the appropriate seed points required to direct the clustering
procedure. It should be possible to re-solve all or part of the route where
the user has made manual changes. A fully developed approach would
require the integration of this type of facility with appropriate user
interface features in a SDSS.
Spatial Decision Support Systems for Large Arc Routing Problems
208
9.6.4 Developments in SDSS
In our examination of existing applications of GIS based systems for
routing, this dissertation has found that the field is a fragmented one and
that many systems make only a trivial use of GIS techniques. This
dissertation has looked at a range of routing problems with respect to
three types of constraint; locations, paths and vehicles. This dissertation
suggests that the first two of these are inherently spatial in nature, and
that path restrictions have been given less attention in traditional
routing applications. This dissertation identified some of the interactions
that can take place between these different types of spatial parameters.
This dissertation identified the type of problems where we believe that a
SDSS may contribute. We further went on to discuss one specific
application, that of a DSS for large sparse capacitated arc routing
problems. We discussed how such a system might be developed using the
TransCAD GIS.
The specific system discussed in this dissertation reaches only a small
part of the potential of routing SDSS in general. It is the firm belief of the
author that more sophisticated SDSS will continue to be developed as
researchers realise the importance of this type of system. Routing will be
one, of many, applications to benefit from advances in SDSS design.
Spatial Decision Support Systems for Large Arc Routing Problems
209
References
1. Abraham, T. and Wankel, C. (1995). “Supporting decision support: Whereinformation on DSS is located”, Decision Support Systems 14(4): 299-312.
2. Adenso-Diaz, B., González, M. and García, E. (1998). “A hierarchical approach tomanaging dairy routing”, Interfaces 28(2): 21-31.
3. Alter, S. (1980). Decision Support Systems: Current Practice and ContinuingChallenges. Reading, USA, Addison-Wesley.
4. Amberg, A., Domschke, W. and Voß, S. (2000). “Multiple center capacitated arcrouting problems: A tabu search algorithm using capacitated trees”, EuropeanJournal of Operational Research 124(2): 360-376.
5. Angehrn, A. A. (1991). “Modelling by Example: A link between users, models andmethods in DSS”, European Journal of Operational Research 55(3): 296-308.
6. Angehrn, A. A. and Lüthi, H.-J. (1990). “Intelligent Decision Support Systems: AVisual Interactive Approach”, Interfaces 20(6): 17-28.
7. Anonymous (2000a). “Europe's Celtic Tiger turns green into gold”, GEOEurope9(10): 37-40.
8. Anonymous (2000b). “OSI: Challenging times for the nation's mapmakers”,GEOEurope 9(10): 34-36.
9. Assad, A. and Golden, B. (1982). “A categorised bibliography of survey articles inmanagement science and operations research”, Management Science 28(4): 425-438.
10. Assad, A., Pearn, W. L. and Golden, B. (1987). “The Capacitated ChinesePostman Problem: Lower bounds and solvable cases”, American Journal ofMathematical and Management Science 7(1): 63-88.
11. Assad, A. A. and Golden, B. L. (1995). Arc Routing Methods and Applications.Handbooks in OR & MS. M. O. Ball, T. Magnanti, C. Monma and G. Nemhauser.Amsterdam, North-Holland, Elsevier. 8: 375-483.
12. Avis, D. (1983). “A survey of heuristics for the weighted matching problem”,Networks 13: 475-493.
13. Barbosa, L. C. and Hirko, R. G. (1980). “Integration of Algorithmic Aids IntoDecision Support Systems”, MIS Quarterly 4(March): 1-12.
14. Basnet, C., Foulds, L. and Igbaria, M. (1996). “FleetManager: a microcomputer-based decision support system for vehicle routing”, Decision Support Systems16(3): 195-207.
15. Basnet, C., Foulds, L. and Wilson, J. (1999). “Heuristics for vehicle routing ontree-like networks”, Journal of the Operational Research Society 50(6): 627-635.
16. Beasley, J. E. (1983). “Route First - Cluster Second methods”, Omega 11(4): 403-408.
17. Begur, S. V., Miller, D. M. and Weaver, J. R. (1997). “An integrated spatial DSSfor scheduling and routing home-health-care nurses”, Interfaces 27(4): 35-48.
18. Belenguer, J. M. and Benavent, E. (1997). A cutting plane algorithm for theCapacitated Arc Routing problem, Valencia, Spain., Universidad de Valencia.
Spatial Decision Support Systems for Large Arc Routing Problems
210
19. Belenguer, J. M. and Benavent, E. (1998). “The Capacitated Arc RoutingProblem: Valid Inequalities and Facets”, Computational Optimization andApplications 10(2): 165-187.
20. Bell, P. C. (1985). “Visual interactive modelling as an Operations Researchtechnique”, Interfaces 15(4): 26-33.
21. Bell, P. C. (1994). “Visualization and Optimization : The future lies together”,ORSA Journal on Computing 6(3): 258-260.
22. Bell, W. J., Dalberto, L. M., et al. (1983). “Improving the distribution ofindustrial gases with an online computerized routing and scheduling optimizer”,Interfaces 13(6): 4-23.
23. Beltrami, E. (1974). “Networks and Vehicle Routing for Municipal WasteCollection”, Networks 4(1): 65-94.
24. Benavent, E., Campos, V., Corberan, A. and Mota, E. (1992). “The CapacitatedArc Routing Problem. Lower Bounds”, Networks 22(7): 669-690.
25. Benavent, E. and Soler, D. (1999). “The Directed Rural Postman Problem withTurn Penalties”, Transportation Science 33(4): 408-418.
26. Beroggi, G. E. G. (1994). “A real time routing model for hazardous materials”,European Journal of Operational Research 75(3): 508-520.
27. Beroggi, G. E. G. and Wallace, W. A. (1994). “A prototype decision supportsystem in hypermedia for operational control of hazardous material shipments”,Decision Support Systems 12(1): 1-12.
28. Bertsimas, D. and Simchi-Levi, D. (1996). “A New Generation of Vehicle RoutingResearch: Robust Algorithms, Addressing Uncertainty”, Operations Research44(2): 286-304.
29. Bocxe, M. A. G. and Tilanus, C. B. (1985). “Testing vehicle scheduling programsfor milk collection”, European Journal of Operational Research 20(1): 25-33.
30. Bodin, L., Fagan, G. and Levy, L. (1992) The Geomod system. The 5th AdvancedPostal Technology Conference Proceedings, Washington, USA., ProInfo
31. Bodin, L., Fagan, G. and Levy, L. (1997). Street routing and schedulingproblems, presented at EURO/INFORMS 97 conference Barcelona.
32. Bodin, L. and Golden, B. (1981). “Classification in Vehicle Routing andScheduling”, Networks 11(2): 97-108.
33. Bodin, L. and Levy, L. (1991). “The arc partitioning problem”, European Journalof Operational Research 53(3): 393-401.
34. Bodin, L. and Levy, L. (1994). “Visualization in Vehicle Routing and SchedulingProblems”, ORSA Journal on Computing 6(3): 261-268.
35. Bodin, L. D., Golden, B. L., Assad, A., A. and Ball, M. O. (1983). “Routing andScheduling of Vehicles and Crews: The State of the Art”, Computers andOperations Research 10(2): 67-211.
36. Bodin, L. D. and Kursh, S. J. (1979). “A Detailed description of a ComputerSystem for the Routing and Scheduling of Street Sweepers”, Computers andOperations Research 6(4): 181-198.
37. Bonczek, R. H., Holsapple, C. W. and Whinston, A. B. (1981). Foundations ofdecision support systems. Orlando, Academic Press.
Spatial Decision Support Systems for Large Arc Routing Problems
211
38. Bott, K. and Ballou, R. (1986). “Research Persepectives in Vehicle Routing andScheduling”, Transportation Research Part A 20A(3): 239-243.
39. Braca, J., Bramel, J., Posner, B. and Simchi-Levi, D. (1997). “A computerisedapproach to the New York City school bus routing problem”, IIE Transactions29(8): 693-702.
40. Brady, E. and Murphy, B. (1998). The Time Capacitated Arc Routing Problem:Feasible solutions for real world problems, MMangtSc Thesis, Dept. of MIS,University College Dublin.
41. Breslin, P. and Keane, A. (1997). The capacitated arc routing problem: Lowerbounds, MMangtSc Thesis, Dept. of MIS, UCD.
42. Camm, J. D., Chorman, T. E., et al. (1997). “Blending OR/MS, Judgement andGIS: Restructuring P&G's Supply Chain”, Interfaces 27(1): 128-142.
43. Chang, Y.-H., Yeh, C.-H. and Cheng, J.-H. (1998). “Decision support for busoperations under uncertainty : A fuzzy expert system approach”, Omega 26(3):367-380.
44. Chapleau, L., Ferland, J. A., Lapalme, G. and Rousseau, J. M. (1984). “A ParallelInsert Method for the Capacitated Arc Routing Problem”, Operations ResearchLetters 3(2): 95-99.
45. Chapleau, L., Ferland, J.-A. and Rousseau, J.-M. (1985). “Clustering for routingin densely populated areas”, European Journal of Operational Research 20(1):48-57.
46. Chrisman, N. (1997). Exploring Geographic Information Systems. New York, J.Wiley and Sons.
47. Christodes, N. (1973). “The Optimum Traversal of a Graph”, Omega 1(6): 719-732.
48. Clarke, G. and Wright, G. (1964). “Scheduling of vehicles from a central depot toa number of delivery points”, Operations Research 12: 568-581.
49. Clossey, J., Laporte, G. and Soriano, P. (2001). “Solving arc routing problemswith turn penalties”, Journal of the Operational Research Society 52(4): 433-439.
50. Cook, W. and Rohe, A. (1999). “Computing minimum-weight perfect matchings”,INFORMS Journal on Computing 11(2): 138-148.
51. Corberán, A. and Sanchis, J. M. (1994). “A polyhedral approach to the ruralpostman problem”, European Journal of Operational Research 79(1): 95-114.
52. Cortez, E., Meek, R. and Koger, J. (1994) GIS to support vehicle routing in publiceducation. Annual Conference of the Urban and Regional Information SystemsAssociation (URISA) Proceedings: 387-399.
53. Coutinho-Rodrigues, J., Rodrigues, N. and Climaco, J. (1993). “Solving an urbanrouting problem using heuristics and a successful case study”, Belgian Journal ofOperational Research, Statistics and Computer Science 33(1): 19-32.
54. Cova, T. and Goodchild, M. (1994) Spatially Distributed Navigable Databases forIntelligent Vehicle Highway Systems. GIS/LIS '94 Proceedings, AmericanSociety for Photogrammetry and Remote Sensing, American Congress onSurveying and Mapping: 191-200.
Spatial Decision Support Systems for Large Arc Routing Problems
212
55. Crossland, M. D., Wynne, B. E. and Perkins, W. C. (1995). “Spatial DecisionSupport Systems: An overview of technology and a test of efficacy”, DecisionSupport Systems 14(3): 219-235.
56. Dantzig, G. B. and Ramser, J. H. (1959). “The Truck Dispatching Problem”,Management Science 6: 80-91.
57. de Silva, F., Gatrell, T., Pidd, M. and Eglese, R. (1993) Spatial Decision SupportSystems for Emergency Planning. European Conference on GeographicalInformation Systems Proceedings: 1276-1285.
58. de Silva, F. N. and Eglese, R. W. (2000). “Integrating simulation modelling andGIS: Spatial decision support systems for evacuation planning”, The Journal ofthe Operational Research Society 41(4): 423-430.
59. Dennis, A. R. and Carte, T. A. (1998). “Using Geographical Information Systemsfor Decision Making: Extending cognitive fit theory to map based presentations”,Information Systems Research 9(2): 194-203.
60. Densham, P. J. (1991). Spatial Decision Support Systems. GeographicalInformation Systems, Volume 1 : Principles. D. J. Maguire, M. F. Goodchild andD. W. Rhind, Longman,. 1: 403-412.
61. Dept of the Environment (1999). Irish Department of the Environment(http://www.environ.ie), Dublin, Ireland.
62. Derigs, U. (1981). “A shortest augmenting path method of solving minimalperfact matching problems”, Networks 11(4): 379-390.
63. DeSanctis, G. (1984). “Computer graphics as decision aids: directions forresearch”, Decision Sciences 15(4): 463-487.
64. Desrochers, M., Jones, C. V., et al. (1999). “Towards a model and algorithmmanagement system for vehicle routing and scheduling problems”, DecisionSupport Systems 25(2): 109-133.
65. Ding, Y., Baveja, A. and Batta, R. (1994). “Implementing Larson and SadiqsLocation Model in A Geographic Information System”, Computers andOperations Research 21(4): 447-454.
66. Djokic, D. (1996). Toward a general-purpose decision support system usingexisting technologies. GIS and Environmental Modeling: Progress and ResearchIssues. M. F. Goodchild, L. T. Steyaert, B. O. Parkset al. Fort Collins, CO, USA,GIS World Books: 353-356.
67. Dror, M., Ed. (2000). Arc routing : theory, solutions, and applications. Boston,MA, Kluwer Academic.
68. Duchessi, P., Belardo, S. and Seagle, J. P. (1988). “Artifical Intelligence and theManagement Science Practitioner: Knowledge Enhancements to a DecisionSupport System for Vehicle Routing”, Interfaces 18(2): 85-93.
69. Edmonds, J. (1965). “The Chinese Postman's problem”, ORSA Bulletin 13: 73.
70. Edmonds, J. and Johnson, E. (1973). “Matching, Euler Tours and the ChinesePostman”, Mathematical Programming 5(1): 88-124.
71. Eglese, R. and Murdock, H. (1991). “Routeing Road Sweepers in a Rural Area”,Journal of the Operational Research Society 42(4): 281-288.
72. Eglese, R. W. (1990). “Simulated Annealing: A tool for operational research”,European Journal of Operational Research 46: 271-281.
Spatial Decision Support Systems for Large Arc Routing Problems
213
73. Eglese, R. W. (1994). “Routing Winter Gritting Vehicles”, Discrete AppliedMathematics 48(3): 231-244.
74. Eglese, R. W. and Li, L. Y. O. (1992). “Efficient Routing for Winter Gritting”,Journal of the Operational Research Society 43(11): 1031-1034.
75. Eiselt, H. A., Gendreau, M. and Laporte, G. (1995a). “Arc Routing Problems. PartI:The Chinese Postman Problem.”, Operations Research 43(2): 231-242.
76. Eiselt, H. A., Gendreau, M. and Laporte, G. (1995b). “Arc Routing Problems. PartII: The Rural Postman Problem”, Operations Research 43(3): 399-414.
77. Eom, H. and Lee, S. (1990). “Decision support systems applications research: Abibliography (1971-1988)”, European Journal of Operational Research 46(3): 333-342.
78. Eom, S., Lee, S. and Kim, J. (1993). “The intellectual structure of DecisionSupport Systems (1971-1989)”, Decision Support Systems 10(1): 19-35.
79. Eom, S., Lee, S., Somarajan, C. and Kim, E. (1997). “Decision support systemsapplications - a bibliography (1988-1994)”, OR Insight 10(2): 18-32.
80. Eom, S. B., Lee, S. M., Kim, E. B. and Somarajan, C. (1998). “A survey ofdecision support applications (1988-1994)”, Journal of the Operational ResearchSociety 49(2): 109-120.
81. Er, M. C. (1988). “Decision Support Systems: A Summary, Problems and FutureTrends.”, Decision Support Systems 4(3): 355-363.
82. Erkut, E. (1996). “The road not taken”, ORMS Today 23(6).
83. ESRI ESRI Corp (http://www.esri.com), Redlands, CA , USA.
84. Euler, L. (1736). “Solutio Problematis ad Geometrian Situs Pertinentis”,Commentarii Academiae Scientarum Petropolitanae 8: 128-140.
85. Evans, J. R. and Minieka, E. (1992). Optimization Algorithms for Networks andGraphs. New York, Dekker.
86. Fernandez de Córdoba, P., Garcia Raffi, L. M. and Sanchis, J. M. (1998). “Aheuristic algorithm based on Monte Carlo methods for the Rural PostalProblem”, Computers and Operations Research 25(12): 1097-1106.
87. Fisher, M. (1995). Vehicle Routing. Handbooks in OR & MS. M. O. Ball, T.Magnanti, C. Monma and G. Nemhauser. Amsterdam, North-Holland, Elsevier.8: 1-33.
88. Fisher, M. and Jaikumar, R. (1981). “A generalized assignment heuristic forvehicle routing”, Networks 11(2): 109-124.
89. Fisher, M. L., Greenfield, A. J., Jaikumar, R. and Lester, J. T. (1982). “Acomputerized vehicle routing application”, Interfaces 12(4): 42-52.
90. Fölsz, F., Mészáros, C. and Rapcsák, T. (1995). “Distribution of gas cylinders”,European Journal of Operational Research 87(3): 613-623.
91. Forgionne, G. A. (1999). “An AHP model of DSS effectiveness”, EuropeanJournal of Information Systems 8: 95-106.
92. Franklin, C. (1992). “An Introduction to Geographic Information Systems:Linking Maps to Databases”, Database 15(2): 12-21.
Spatial Decision Support Systems for Large Arc Routing Problems
214
93. Freckmann, P. (1993) Route Calculation for Dangerous Goods Transports with aGraphical Information System. European Conference on GeographicalInformation Systems Proceedings: 1132-1138.
94. Fung, D. S. and Remsen, A. P. (1997). “Geographic Information Systemstechnology for business applications”, Journal of Applied Business Research13(3): 17-23.
95. Gamma Geographical and Multimedia Applications Ltd(http://www.gamma.ie), Dublin, Ireland.
96. Garey, M. R. and Johnson, D. S. (1979). Computers and Intractibility: a guide tothe theory of NP-Completeness. San Francisco, Freeman.
97. Gendreau, M., Laporte, G. and Yelle, S. (1997). “Efficient routing of servicevehicles”, Engineering Optimization 28(4): 263-271.
98. Georoute (1998). GIRO Inc (http://www.giro.ca), Montréal , Canada.
99. Ghiani, G. and Improta, G. (2000). “An efficient transformation of thegeneralized vehicle routing problem”, European Journal of Operational Research122(1): 11-17.
100. Gillet, B. and Miller, L. (1974). “A heuristic algorithm for the vehicle-dispatchproblem”, Operations Research 22(2): 340-349.
101. Golden, B. L., Bodin, L. and Goodwin, T. (1986). “Microcomputer-Based VehicleRouting and Scheduling Software”, Computers and Operations Research 13(2):277-285.
102. Golden, B. L., DeArmon, J. S. and Baker, E. K. (1983). “ComputationalExperiments With Algorithms for a Class of Routing Problems”, Computers andOperations Research 10(1): 47-59.
103. Golden, B. L., Hevner, A. R. and Power , D. J. (1986). “Decision Insight Systemsfor Microcomputers: A Critical Evaluation”, Computers & Operations Research13(2): 287-300.
104. Golden, B. L. and Wong, R. T. (1981). “Capacitated Arc Routing Problems”,Networks 11(3): 305-315.
105. Gorry, A. and Scott-Morton, M. (1971). “A Framework for Information Systems”,Sloan Management Review 13(Fall 1971): 56-79.
106. Grace, B. F. (1977). “Training Users of a prototype DSS”, Data Base 8(3): 30-36.
107. Greistorfer, P. (1994). Computational Experiments with heuristics for aCapacitated Arc Routing Problem, University of Graz - Working Paper 32.
108. Grimshaw, D. J. (1996) Towards a Taxonomy of Geographical InformationSystems. 29th Hawaii International Conference on System Sciences (HICSS-29)Proceedings, Hawaii: 547-556.
109. Grimshaw, D. J. (2000). Bringing Geographical Information Systems intoBusiness, Wiley.
110. Grimshaw, D. J. and Clarke, M. (1996). “The use of spatial models to supportdecision making: an empirical investigation of financial services and retailsectors”, Journal of Targeting, Measurement and Analysis for Marketing 4(4):314-325.
Spatial Decision Support Systems for Large Arc Routing Problems
215
111. Grimshaw, D. J., Mott, P. L. and Roberts, S. A. (1997). “The role of context indecision making: some implications for database design”, European Journal ofInformation Systems 6(2): 122-128.
112. Hall, R. W. and Partyka, J. G. (1997). “On the Road to Efficiency”, Or/MS TodayOnline 24(3).
113. Hamers, H., Borm, P., van de Leensel, R. and Tijs, S. (1999). “Cost allocation inthe Chinese postman problem”, European Journal of Operational Research119(3).
114. Harding, S. M. and Wilkinson, G. G. (1996). A strategic view of GIS research andtechnology development for Europe, Ispra, European Space ApplicationsInstitute.
115. Harrison, H. (1979). “A Planning System for Facilities and Resources inDistribution Networks”, Interfaces 9(2): 6-22.
116. Harrison, H. C. and Deegan, A. J. D. (1992) A geographical database approach tothe optimization of the Irish postal networks. USPS Advanced TechnologyConference Proceedings, United States Postal Service: 981-990.
117. Harrison, H. C. and Wills, D. R. (1983). “Product assembly and distributionoptimization in an agribusiness cooperative”, Interfaces 13(2): 1-9.
118. Hertz, A., Laporte, G. and Hugo, P. H. (1999). “Improvement procedures for theundirected rural postman problem”, INFORMS Journal on Computing 11(1): 53-62.
119. Hertz, A., Laporte, G. and Mittaz, M. (2000). “A tabu search heuristic for thecapacitated arc routing problem”, Operations Research 48(1): 129-135.
120. Hirabayashi, R., Saruwatari, Y. and Nishida, N. (1992). “Tour ConstructionAlgorithm for the Capacitated Arc Routing Problem”, Asia-Pacific Journal ofOperational Research 9(2): 155-175.
121. Hobieka, A., Kim, S. and Beckwith, R. (1994). “A Decision Support System forDeveloping Evacuation Plans around Nuclear Power Stations”, Interfaces 24(5):22-35.
122. Holsapple, C. W. and Whinston, A. (1996). Decision Support Systems: AKnowledge Based Approach, West Publishing.
123. Hong, Y. and Thompson, G. L. (1998). “Finding postal carrier walk paths inmixed graphs”, Computational Optimization and Applications 9(3): 229-247.
124. Hurrion, R. D. (1986). “Visual interactive modelling”, European Journal ofOperational Research 23(3): 281-287.
125. Imielinski, T. and Navas, J. C. (1999). “GPS-based geographic addressing,routing, and resource discovery”, Communications of the ACM 42(4): 86-92.
126. IMO (1995). Geographic Information Systems in Europe : Problems andPotential, IMO Working Paper 95/2, Information Market Observatory (IMO),European Commission, Luxembourg.
127. IRIS Irish Regional Information Systems Ltd (http://www.iris.ie), Dublin,Ireland.
128. Ives, B. (1982). “Graphical user interfaces for business information systems”,MIS Quarterly(Dec. (special issue)): 15-47.
Spatial Decision Support Systems for Large Arc Routing Problems
216
129. Jansen, K. (1993). “Bounds for the general capacitated routing problem”,Networks 23(3): 165-173.
130. Jones, C. (1991). User Interface Development and Decision Support Systems.Recent Developments in Decision Support Systems. C. Holsapple and A.Whinston, Nato ASI Series, Springer-Verlag: 181-209.
131. Jones, C. (1994a). “Anchoring and Cross-Fertilization”, ORSA Journal onComputing 6(3): 278-280.
132. Jones, C. (1994b). “Visualization and Optimization”, ORSA Journal onComputing 6(3): 221-257.
133. Jones, C. V. (1999). “Visualization and Modeling”, Interactive Transactions ofOR/MS 2(1).
134. Keen, P. (1986). Decision Support Systems: The Next Decade. Decision SupportSystems: a decade in perspective. E. McLean and H. G. Sol, North-Holland.
135. Keen, P. (1998). Puzzles and Dilemmas: An agenda for value adding IS research,Keynote Address at ECIS 98, Aix-en-Provence,(http://www.peterkeen.com/Downloads/ecis.ppt).
136. Keenan, P. and Naughton, M. (1996). Arc routing for rural Irish networks.System Modelling and Optimization. J. Dolezal and J. Fidler. London, ChapmanHall.: 599-606.
137. Keenan, P. B. (1997) Geographic Information Systems : Their contribution to theIS mainstream. AIS Americas Conference on Information Systems Proceedings,Indianapolis, USA, Association of Information Systems: 691-693.
138. Keenan, P. B. (1998a). “Spatial Decision Support Systems for Vehicle Routing”,Decision Support Systems 22(1): 65-71.
139. Keenan, P. B. (1998b) Spatial Decision Support Systems: Extending thetechnology to a broader user community. Context Sensitive Decision SupportSystems Proceedings, Bled, Slovenia: 21-30.
140. Keenan, P. B. (1998c). “When the Question is 'Where'? Integrating geographicinformation systems and management science”, OR Insight 11(1): 23-28.
141. Keenan, P. B. and Brodie, S. (2000) A Prototype Web-based Carpooling System.Americas Conference on Information Systems (AMCIS) Proceedings, Long Beach,California, USA, Association for Information Systems: 362-364.
142. Kiuchi, M., Shinano, Y., Hirabayashi, R. and Saruwatari, Y. (1995) An exactalgorithm for the Capacitated Arc Routing Problem using the Parallel Branchand Bound method. 30th SSOR, The Operations Research Society of JapanProceedings, Japan: 18-23.
143. Kwan, M.-K. (1962). “Graphic programming using odd or even points”, ChineseMathematics 1: 273-276.
144. Lang, L. (1999). Transportation GIS. Redlands, CA, Environmental SystemsResearch.
145. Lapalme, G., Rosseau, J.-M., et al. (1992). “Georoute : A Geographic InformationSystem for Transportation Applications”, Communications of the ACM 35(1): 81-88.
Spatial Decision Support Systems for Large Arc Routing Problems
217
146. Laporte, G., Asef-Vaziri, A. and Sriskandarajah, C. (1996). “Some applications ofthe generalized travelling salesman problem”, Journal of the OperationalResearch Society 47(12): 1461-7.
147. Laporte, G. and Osman, I. (1995). “Routing Problems: A bibliography”, Annals ofOperations Research 61: 227-262.
148. Lee, Z., Gosain, Sanjay and Im, I. (1999). “Topics of interest in IS: evolution ofthemes and differences between research and practice”, Information &Management 36(5): 233-246.
149. Lenstra, J. K. and Rinnooy Kan, A. H. G. (1976). “On general routing problems”,Networks 6(3): 273-280.
150. Lenstra, J. K. and Rinnooy Kan, A. H. G. (1981). “Complexity of vehicle routingand scheduling problems”, Networks 11(2): 221-227.
151. Letchford, A. and Eglese, R. (1998). “The rural postman problem with deadlineclasses”, European Journal of Operational Research 105(3): 390-400.
152. Letchford, A. N. (1999). “The general routing polyhedron: A unifyingframework”, European Journal of Operational Research 112(1): 122-133.
153. Levy, L. and Bodin, L. (1989). “The arc oriented location routing problem”,INFOR 27(1): 74-94.
154. Li, L. Y. O. (1992). Vehicle routeing for winter-gritting, PhD Thesis, Dept. ofManagement Science, Lancaster University.
155. Li, L. Y. O. and Eglese, R. W. (1992). A lower bound for the time constrainedrouting problem.
156. Li, L. Y. O. and Eglese, R. W. (1996). “An interactive algorithm for vehiclerouteing for winter-gritting”, Journal of the Operational Research Society 47(2):217-228.
157. List, G., Mirchandani, P., Turnquist, M. and Zografos, K. (1991). “Modeling andAnalysis for Hazardous Materials Transportation: Risk Analysis,Routing/Scheduling and Facility Location”, Transportation Science 25(2): 100-114.
158. Little, J. D. C. (1971). “Models and managers: the concept of a decision calculus”,Management Science 16(8): 466-485.
159. Maguire, D. J. (1991). An Overview and definition of GIS. GeographicalInformation Systems, Volume 1 : Principles. D. J. Maguire, M. F. Goodchild andD. W. Rhind, Longman. 1: 9-20.
160. Malandraki, C. and Daskin, M. S. (1993). “The maximum benefit ChinesePostman problem”, European Journal of Operational Research 65(2): 218-234.
161. Mallach, E. G. (1994). Understanding Decision Support Systems and ExpertSystems, Irwin.
162. Mapinfo MapInfo Corp (http://www.mapinfo.com), Troy, NY, USA.
163. Marakas, G. M. (1998). Decision Support Systems in the 21st Century, PrenticeHall.
164. Martell, D. L., Gunn, E. A. and Weintraub, A. (1998). “Forest managementchallenges for operational researchers”, European Journal of OperationalResearch 104(1): 1-17.
Spatial Decision Support Systems for Large Arc Routing Problems
218
165. McBride, R. (1982). “Controlling left and U-turns in the routing of refusecollection vehicles”, Computers and Operations Research 9(2): 145-152.
166. McCosh, A. M. and Scott Morton, M. S. (1978). Management decision supportsystems. London, Macmillan.
167. Mejia-Navarro, M. and Garcia, L. A. (1995). Integrated Planning DecisionSupport System (IPDSS). R. Report, Integrated Decision Support Group,Colorado State University.
168. Mennecke, B. E. (1997). “Understanding the Role of Geographic InformationTechnologies in Business: Applications and Research Directions”, Journal ofGeographic Information and Decision Analysis 1(1): 44-68.
169. Mennecke, B. E., Crossland, M. D. and Killingsworth, B. L. (2000). “Is a mapmore than a picture? The role of SDSS technology, subject characteristics, andproblem complexity on map reading and problem solving”, MIS Quarterly 24(4):601-4, 625-9.
170. Mentzas, G. (1994). “A functional taxonomy of computer based informationsystems”, International Journal of Information Management 14(6): 397-410.
171. Minieka, E. (1979). “The Chinese postman problem for mixed networks”,Management Science 25(7): 643-648.
172. Mole, R. H. (1979). “A survey of local delivery vehicle routing methodology”,Journal of the Operational Research Society 30(3): 245-252.
173. Montwill, P. and Naughton, M. (1994). The Capacitated Chinese PostmanProblem, MMangtSc Thesis, Dept. of MIS, University College Dublin.
174. Mtenzi, F. J. (2000). The sparse travelling salesman problem, PhD Thesis, Dept.of Management Information Systems, University College Dublin.
175. Muller, J.-C. (1993). “Latest developments in GIS/LIS”, International Journal ofGeographical Information Systems 7(4): 293-303.
176. Murphy, L. (1996) Competing in Space: The Strategic Roles of GeographicInformation Systems. The conference of the Association of Information SystemsProceedings, Las Vegas
177. Navtech (1999). Navigation Technologies Corp, Rosemont, Illinois, USA.
178. Naylor, T. (1982). “Decision support systems or whatever happened to MIS?”,Interfaces 12(4): 92-97.
179. Newcomb, M. and Medan, J. (1993) Data Requirements for Route Guidance.Annual Conference of the Urban and Regional Information Systems AssociationProceedings: 200-211.
180. Önal, H., Jaramillo, B. M. and Mazzocco, M. A. (1996). “Two formulations of theVehicle Routing Problem: An empirical application and computationalexperience”, Logistics and Transportation Review 32(2): 177-190.
181. Orloff, C. S. (1974). “A fundamental problem in vehicle routing”, Networks 4: 35-64.
182. Partyka, J. G. and Hall, R. W. (2000). “On the Road to Service”, OR/MS Today27(4).
183. Patel, M. and Horowitz, A. (1994). “Optimal routing of hazardous materialsconsidering risk of spill”, Transportation Research: Part A 28A(2): 119-132.
Spatial Decision Support Systems for Large Arc Routing Problems
219
184. Pearn, W. L. (1988). “New Lower Bounds for the Capacitated Arc RoutingProblem”, Networks 18(3): 181-191.
185. Pearn, W. L. (1989). “Approximate solutions for the capacitated arc routingproblem”, Computers and Operations Research 16(6): 589-600.
186. Pearn, W. L. (1991). “Augment-Insert Algorithms for the Capacitated RoutingProblem”, Computers and Operations Research 18(2): 189-198.
187. Pearn, W. L., Assad, A. and Golden, B. L. (1987). “Transforming Arc Routing intoNode Routing Problems”, Computers and Operations Research 14(4): 285-288.
188. Pearn, W. L. and Chou, J. B. (1999). “Improved solutions for the ChinesePostman problem on mixed networks”, Computers and Operations Research26(8): 819-827.
189. Pearn, W. L. and Li, M. L. (1994). “Algorithms for the Windy Postman Problem”,Computers and Operations Research 21(6): 641-651.
190. Pearn, W. L. and Wu, T. C. (1995). “Algorithms for the rural postman problem”,Computers and Operations Research 22(8): 819-828.
191. Portier, M. A., Berthet, P. and Moreno, J. (1994) Optimized Network Modellingfor Route Planning. European Conference on Geographical Information SystemsProceedings: 1807-1816.
192. Potvin, J., Lapalme, G. and Rousseau, J. (1989). “ALTO:A computer system forthe design of vehicle routing algorithms”, Computers and Operations Research16(3): 451-470.
193. Potvin, J.-Y., Lapalme, G. and Rousseau, J.-M. (1994). “A microcomputerassistant for the development of vehicle routing and scheduling heuristics”,Decision Support Systems 12(1): 41-56.
194. Power, D. J. and Kaparthi, S. (1998). The changing technological context ofDecision Support Systems. Context Sensitive Decision Support Systems. D.Berkeley, G. Widmeyer, P. Brézillion and V. Rajkovic, Chapman Hall: 41-54.
195. Psaraftis, H. N. (1995). “Dynamic Vehicle Routing : Status and Prospects”,Annals of Operations Research 61: 143-164.
196. Ralston, B. and Zhu, X. (1991) Interfacing Stand Alone Transport AnalysisSoftware with GIS. GIS/LIS '91 Proceedings: 209-218.
197. Rasmussen, L. H. (1997). “GIS in the military,yesterday, today and tomorrow”,GIS Europe 6(1): 26-28.
198. Rolland, E. and Gupta, R. (1996) New linkages between GIS and combinatorialoptimization. AIS Americas Conference on Information Systems Proceedings,Phoenix, Arizona, USA
199. Roy, S. and Rousseau, J.-M. (1989). “The Capacitated Canadian PostmanProblem”, INFOR 27(1): 58-73.
200. Sahay, S. and Walsham, G. (1996). “Implementation of GIS inIndia:organizational issues and implications”, International Journal ofGeographical Information Systems 10(4): 385-404.
201. Santana, M. (1995) Managerial learning: A neglected dimension in decisionsupport systems. 28th Annual Hawaii International Conference on SystemSciences Proceedings
Spatial Decision Support Systems for Large Arc Routing Problems
220
202. Saruwatari, Y., Hirabayashi, R. and Nishida, N. (1992). “Node Duplication LowerBounds for the Capacitated Arc Routing Problem”, Journal of the OperationsResearch Society of Japan 35(2): 119-133.
203. Sauder, R. L. and Westerman, W. M. (1993). Computer aided train dispatching:decision support through optimization. Decision Support Systems: Putting theoryinto practice. R. H. Sprague and H. J. Watson, Prentice Hall: 85-95.
204. Sauter, V. (1997). Decision Support Systems. New York, John Wiley & Sons.
205. Silver, M. S. (1991). “Decisional guidance for computer-based decision support”,MIS Quarterly 5(1): 105-22.
206. Simon, H. A. (1977). The new science of management decision, Prentice-Hall.
207. Smelcer, J. B. and Carmel, E. (1997). “The effectiveness of differentrepresentations for managerial problem solving: Comparing maps and tables”,Decison Sciences 28(2): 391-420.
208. Sprague, R. (1980). “A Framework for the development of Decision SupportSystems”, MIS Quarterly 4(1).
209. Sprague, R. H. and Carlson, E. D. (1982). Building Effective Decision SupportSystems, Prentice Hall International.
210. Stabell, C. B. (1986). Decision Support Systems: Alternative Perspectives andSchools. Decision Support Systems: a decade in perspective. E. McLean and H. G.Sol, North-Holland: 173-182.
211. Stern, H. I. and Dror, M. (1979). “Routing electric meter readers”, Computersand Operations Research 6(4): 209-223.
212. Stocx, C. F. M. and Tilanus, C. B. (1991). “Deriving route lengths from radialdistances: Emphirical evidence”, European Journal of Operations Research 50(1):22-26.
213. Sussams, J. (1984). “The future for computerised vehicle-load planning systems”,Journal of the Operational Research Society 35(11): 963-966.
214. Swink, M. and Speier, C. (1999). “Presenting geographic information: Effects ofdata aggregation, dispersion, and users' spatial orientation”, Decision Sciences30(1).
215. Sylvan Sylvan Ascent, Inc., Santa Fe, NM, USA.
216. Tactician “Tactician Corporation (http://www.tactician.com)”, .
217. Teng, J. T. C. and Galletta, D. F. (1990). “MIS research directions: A survey ofResearchers' Views”, Data Base 21(3-4): 1-10.
218. Thill, J.-C. (2000). “Geographic information systems for transportation inpersepective”, Transport Research Part C 8C(1-6): 3-12.
219. Tracey, M. and Dror, M. (1997). “Interactive graphical computer application forlarge-scale cattle feed distribution management”, Decision Support Systems19(1): 61-72.
220. Transcad (1996). Transcad User's Guide Version 3.0 for Windows, Newton, Mass.USA, Caliper Corporation.
221. Turban, E. (1995). Decision Support and Expert Systems, Prentice-HallInternational.
Spatial Decision Support Systems for Large Arc Routing Problems
221
222. Turban, E. and Aronson, J. E. (1998). Decision Support and Intelligent Systems,Prentice-Hall International.
223. van der Knapp, W. (1993) GIS and Planning of Route Selection made by TouringCyclists. European Conference on Geographical Information Systems '93Proceedings: 1580-1581.
224. Walsham, G. and Sahay, S. (1999). “GIS for district-level administration inIndia: Problems and opportunities”, MIS Quarterly 23(1): 39-65.
225. Watson, H. J. and Hill, M. M. (1983). “Decision Support Systems or what didn'thappen with MIS”, Interfaces 13(3): 81-88.
226. Weigel, D. and Cao, B. (1999). “Applying GIS and OR techniques to solve Searstechnician dispatching and home delivery problems”, Interfaces 29(1): 112-130.
227. Wilson, R. D. (1994). “GIS & decision support systems”, Journal of SystemsManagement 45(11): 36-40.
228. Win, Z. (1988). Contributions to Routing Problems, PhD Thesis, UniversitätAugsburg.
229. Wunderlich, J., Collette, M., Levy, L. and Bodin, L. (1992). “Scheduling meterreaders for Southern California Gas Company”, Interfaces 22.(3): 22-30.
230. Zhan, F. B. and Noon, C. E. (1998). “Shortest Path Algorithms: An evaluationusing real road networks”, Transportation Science 32(1): 65-73.