vms evaluation

8/13/2019 vms evaluation

1/17

Nonlinear Dyn (2007) 49:493509

DOI 10.1007/s11071-006-9111-3

O R I G I N A L A R T I C L E

Modelling drivers compliance and route choice behaviour

in response to travel informationHussein Dia Sakda Panwai

Received: 3 November 2005 / Accepted: 12 March 2006 / Published online: 12 January 2007C Springer Science+Business Media B.V. 2007

Abstract This paper addresses commuters route

choice behaviour in response to traveller information

systems. The data used in this study was obtained

from a field behavioural survey of drivers that was

conducted on a congested commuting corridor in

Brisbane, Australia. Agent-based neural network

models (Neugents) were used to analyse the impacts

of socio-economic, context and information variables

on individual behaviour and propensity to change

route and adjust travel patterns. The results from these

models clearly indicate that prescriptive, predictiveand quantitative real-time delay information provided

for both the usual and best alternate routes are most

effective in influencing commuters to change their

routes. The Neugent behavioural models describing

drivers dynamic route choice decision making

were also implemented within a microscopic traffic

simulation tool to evaluate the corridor-wide impacts

of providing drivers with real-time traffic informa-

tion. The simulation results support the notions that

commuters decisions to divert to alternate routes are

influenced by their socio-economic characteristics;the degree of familiarity with network conditions and

the expectation of an improvement in travel time that

exceeds a certain delay threshold associated with each

H. Dia S. PanwaiSchool of Engineering, Hawken Engineering Building,Staff House Road, The University of Queensland,Brisbane, Queensland 4072, Australiae-mails: [email protected]; [email protected]

commuter. An evaluation of the benefits of the Neugent

model over static route choice algorithms which do not

consider dynamic driver behaviour and compliance

with travel advice showed improvements of 47% in

network speeds; 58% in network delays; 711% in

stop time per vehicle and 13% in network travel times.

Keywords Behavioural route choice models.

Artificial neural networks.Microscopic traffic

simulation.Advanced traveller information systems

Introduction

Advanced traveller information systems (ATIS) are in-

creasingly being recognised as a potential strategy for

influencing driver behaviour on route choice, trip mak-

ing, time of departure and mode choices. The provi-

sion of real-time travel information allows travellers to

make informed travel decisions and has the potential

to improve network efficiency, reduce congestion and

enhance environmental quality. The successful imple-mentation of these systems, however, will depend to

a large extent on understanding how motorists adjust

their travel behaviour in response to the information

received.

This paper presents findings based on a travel be-

haviour survey of commuters on a congested commut-

ing corridor in Brisbane. The survey results provided

a useful insight into the factors influencing driver be-

haviour and route choice decisions under the presence

Springer


2/17

494 Nonlinear Dyn (2007) 49:493509

of real-time information. A number of agent-based

neural network dynamic driver behaviour models that

were developed to predict drivers compliance and

route choice decisions under the influence of variable

message sign (VMS) information are also presented.

The applicability and feasibility of the approach was

demonstrated by interfacing the neural network driverbehavioural models to a microscopic traffic simula-

tor to evaluate the impacts of providing travel in-

formation on driver compliance rates and network

performance.

Information requirements

The quality of the information provided by ATIS is

critical to the credibility of these systems and needs

to be perceived as accurate and reliable. The effect

of information on travellers decisions depends on its

content, format or presentation style, and nature (i.e.

static, dynamic or predictive). This information can be

either qualitative, quantitative or predictive delay infor-

mation. Prescriptive information advises travellers to

make a specific choice (i.e. take an alternative route).

Descriptive information can be either qualitative or

quantitative and provides information on travel times,

incident locations and expected delays. A study con-

ducted by Polydoropoulou et al. [15] found that trav-

ellers propensity to take alternative routes increased

with prescriptive information and the most significant

increase occurred when quantitative, real-time or pre-

dictive information on both the usual route and the

alternative routes were provided. This is consistent

with the study by Khattak et al. [7] which concluded

that with accurate quantitative delay information, com-

muters may overcome their behavioural inertia when

faced with unexpected delays.

Travel behaviour framework

For the purpose of this study, it is assumed that a

travel behaviour framework similar to that proposed

by Khattak et al. [7] applies. This framework takes into

account that individuals attempt to find the best alter-

native route using the limited information available to

them. Various aspects of travel information influence

travellers decisions. The processing of information de-

pends on its content or meaning, format or presentation

style, its nature and type [18].

An example of a dynamic driver behaviour mod-

elling framework is shown in Fig. 1 [1]. Drivers set out

with goals to travel between an origin and destination

within a given period of time while incurring the low-

est possible cost. They acquire information about the

performance of the road system through direct obser-

vation or by having access to electronic information

systems. Drivers process and interpret the information

in light of their current knowledge and in accordance

with their ability to combine and process a variety of

information concerning road conditions. The interpre-

tation translates into perceptions of travel times and

Goals

Decision Rules Information Acquisition

Processing Capacity

Computational Ability

acquire info

including memory

Combining info

Actual Decisions

Reviewing

using info / prediction

Review

decision rules

Review

decisions

Choose travelpattern

Fig. 1 Driver decision process [1]

Springer


3/17

Nonlinear Dyn (2007) 49:493509 495

delays. Perceptions, restrictions and individual charac-

teristics also form preferences for certain alternatives

(modes, routes and departure times). These preferences

will also depend on the previously acquired knowl-

edge, stored in the memory, and on certain thresholds

whereby motorists only switchfrom their current path if

the improvement in travel time exceeds some thresholdlevel. These preferences result in observable choices

that have outcomes (e.g. arrival time at work). If the

outcome is satisfactory, then the same choice is likely

to be repeated on the following trip forming a commute

pattern [1]. The repetition of a choice in the commute

context also depends on the future or anticipated out-

comes. These outcomes also provide feedback to the

memory in the form of knowledge updates. In unex-

pected delay situations, the anticipated outcomes are

often unsatisfactory triggering review of preferences

and changes in normal travel patterns on a real-timeand day-to-day basis [1].

As was mentioned earlier, various aspects of travel

information influence drivers decisions. The process-

ing of information depends on its content or meaning,

nature, type (whether it is quantitative or qualitative)

and presentation style. In addition, drivers are more

likely to rely on relevant and accurate information. For

example, under incident congestion, the perception of

delay and the quality of real-time information are crit-

ically important in influencing travel behaviour. Ob-

viously, the success of driver behavioural models willdepend to a large extent on capturing the different pa-

rameters mentioned above. Most previous research on

driver behaviour modelling has focused on modelling

travel response decisions based on data from travel sur-

veys or travel simulators [8]. These behavioural mod-

els mostly used drivers socio-economic characteris-

tics and attributes of usual travel patterns as explana-

tory variables. The work reported here extends previous

research efforts and will be based on real data compris-

ing the complete set of drivers choices as determined

from a driver behavioural survey.

Behavioural survey of drivers on a congested

commuting corridor

Behavioural surveys of drivers during congested traf-

fic conditions are best suited to developing behavioural

models under the influence of travel information. Prop-

erly designed surveys that capture the interactions in the

travel behaviour model would allow for the investiga-

tion of the influence of (a) unexpected and expected

congestion, (b) the various types and quality of in-

formation received about congestion and (c) drivers

experiences with congestion and related information

on the whole spectrum of pre-trip and en route deci-

sions. In particular, these behavioural surveys allowfor the relationship between a drivers response to in-

formation to be modelled in combination with actual

behaviour.

Data collection

User response to travel information is typically mod-

elled using data collected from a behavioural survey of

congestion [7]. For this study in Australia, mail-back

questionnaires were distributed to peak-periodautomo-

bile commuters travelling along a traffic commuting

corridor in Brisbane. The survey comprised questions

in the following five categories:

(1) Personal information. A drivers age, occupation and

gender may influence certain travel behaviours.

(2) Normal travel patterns. These include day-to-day be-

haviour such as work schedule, route choice and re-

sponse to recurring congestion.

(3) Pre-trip response to unexpected congestion informa-

tion.The knowledge that road conditions are abnor-

mal may influence drivers decisions regarding de-parture time and route choice before setting out on

their journeys.

(4) En route response to unexpected congestion informa-

tion. When drivers learn of abnormal road conditions

while driving, they may change certain travel deci-

sions.

(5) Willingness to change driving patterns. Given some

incentive (e.g. reduction in travel time), drivers may

be willing to leave early or take an alternative

route.

Two questionnaire forms were distributed: the first

form included questions from Categories 1, 2, 3 and 5

(for pre-trip information) whereas the second form in-

cludedall questionsfromCategories 1,2, 4 and 5 (for en

route information). Users responses to ATIS are based

on (a) revealed preference (RP) data in which the actual

behavioural response to expected/unexpected delay is

reported and (b) stated preference (SP) data in which a

drivers behaviour under a hypothetical ATIS scenario

Springer


4/17

496 Nonlinear Dyn (2007) 49:493509

Table 1 Vehicle use patterns

Frequency (%)

Vehicle use Never Occasionally 1 day 2 days 3 days 4 days 5 + days

Alone by car 6.4 6.4 2.9 2.3 4.7 5.8 71.3

In a car-pool 80.7 7.0 1.8 1.2 0.0 2.3 7.0

Using park n ride facilities 91.2 4.1 0.6 1.2 1.2 0.0 1.8Using public transport 80.1 13.5 2.3 1.2 1.8 0.0 1.2

Other 94.7 2.3 0.0 0.6 1.2 0.0 1.2

Table 2 Stated commuting patterns

Response

Commuting pattern Strongly disagree Disagree Agree Strongly agree Cant say

Willing to discover new routes to get somewhere 8.8 27.1 41.8 20.6 1.8

Willing to take unfamiliar routes to avoid traffic delays 5.3 18.2 44.7 29.4 2.4

is reported. Each questionnaire comprised 40 questions

and took about 20 min to complete.

A total of 490 questionnaires were distributed to

drivers on a traffic commuting corridor in Brisbane

during the morning and afternoon peak periods.

South-west Brisbane was selected as the study corridor

because it met a number of criteria including the

availability of alternate routes to the CBD area, the

availability of public transport and real-time traffic

information and variable message signs providing

information about traffic delays to the CBD. The mail-

back questionnaire had a response rate of 34% (167

questionnaires) comprising a total of 82 pre-tip and 85

en route questionnaires. This response rate compares

favourably with the results obtained from overseas

research [7] in which substantial financial incentives

were offered to respondents. The format and presenta-

tion of the questionnaire is believed to be a key factor

in achieving this response rate, considering it was

anticipated that it would take 20 min to complete eachquestionnaire.

Overview of selected survey results

The main objective of the behavioural survey was to

determine the factors that influence route change; the

frequency of route change and traffic information pref-

erences by respondents. Some selected results relevant

to the dynamics of commuter route choice behaviour

are summarised below. A more detailed discussion of

survey results can be found in Dia et al. [2].

Vehicle use and commuting patterns

Table 1 summarises the respondents vehicle use pat-

terns. These results clearly indicate a high level of car

dependency amongst the respondents travelling on thechosen corridor with 71% choosing to drive to work in

single occupancy vehicles five or more times a week.

About 80% of respondents never used car-pools;

91% never used park n ride facilities; 80% never used

public transport and 95% never used other alternative

modes of transport. These results suggest a tendency

towards habit formation or mode choice commitment

amongst the respondents. Further exploration of the re-

spondents commuting patterns (Table 2) revealed that

62%of respondents were willing to discover new routes

and 74% were willing to take unfamiliar routes to avoidtraffic delays. These results are significant as they in-

dicate the potential of traveller information systems to

influence drivers route choice decisions.

Socio-economic attributes of respondents

Socio-economic and travel attributes of respondents

have an important effect on travel behaviour. As part

of this survey, a rich source of individual data has been

Springer


5/17

Nonlinear Dyn (2007) 49:493509 497

Table 3 Characteristics of travel behaviour surveys for Brisbane, San Francisco and Chicago

Description Brisbane, Australia San Francisco, USA Chicago, USA

Sample size 171 3238 700

Gender (%)

Male 56.1 64.9 54.3

Female 43.9 35.1 45.7

Age (%)

Less than 19 0.6 0.2 0

2029 23.5 11.4 23.6

3039 22.4 31.1 33

4049 28.8 33.7 29.7

5064 22.4 21.4 12.5

Above 65 2.4 2.2 1.2

Education (%)

High school or less 15.9 4.3 4.6

Vocational/technical school 20.0 1.3 19.4

Undergraduate degree 33.5 61.3 36.1

Post-graduate degree 30.6 33.1 39.9

Annual personal income (%)Under $20,000 4.9 3.1 4.3

$20,00040,000 25.8 18.3 33.1

$40,00060,000 27.0 23 26.3

$60,00080,000 15.3 14.2 14

$80,000100,000 11.0 14.2 6.6

Above $100,000 16.0 27.2 15.7

Travel time (min)

On usual route 31 40.6 43

On best alternative route 33 44.8 53.4

Average duration of residence in the area (years) 8.3 6.7 5.3

collected which will be important in modelling the fac-tors that affect trip change behaviour,willingness to pay

for ATIS services and compliance with travel informa-

tion and route directives. Table 3 presents a summary

of selected socio-economic attributes of respondents.

The average respondent is middle-aged and has resided

in the area for 8.3 years. The respondents are divided

fairly equally between males and females with 56%

of respondents being males. The average annual in-

come is $62 000 and 64% of respondents have either

undergraduate or postgraduate qualifications. The pri-

mary occupations of this sample are professional, cler-ical/service, executive and managerial/administration.

As suggested by these results, upper-income groups

and well-educated individuals were over-represented

in this survey when compared to census demographic

profiles of the Brisbane population. However, this

problem has also been experienced in other similar

studies conducted in the United States as presented

by Polydoropoulou and Khattak [16] and shown in

Table 3.

Respondents preferences for trafficinformation sources

Travellers on this corridor received information from

a variety of sources. Table 4 lists the current sources

of traffic information as indicated by the respondents.

Most respondents indicated that their primary source

of traffic information was radio traffic reports (74%)

and their own observation (64%). About 23% indicated

that they relied on variable message signs (VMS) as a

source of traffic information on their usual route.

These results clearly indicate that the implemen-tation of strategically located and credible VMS has

the potential to influence drivers route choice deci-

sions. Other traffic information sources such as the In-

ternet and in-vehicle navigation systems were rarely

used by respondents. This may be attributed to the lim-

ited market penetration rates of in-vehicle navigation

systems and the lack of Internet-based traffic informa-

tion systems in Brisbane at the time the survey was

conducted.

Springer


6/17

498 Nonlinear Dyn (2007) 49:493509

Table 4 Current traffic information sources

Information source Frequencya (%)

Radio traffic reports 74

Own observation 64

Electronic message signs 23

Conversations with other people 10

Mobile phone 4

Printed matter 2

Television 2

Home/office telephone 1

In-vehicle navigation system 1

Internet 1

Other 0

aMultiple responses allowed

Table 5 Respondents frequency of route change in theprevious month

No. of times respondentschanged route in

previous month No. of respondents Percentage

None 37 25.5

1 14 9.7

2 25 17.2

3 19 13.1

4 17 11.7

5 17 11.7

6+ 16 11.0

Frequency of route change and causesof unexpected congestion

The frequency of route change was examined by asking

the respondents to provide information on the number

of times they changed routes in the past month, in re-

sponse to the presence of congestion on their usual

routes. The results are displayed in Table 5.

On average respondents took an alternative route

3.6 times in a month (about 20 working days). How-

ever, it is believed that the proportion of drivers com-

plying with travel information can be increased by de-signing effective and credible traveller information sys-

tems. This is discussed in more detail in the next sec-

tions dealing with the content of the messages and the

type of information provided.The causes of unexpected

congestion experienced by respondents is presented in

Table 6.

Construction and roadworks were reported as the

major causes of unexpected congestion on both the pre-

trip and en route questionnaires. Accidents were also

Table 6 Causes of unexpected congestion

Cause of congestion Pre-trip (%) En-route (%)

Construction/roadworks 58.6 42.2

Accident 44.8 40.6

Disabled vehicle 6.9 10.9

Dont know 6.9 21.9

Bad weather 3.4 7.8

Other 0.0 0.0

a significant source of unexpected congestion. During

the time of the distribution of this survey, major main-

tenance and construction works were being conducted

on Coronation Drive and adjoining networks possibly

accounting for the severity of unexpected congestion

caused by roadworks. This result confirms the need

for advanced freeway and arterial incident detection

systems capable of detecting incidents in the shortestpossible time and making this information available to

the travelling public through advanced traveller infor-

mation systems.

En route responses to hypothetical ATIS messages

The effect of providing alternative forms of en route

information was evaluated. Respondents were asked

to state whether they would change travel decisions if

they were alerted of delays. Each of the five different

information types being tested were then presented to

respondents.

Qualitative information

In this situation the device only offered a simple mes-

sage: Unexpected Congestion on your usual route

(Fig. 2), where your usual route is the road that re-

spondents indicated they normally used for their travel.

While this is simple information, it represents the type

of information that is commonly available to travellersvia radio or VMS.

Fig. 2 Qualitative information message

Springer


7/17

Nonlinear Dyn (2007) 49:493509 499

Fig. 3 (a) Quantitative information messageusual route. (b)Quantitative informationmessageusual and best alternateroutes

Fig. 4 Predictive information message

Quantitative information

For this scenario, respondents were provided with the

same message as before, but the device also displayed

the expected delays on the usual route and the traveltime on the best alternate route (as specified by respon-

dents), as shown in Fig. 3.

Predictive information

For this scenario, respondents were asked how they

would modify their travel choices if thedevice provided

them with the delays at the present time, and accurately

predicted the expected delays in 15 and 30 min into the

future, as shown in Fig. 4.

Prescriptive information

This scenario explored the response to recommenda-

tions offered by the ATIS which suggested taking the

route which the respondents indicated was their best

alternative route to their destination (Fig. 5).

Respondents were asked to indicate their prefer-

ences when presented with hypothetical ATIS informa-

tion by choosing from a set of finite responses which

Fig. 5 Prescriptive information message

included definitely take my usual route; probably

take an alternative route; definitely take best alterna-

tive route; probably take best alternative route and

cant say. A summary of respondents choices is pre-

sented in Table 7. These results provide one of the most

significant findings from the ATIS experiment. They

clearly indicate that prescriptive, predictive and quan-

titative real-time delay information provided for both

the usual and best alternate routes are most effective ininfluencing respondents to change their routes. There-

fore, detailed route choice decision models were devel-

oped and investigated for each of these ATIS scenarios

as discussed next.

Model structure and development tools

Driver behaviour models that can be used to dynami-

cally estimate or predict the degree of drivers compli-

ance with traffic information can be thought of as a clas-sification problem. The inputs to the model comprise

drivers individual socio-economic characteristics and

other variables that may influence their route choice

behaviour; and the output a binary integer (1,0) repre-

senting whether drivers comply with travel advice or

not, respectively. Two approaches are available for de-

veloping such models: discrete choice models and arti-

ficial neural network techniques. Detailed information

about the theoretical foundations of these techniques

is beyond the scope of this paper. However, a compre-

hensive treatment of discrete choice models and theirapplications can be found in [4] and a detailed review

of artificial neural networks and their applications in

the transportation field can be found in [3]. Hawas [5]

provides a state-of-the-art review of route choice mod-

els where he discusses the strengths and weaknesses

of the two approaches and presents recent advances in

model formulations aimed at addressing their limita-

tions. The review clearly points to the limitation of the

discrete choice approach in modelling the vagueness

Springer


8/17

500 Nonlinear Dyn (2007) 49:493509

Table 7 En-route stated preferences for unexpected congestion

Qualitative Quantitative Quantitative RT Predictive Prescriptive

delay real-time delay delay on best real-time delay best alternate

Attributes information information alternate route information route

Definitely take my usual route 12.0 10.3 8.0 9.3 6.3

Probably take my usual route 28.9 29.5 12.0 16.0 13.9

Definitely take my best alternate route 32.5 29.5 41.3 41.3 53.2

Probably take my best alternate route 25.3 24.4 33.3 29.3 22.8

Cant say 2.3 7.7 5.3 5.3 3.8

(or fuzziness) of driver behaviour. The author cites a

number of studies where fuzzy logic was effectively

used to capture the variability of drivers appraisal of

the different route attributes as well as the variabil-

ity in their perceptions to the various attribute levels.

In particular, the use of neuro-fuzzy systems, where

fuzzy logic is coupled with the neural networks train-ing capability to replicate complex system behaviour,

was highlighted as a technique with potential promise.

The neural network training basically uses examples

of real-life behaviour to calibrate the fuzzy sets param-

eters and knowledge base with the aim of replicating

drivers route choice behaviour. Evidence of the poten-

tial for using neural network models for route choice

can be found in a number of studies which conducted

comparative evaluations of the discrete choice and neu-

ral network approaches [6, 12]. These studies showed

that neural networks can provide comparable if notbetter approximations to route choice problems. For

this study, neural network techniques were selected as

the modelling paradigm to predict drivers compliance

with travel information.

Neural network route choice modelling framework

The development of neural network models involves

a number of steps which include data collection, se-

lection of input variables, assignment of desired out-

put states, creation of training and validation data sets,selection of neural network architecture (e.g. number

of nodes in the hidden layer, number of hidden lay-

ers, objective function, learning algorithms and node

transfer functions, etc.), training strategy and selection

of key performance indicators to be used for evaluat-

ing the performance of the model. The neural network

route choice models presented in this paper will only

address drivers en route responses to traffic delay in-

formation. The treatment of drivers compliance with

Table 8 Data sets for model development

Data set ATIS message types Sample size

1 Qualitative delay information 80

2 Quantitative real-time delay

information 74

3 Quantitative real-time delay on best

alternative route 714 Predictive delay information 70

5 Prescriptive best alternative route 71

pre-trip travel information is the subject of a separate

study and will not be reported here.

Data for model development

The five data sets that were used for model development

are shown in Table 8 below. Each data set comprisesrespondents replies to each of the ATIS message types

described before (Table 7 and Figs. 25).

Selection and coding of input and output variables

As was mentioned before, drivers route choice deci-

sions are related to their socio economic characteris-

tics and travel habits. The inputs selected for this study

comprised income, education level, age, gender, years

at residence (a surrogate for familiarity with road net-

work conditions) and work schedule flexibility. A num-ber of studies [5] also found that these inputs have di-

rect impact on drivers responses to travel advice. Trav-

ellers responses to each of these inputs were coded ac-

cording to the categories shown in Table 9. Work sched-

ule flexibility, age and awareness or familiarity with

road network conditions are considered factors that im-

pact drivers aggressiveness and hence may influence

their decisions on route choices when provided with

real-time traffic information [9]. Drivers willingness

Springer


9/17

Nonlinear Dyn (2007) 49:493509 501

Table 9 Variables used for model development

Input data sets

Input variable Category description Categorised value Normalised value

Work schedule flexibility Flexible 1 0.00

Fixed 2 0.50

Variable 3 1.00Age Under 18 years 1 0.00

Between 18 and 29 years 2 0.20




Over 65 years 6 1.00

Gender Male 1 0.00

Female 2 1.00

Income level Under 20 thousand $AUD 1 0.00

Between AUD $20,000 and AUD $40,000 2 0.20

Between AUD $40,000 and AUD $60,000 3 0.40

Between AUD $60,000 and AUD $80,000 4 0.60Between AUD $80,000 and AUD $100,000 5 0.80

More than AUD $100,000 6 1.00

Education level High school or less 1 0.00

Vocational or technical school 2 0.33

Undergraduate degree 3 0.66

Post graduate degree 4 1.00

Years at residence Under 5 years 1 0.00




More than 20 years 5 1.00

to pay for travel information using ATIS technologies

and services is clearly a function of their income and

has also been found to be related to age and gender [17].

Having determined the relevant inputs, the neural

network architecture can be formulated as shown in

Fig. 6 where the outputof the model represents whether

a driver complies or not with the travel information.

Drivers compliance

Drivers compliance was captured in the behaviouralfield survey through the respondents answers to the

different ATIS scenarios. For example, drivers are said

to comply with travel information if their response to

the specific ATIS scenario was definitely take my best

alternate route. Similarly, they are said to ignore the

travel advice if their responses were definitely take

my usual route. For the purpose of model develop-

ment, these two response or model output categories

were coded as 1.0 and 0.0, respectively. Respondents

also had two other options to choose from if they were

undecided on which route to choose. If the respon-

dent selected either probably take my usual route or

probably take my best alternate route, then these two

response categories will need to be coded as values be-

tween (0, 0.5) and (0.5, 1.0), respectively. A number

of experiments were conducted where different values,

e.g. (0.33, 0.67), (0.3, 0.7) and (0.4, 0.6) were coded for

these categories but no differences were found in terms

of the neural network model performance. The four-

point semantic scale of drivers route choice selectedin this study is shown in Table 10.

Selection of neural network architecture

Four neural network architectures are typically used for

classification problems [10]: Fuzzy ARTMAP, Back

Propagation (Logicon Projection Network), Reinforce-

ment Network and Radial Basis Function Network

Springer


10/17

502 Nonlinear Dyn (2007) 49:493509

Fig. 6 Basic structure of neural network agent-based (Neugent) route choice model

(RBFN). A brief description of each of these archi-

tectures is provided below.

(1) Fuzzy ARTMAPis a general-purpose classification

network. It has two fuzzy ART networks (ARTa

and ARTb) which are connected by a subsystem

referred to as a match tracking system which in-

creases the vigilance of the Fuzzy ART network

until a match is made or a new category is formed.

(2) Back-Propagation is a general-purpose network

paradigm. Back-prop calculates an error between

desired and actual outputs and propagates the error

information back to each node in the network. The

back-propagated error drives the learning at each

node.

(3) Radial Basis Function Networks (RBFN) are

general-purpose networks which can be used for

a variety of problems including classification, sys-

tem modelling and prediction. In general, an RBFN

Table 10 Coding of route choice preferences

Strength of Choice transformation

preference scale

Definitely take my usual route 0.00

Probably take my usual route 0.33

Probably take my best alternate route 0.67

Definitely take my best alternate route 1.00

is any network which makes use of radially sym-

metric and radially bounded transfer functions in

its hidden (pattern) layer.

(4) Reinforcement Networksrefer to learning schemes

which iterate through a number of steps until per-

formance no longer improves. A set of weights

is selected in some methodical or semi-random

way depending on previously selected and saved

weights. Then, the performance of the network is

assessed by running the training data through the

network and evaluating an objective function. If the

weights show an improved performance, they are

saved, replacing some previously saved weights.

Performance measures

One of the main indicators to evaluate the performance

of a neural network classifier is the Classification Rate.

This indicator provides a measure of the correctly clas-sified inputs and is best depicted using the classification

rate matrix shown in Fig. 7.

The columns of the matrix represent the actual or

desired results whereas the rows represent the neural

network estimates or outputs. The body of the Classifi-

cation Rate matrix represents the intersection between

desired results and actual network predictions. A value

of 1.0 for any given cell means that all desired out-

puts for that category were correctly predicted by the

Springer


11/17

Nonlinear Dyn (2007) 49:493509 503

Fig. 7 Classification ratematrix

network. Similarly, a value of 0.0 in a cell means that

none of the desired outputs were predicted to be a mem-

berof that category. A perfect classifierwould comprise

a value of 1.0 for the correctly classified states and zero

for the Type I and Type II errors. Classification rates ob-

tained during the training or model development stageprovide a measure of the calibration results while clas-

sification rates obtained during testing provide an indi-

cation of the generalisation ability and validity of the

model.

Learning rules and transfer functions

The learning rule represents the mathematical algo-

rithm used to determine the increment or decrement

by which weights of a processing element (PE) change

during the learning phase. Examples of such algo-rithms include Maxprop, Delta Prop and Genetic Al-

gorithms [11]. A transfer function is typically a non-

linear function that transforms the weighted sum of the

effective inputs to a potential output value [11]. The se-

lection of learning rules and transfer functions depends

on the individual ANN architecture and are typically

arrived at through comprehensive testing. The learn-

ing rules, transfer functions and other parameters that

were selected and tested in this study are shown in

Table 11.

A large number of experiments were run to de-termine the best architecture and set of parameters

for the route choice compliance problem. These in-

cluded five different Fuzzy ARTMAP architectures

(each with a different mapping layer); 10 different

back-propagation and Radial Basis architectures and

12 Reinforcementnetwork models. In total,185 models

with different architectures were developed and tested.

Different experiments were run using the categorised

and normalised data. The best performance models are

shown in Tables 12 and 13 for the categorised and nor-

malised input data, respectively.

The results in Tables 12 and 13 show the superior

performance of the Fuzzy ARTMAP and RBFN archi-

tectures over other neural network architectures. The

results also show an improved classification rate whendata is categorised. These results suggest that predic-

tion errors between the ARTMAP and RBFN models

are negligible and that adoption of either model is ac-

ceptable for each of the five ATIS scenarios. Closer

inspection of the model formulations, however, re-

vealed that the RBFN model had fewer parameters

to calibrate than the Fuzzy ARTMAP model and was

hence selected as the architecture for implementation

in the traffic simulator. In the remainder of this docu-

ment, this model will be referred to as the Neugent

model.

Interface to microscopic simulation

To demonstrate the feasibility of the approach, the Neu-

gent driver behaviour model was interfaced to a traf-

fic simulation tool. The traffic simulator was used to

govern individual vehicular movements on the traffic-

commuting corridor using car following, lane chang-

ing and gap acceptance behaviour. Given the static

characteristics of the commuting corridor network, itslink capacities, connections and speeds, the simulator

takes a time-dependent loading pattern and handles the

movement of individual vehicles on links as well as

the transfer between links. The routing of individual

vehicles under the influence of real-time information

is then provided by the driver behaviour model devel-

oped in this study. This model specifies how a particu-

lar drivervehicleunit (DVU) with given goals, char-

acteristics, preferences and knowledge of prevailing

Springer


12/17

504 Nonlinear Dyn (2007) 49:493509

Table 11 Neural network architectures

Model Model Learning Transfer Mapping

architecture description rule function layer (F2)

Fuzzy ARTMAP network Baseline vigilance = 0.0 (between 01.0) N.A. N.A. 50

Recode rate = 0.5 (between 0.21.0) 100

Choice parameter= 0.1 150

Error tolerance = 0.01 200

250

Back-propagation network LCoef (Learning rate) 0.3 for 1st hidden,

0.25 for 2nd hidden, 0.15 for output Delta rule TanH N.A.

Momentum = 0.4 Nom-Cum Linear

Trans. Pt. = 10,000 Ext DBD Sigmoid

LCoef Ratio = 0.5 QuickProp DNNA

F offset = 0.1 MaxProp Sine

DBD

Radial basis function networks LCoef (Learning rate) 0.3 for 250 Proto,

0.25 for hidden, 0.15 for output Delta rule TanH N.A.

Momentum = 0.4 Nom-Cum Linear

Trans. Pt. = 10000 Ext DBD SigmoidLCoef Ratio = 0.5 QuickProp DNNA

Max. Trans. = 2000 MaxProp Sine

DBD

Reinforcement networks LCoef (Learning rate) 0.3 for 250 Proto,

0.25 for hidden 0.15 for output Genetic TanH N.A.

Momentum = 0.4 DRS Linear

Trans. Pt. = 10000 Sigmoid

LCoef Ratio = 0.5 Exp

Max. Trans. = 2000 Perceptron

Sine

conditions, selects the next link under the influence oftraffic information.

The commercially available microscopic traffic sim-

ulation model AIMSUN [19] was selected for model

testing. The selection was based on a comparative eval-

uation of the car following behaviour in a number of

simulation tools which showed that the Gipps model

implemented in AIMSUN performed better than the

family of psychophysical car following models imple-

mented in PARAMICS and VISSIM [13]. This sim-

ulation software also has a number of unique fea-

tures which allow the user to customise many fea-tures of the underlying simulation model through an

application programming interface (API). This feature

is critical for this study, as it will allow for interfac-

ing the driver behaviour model with the traffic simu-

lation tool. The traffic simulation will follow a deter-

ministic, fixed time-step approach. At each time step

(e.g. 1 s), vehicles will be moved at the prevailing lo-

cal speed on the same link or transferred to another

link. The latter is determined by the driver behaviour

model according to user preferences, perceptions, path-switching thresholds and other factors identified and

discussed in this study. Travel times are then calcu-

lated from each node based on the current link speeds

and the simulation continues until all DVUs reach their

destinations.

For each of the ATIS scenarios presented in Ta-

ble 7, the Neugent route choice model was coded

and interfaced to a hypothetical AIMSUN network to

evaluate the impacts of providing drivers with travel

information.

A hypothetical case study

The Neugent model developed in this study was tested

on a simple hypothetical network (Fig. 8) in which

three routes between zones A and D are clearly marked

(main route ABCD and alternative routes AEC

D and AEFD). Intersections A, B, C and D are all

signalised.

Springer


13/17

Nonlinear Dyn (2007) 49:493509 505

Table 12 Best performance models using categorised input data

ATIS scenario ANN architecture Learning rule Transfer function Classification rate (CR)

Qualitative delay information Fuzzy ARTMAP (150a) N.A. N.A. 0.9636b

Back-Prop MaxProp Sine 0.5919

RBFN Delta rule Sine 0.9636b

Reinforcement Genetic Perceptron 0.7237

Quantitative real-time delay information Fuzzy ARTMAP (100a) N.A. N.A. 0.9661b


RBFN DBD Sine 0.9412c

Reinforcement Genetic Sine 0.7802

Quantitative real-time delay on best Fuzzy ARTMAP (250a) N.A. N.A. 0.9545c

alternative route

Back-Prop MaxProp TanH 0.6162

RBFN Delta Rule Sine 0.9589b

Reinforcement DRS Perceptron 0.7537

Predictive delay information Fuzzy ARTMAP (50a) N.A. N.A. 0.9615b

Back-Prop Delta rule TanH 0.5844

RBFN Delta rule Sine 0.9252c

Reinforcement Genetic Sine 0.7436Prescriptive best alternative route Fuzzy ARTMAP (250a) N.A. N.A. 0.9732b

Back-Prop DBD Linear 0.6375

RBFN Delta rule TanH 0.9488c

Reinforcement Genetic Linear 0.7815

aNumber of mapping layers (F2 units) for the Fuzzy ARTMAP networkbBest model based on classification ratecSecond best model based on classification rate

Fig. 8 A hypotheticalnetwork with three

alternative routes

Springer


14/17

506 Nonlinear Dyn (2007) 49:493509

Table 13 Best performance models using normalised input data

ATIS scenario ANN architecture Learning rule Transfer function Classification rate (CR)

Qualitative delay information Fuzzy ARTMAP (250a) N.A. N.A. 0.9663b


RBFN Delta rule Sine 0.9636c

Reinforcement Genetic Perceptron 0.7205

Quantitative real-time delay information Fuzzy ARTMAP (200a) N.A. N.A. 0.9661b


RBFN DBD Sine 0.9412c

Reinforcement Genetic Sine 0.7548

Quantitative real-time delay on best Fuzzy ARTMAP (150a) N.A. N.A. 0.9589b

alternative route

Back-Prop Nom-Cum TanH 0.7711c

RBFN Nom-Cum TanH 0.7711c

Reinforcement DRS Perceptron 0.6902

Predictive delay information Fuzzy ARTMAP (100a) N.A. N.A. 0.9615b

Back-Prop Delta rule TanH 0.5919

RBFN Nom-Cum TanH 0.8889c

Reinforcement Genetic Sine 0.8889Prescriptive best alternative route Fuzzy ARTMAP (200a) N.A. N.A. 0.9732b

Back-Prop DBD Linear 0.5667

RBFN Delta rule TanH 0.8667c

Reinforcement Genetic Linear 0.7929

aA total number of F2 units for Fuzzy ARTMAPbBest model based on classification ratecSecond best model based on classification rate

The main route commonly travelled by drivers

is route ABCD. Drivers also have the option of

diverting through routes AECD or AEFD incase of congestion on the main route. Familiar drivers

are assumed more likely to choose route AEFD

to avoid the delays caused by the traffic signals on the

main route. A variable message sign (VMS) is also sim-

ulated in this experiment and was strategically located

at a control point before node A to provide drivers with

information about delays along the main and alternative

routes. DVUs approaching the control point will assess

their environments (using the driver behaviour mod-

els) and decide whether they are going to comply with

the information provided on the VMS. Drivers whohave in-vehicle devices can access real-time traffic in-

formation through the device while driving. Drivers

responses and compliance with traffic information (for

both VMS and in-vehicle device) is provided by the

Neugent driver behavioural model developed in this

study. The following factors affecting drivers compli-

ance were considered in this case study: driver aggres-

sion, driver awareness or familiarity with the network,

VMS credibility and delay tolerancethresholds for both

familiar and unfamiliar drivers. An interface which ac-

tivates route guidance within the model via variable

message signs (VMS) or in-vehicle devices was alsodeveloped for this experiment (Fig. 9).

The plug-in allows manipulation of the ITS infor-

mation strategies influencing driver compliance via the

graphical user interface (GUI). Under non-congested

conditions, the VMS would be blank, indicating that

conditions are free-flowing. As congestion builds on

the main route and reaches a threshold (determined

by the minimum delay and minimum travel time),

the VMS is activated and displays delay informa-

tion to drivers. Each driver is assigned a set of val-

ues (from a distribution function) representing age,gender, aggressiveness, awareness, acceptable delay

threshold and familiarity with network conditions as

determined from the behavioural survey. The plug-

in further assumes that drivers are classified into one

of the following three categories: informed drivers

(e.g. drivers who subscribe to en route traffic infor-

mation services); familiar drivers (e.g. drivers who

have lived in the designated area for an extended pe-

riod of time and are familiar with traffic conditions on

Springer


15/17

Nonlinear Dyn (2007) 49:493509 507

Table 14 Additional benefits from implementing dynamic driver compliance (Neugent) models

Percentage improvements over static C-Logit

Network Network Stop time Network travel

ATIS scenario speed delay per vehicle time

Qualitative delay information 4.12 5.43 6.86 1.74

Quantitative real-time delay information 4.12 6.09 8.28 0.73Quantitative real-time delay on best alternative route 7.06a 6.90 8.12 3.10

Predictive delay information 4.71 6.90 8.52 3.67a

Prescriptive best alternative route 5.29 8.22a 10.57a 2.09

aBest model result

Fig. 9 GUI linking the neugent model and the traffic simulator

the network); and unfamiliar drivers (e.g. tourists or

drivers not familiar with the road network). The last

two types of drivers are assumed to have no access to

in-vehicle route guidance and that the only information

they receive would be from VMS on the side of the

road. It was further assumed that the three driver cat-egories were equally represented. The simulation was

run for 1 h including a start-up period of 5 min. Default

car following and lane changing parameters were used

to govern vehicular movement between origins and

destinations.

Simulation results

The aim of the simulation experiments was to demon-

strate the additional benefits that can be obtained using

models of driver compliance with traffic informationcompared to what can be obtained using default route

choice models. The base case scenario was therefore

based on the simulators default route choice model

(C-Logit) which does not take into account drivers

compliance with information. The comparative evalu-

ation results are reported in Table 14.

These aggregate results show that quantitative real-

time delay; predictive real-time delay and prescrip-

tive information on best alternative route provide the

best overall benefits in terms of network speed, delay

and network travel times. In other words, these ATISstrategies or scenarios are best used to influence driver

behaviour and impact their compliance rates. These

findings are also supported by a number of other stud-

ies in the literature [14]. Furthermore, the evaluation

results which compared the benefits of the Neugent

model over C-Logit clearly show improvements of 4

7% in network speeds, 58% in network delays, 711%

in stop time per vehicle and 13% in network travel

time.

Springer


16/17

508 Nonlinear Dyn (2007) 49:493509

Conclusions

The results reported in this paper provide a useful in-

sight into commuters needs and preferences for traffic

information in the Brisbane metropolitan area. Unlike

other studies which relied on generalised driver char-

acteristics, this study was based on a field behaviouralsurvey of drivers which was conducted on a congested

real-world commuting corridor. A substantial amount

of information regarding driver preferences and be-

haviour were collected and a number of neural network

agent-based dynamic driver behaviour models were de-

veloped to determine the factors affecting driver com-

pliance and response to a variety of traffic information

sources. These models clearly indicated that prescrip-

tive, predictive and quantitative real-time delay infor-

mation provided for both the usual and best alternate

routes were most effective in influencing respondentsto change their routes. The behavioural models describ-

ing drivers route choice decisions were interfaced to

a microscopic traffic simulation tool which was cali-

brated to simulate the movement of individual vehicles

between their origins and destinations. The simula-

tion results supported the notions that commuters de-

cisions to divert to alternate routes are influenced by

their socio-economic characteristics; the degree of fa-

miliarity with network conditions and the expectation

of an improvement in travel time that exceeds a certain

delay threshold associated with each individual. Theimplementation of the Neugent models developed in

this study have been shown to produce benefits over

the C-Logit route choice algorithm with improvements

of 47% in network speeds, 58% in network delays,

711% in stop time per vehicle and 13% in network

travel times.

Acknowledgements The authors would like to thank theanony-mous reviewers for their valuable comments and constructivefeedback on the paper. The authors also acknowledge the as-

sistance provided by David Lockie, Katherine Johnston, DanielHarney and Andrew Boyle with the behavioural survey, data col-lection and pre-processing.

Disclaimer

This paper presents the opinion of the authors based on

their research results and does not necessarily represent

the views or policy of the ITS Research Laboratory

at the University of Queensland. This paper does not

constitute a recommendation or standard, and it is for

general information only and should not be relied upon

situationally or circumstantially. Neither the authors

nor the University are responsible for the use which

might be made of these results.

References

1. Ben-Akiva, M., Palma, A.D., Kaysi, I.: Dynamic networkmodels and driver information systems. Transportation Res.Part A25A(5), 251266 (1991)

2. Dia, H., Harney, D., Boyle, A.: Analysis of commutersresponses to travel information: Implications for ATIS de-ploymentin Brisbane.In: The22ndConferenceof AustralianInstitutes of Transport Research (CAITR), The AustralianNational University, Canberra, ACT, Australia

3. Dougherty, M.:A review of neuralnetworksapplied to trans-

port. Transport. Res. Part C: Emerg. Technol.3(4), 247260(1995)

4. Greene, W.H.: Econometric Analysis. Prentice-Hall, NewJersey (2000)

5. Hawas, Y.E.: Development and calibration of route choiceutility models: factorial experimental design approach.J. Transport. Eng. Am. Soc. Civil Eng. 30(2), 159170(2004)

6. Hensher, D.A., Ton, T.T.: A comparison of the predictivepotentialof artificialneural networksand nestedlogit modelsforcommuter mode choice. Transport. Res. Part E: LogisticsTransport. Rev.36(3), 155172 (2000)

7. Khattak, A., Polydoropoulou, A., Ben-Akiva, M.: Model-ing revealed and stated pretrip travel response to advanced

traveler information systems. Transport. Res. Rec. (1537),4654 (1996)

8. Koutsopoulos, H., Polydoropoulou, A., Ben-Akiva, M.:Travel simulators for data collection on driver behavior inthe presence of information. Transport. Res. Part C 3(3),143159 (1995)

9. Lai, K.-H., Wong, W.-G.: SP approach toward driver com-prehension of message formats on VMS. ASCE J. TransportEng.126(3), 221227 (2000)

10. NeuralWare: Advanced Reference Guide: A Software Ref-erence for NeuralWorks Professional II/PLUS, Carnegie, PA(2001)

11. NeuralWare: NeuralWorks Professional II/PLUS: GettingStarted A Tutorial for Microsoft Windows Computers,Carnegie, PA (2001)

12. Pang, G.K.H., Takahashi, K., Yokota, T., Takenaga, H.:Adaptive route selection for dynamic route guidance systembased on fuzzy-neural approaches. IEEE Trans. VehicularTech.48(6), 20282041 (1999)

13. Panwai, S., Dia, H.: Comparative evaluation of microscopiccar-following behavior. IEEE Trans. Intell. Transport. Syst.6(3), 314325 (2005)

14. Peeta, S., Poonuru, K., Sinha, K.: Evaluation ofmobility impacts of advanced information sys-tems. ASCE J. Transport. Eng. 126(3), 212220(2000)

Springer


17/17

Nonlinear Dyn (2007) 49:493509 509

15. Polydoropoulou, A., Ben-Akiva, M., Khattak, A., Lauprete,G.: Modeling revealed and stated en-route travel response toadvanced traveler information systems. Transport. Res. Rec.(1537), 3845 (1996)

16. Polydoropoulou, A., Khattak, A.: A comparative analysis ofspatial knowledge and en route diversion behavior acrossChicago and San Francisco: implications for advanced trav-

eler information systems. Transport. Res. Rec. (1621), 2735(1998)

17. Ramming, M.S.: Network Knowledge and Route Choice.Ph.D. thesis, Massachusetts Institute of Technology (MIT),Cambridge, MA (2002)

18. Schofer, J., Khattak, A., Koppelman, F.: Behavioral is-sues in the design and evaluation of advanced travellerinformation systems. Transport. Res. Part C1(2), 107117(1993)

19. TSS: Transport Simulation Systems, GETRAM/AIMSUNversion 4.1 Users Manuals (2002)

Springer

vms evaluation

Documents