using intelligent agents for transportation regulation support system design

Transportation Research Part C 18 (2010) 140–156

Contents lists available at ScienceDirect

Transportation Research Part C

journal homepage: www.elsevier .com/locate / t rc

Using intelligent agents for Transportation Regulation SupportSystem design

Flavien Balbo a,b,*, Suzanne Pinson a

a University Paris-Dauphine – Lamsade, Place du Maréchal de Lattre de Tassigny, F-75775 Paris 16 cedex, Franceb Inrets – Gretia, Le Descartes 2, 2, Rue de la Butte Verte, F-93166 Noisy-le-Grand cedex, France

a r t i c l e i n f o

Article history:Received 15 October 2008Received in revised form 5 March 2009Accepted 29 April 2009

Keywords:Agent-based applicationsDecision Support SystemPublic transportation network managementBus network

0968-090X/$ - see front matter � 2009 Elsevier Ltddoi:10.1016/j.trc.2009.04.019

* Corresponding author. Address: University ParisE-mail addresses: [email protected] (F. B

1 http://www.ewh.ieee.org/tc/its/.

a b s t r a c t

This paper presents an agent-based approach used to design a Transportation RegulationSupport System (TRSS), that reports the network activity in real-time and thus assiststhe bus network regulators. The objective is to combine the functionalities of the existinginformation system with the functionalities of a decision support system in order to pro-pose a generic model of a traffic regulation support system. Unlike the other approachesthat only deal with a specific task, the original feature of our generic model is that it pro-poses a global approach to the regulation function under normal conditions (network mon-itoring, dynamic timetable management) and under disrupted conditions (disturbanceassessment and action planning of feasible solutions). Following the introduction, the sec-ond section presents the notions of the domain and highlights the main regulation prob-lems. The third section details and motivates our choice of the components of thegeneric model. Based on our generic model, in the fourth section, we present a TRSS pro-totype called SATIR (Système Automatique de Traitement des Incidents en Réseau – Auto-matic System for Network Incident Processing) that we have developed. SATIR has beentested on the Brussels transportation network (STIB). The results are presented in the fifthsection. Lastly, we show how using the multi-agent paradigm opens perspectives regardingthe development of new functionalities to improve the management of a bus network.

� 2009 Elsevier Ltd. All rights reserved.

1. Introduction

The development of surface public transportation networks is a major issue in terms of ecology, economy and society. Toimprove its attractiveness, the urban networks must increase their quality in terms of punctuality and vehicle frequencywhile at the same time they must decrease management costs. A project like the Bus Rapid Transit shows the benefits ofimproving infrastructures; but better management of the available resources is less costly than improving network infra-structures. Intelligent Transportation Systems1 (ITS), based on synergy between new information technologies for simulation,real-time control, and communications networks are an alternative to improve available resource management. Urban trafficcontrol (UTC) systems are ITS enabling a better real-time management of available resources. The usability and the effectivenessof the UTC systems greatly depends on their ability to locate, assess and react to traffic disturbances.

In order to automate the transportation activity, the theoretical bus supply is computed. It gives the transportation planwhich represents the optimum supply in a theoretical context. It may become obsolete as the urban traffic conditions evolve.Regulators (the staff in charge of monitoring the bus networks) have to ensure the success of the transportation plan, in the

. All rights reserved.

-Dauphine – Lamsade, Place du Maréchal de Lattre de Tassigny, F-75775 Paris 16 cedex, France.albo), [email protected] (S. Pinson).

http://dx.doi.org/10.1016/j.trc.2009.04.019

mailto:[email protected]

mailto:[email protected]

http://www.ewh.ieee.org/tc/its/

http://www.sciencedirect.com/science/journal/0968090X

http://www.elsevier.com/locate/trc

F. Balbo, S. Pinson / Transportation Research Part C 18 (2010) 140–156 141

sense of adapting theoretical supply to satisfy the passenger demand according to the urban traffic disturbances. Regulatorsuse UTC systems known as Automatic Vehicle Monitoring (AVM) systems in order to collect and display data. The use of anAVM system is the first step to the computerization of the transportation network activity. However this system is limited todetecting disturbances linked to unanticipated demands and to traffic conditions but is not able to deal with difficulties re-lated to the real-time management of the bus network: managing the inconsistencies of data coming from sensors that lo-cate the vehicles, assessing a disturbance according to its context as well as proposing feasible solutions. These limits are dueto the inadequacy of the data collecting, shaping and displaying processes. To cope with these limits, we propose to completethe AVM system with a Decision Support System (DSS) in order to analyze the data so as to give a dynamic and contextualassessment of the disturbances in real-time, as well as action planning and decision-making aid.

In this paper, we study the different ways to integrate an AVM system and a DSS into what we call a Transportation Reg-ulation Support System (TRSS) that is to say how to build a TRSS with the functionalities of both the AVM system and theDSS. Then, we show that the best way to model efficiently these functionalities is to use a multi-agent paradigm. Agent-based DSS are particularly relevant in domains where human operators have to make operational decisions regarding themanagement of complex organizational processes that are inherently distributed (spatially, logically and/or physically)(Ossowski et al., 2004; Vahidov and Fazlollahi, 2004). The autonomy of a multi-agent system where distributed entities,called agents, interact with each other and its ability to adapt and react to the changes in the environment are essentialin the field of transportation where the environment is dynamic, open and uncertain.

Section 2 describes real-time management of urban transportation networks and underlines their advantages and limits.Section 3 highlights different ways of integrating an AVM and a DSS to give a generic Transportation Regulation Support Sys-tem. It motivates the different options that we have taken based on drawbacks of existing systems: the choice of an efficientarchitecture, the choice of a multi-agents model to support this architecture and the choice of a specific environment modelto cope with the topological and temporal characteristics of the transportation domain. This generic model is then used inSection 4 to design a traffic regulation system in the case of the real-time management of the Belgium urban network (STIBnetwork). Section 5 presents our experimentation and conclusions are drawn in Section 6.

2. Urban network regulation: notions of the domain

This section describes how information processing and task processing are performed in Urban Traffic Control systems.The first part describes the functionalities of the AVM system as well as the general data model based on it. The second partpresents the regulation tasks and the drawbacks of existing regulation processes.

2.1. The Automatic Vehicle Monitoring system (AVM)

2.1.1. The AVM functionalitiesIn urban transportation control domain, human regulators are located in a control center. They have to manage the trans-

portation network under normal operating conditions (where are the buses located?) and also under disturbed conditions(where are disturbances – bus delays, bus advances – located?). What action has to be taken to solve the problem?

In most networks, vehicles are located through sensors which provide real-time information. This information representsa huge amount of data (for example data arrives every 40 s in the STIB network). Furthermore it may be incomplete (a sensormay break down) or uncertain (the quality of the data may be poor). This data is collected through the Automatic VehicleMonitoring system (AVM). The AVM system compares the actual positions of the vehicles (captured by the sensors) withtheir theoretical positions given by pre-registered timetables in order detect disturbances representing by alarms on thescreen (color code in Fig. 1). In this way, the regulator can see whether the vehicles are running ahead of timetable or arerunning late.

Fig. 1 shows the AVM management of real-time information coming from sensors and the output of the system. Each lineis represented two ways with its stops and its running buses. Each bus location is represented by (1) a number for its the-oretical position coming from theoretical timetables and (2) a colored square for its real location detected by the system. Thisreal location may be erroneous due to sensor break downs. Stops are represented by black dots. The gap between the the-oretical position and the real position gives an information about the bus delays or advances. Colors give the importance ofthe delays or advances. Some AVM systems incorporate geographical criteria such as delay/advance alarms in a town-center,time criteria such as detecting a delay on the next scheduled departure. The role of the AVM system is to compute onlinebasic information, organizes data collecting and displaying and computes alarms. Its screen interface facilitates the accessto this data by allowing to click on a bus number to get more information on this bus or to click on a given stop to get moreinformation on this stop (see Fig. 1). Given this information, the regulators have to rely on their own experience to decideupon the regulation actions to be taken. In this paper, we propose to automate the decision-making process to prevent it torely on regulators’ experience giving a generic model of urban Transportation Regulation Support Systems (TRSS).

2.1.2. The data modelA significant amount of work has been done to model data in automatic traffic regulation systems. The objectives were to

represent both the physical network configuration and the bus timetables in a same model. The first project of a data model

Fig. 1. The AVM interface.

142 F. Balbo, S. Pinson / Transportation Research Part C 18 (2010) 140–156

in Europe was CASSIOPE (1989–1992), which led to the Transmodel.2 This is the European reference for conceptual data mod-eling in public transportation domains. Work on subsequent European projects (Eurobus,2 1992–1994 and Harpist,2 1995) cul-minated in a finished and fully validated product, Titan (Transmodel-based Integration of Transport Applications andNormalization, 1996–1998). Recently, more specific standard like IFOPT3 (Identification of Fixed Objects in Public Transport)have been developed. The result is a modeling approach based on the Entity/Association or UML formalism. A hierarchicaldecomposition of the information has been adopted, i.e. the physical configuration is divided into lines, each of which is sub-divided into routes. A route is broken down into sections, and each one has a stop and an inter-stop distance. A timetable isdivided into missions, each of which is subdivided into runs. This model has been chosen in order to satisfy specific needs: easyaccess to a specific component of a timetable, for example the description of the route of a run, or of a section to get the list ofmissions in a given time period. Recently, the standard model was completed by an XML protocol to take into account data ex-change and specifications SIRI4 allowing distributed computers to exchange real-time information about public transport ser-vices and vehicles.

2.2. The regulation task

This section describes the regulation process and highlights the limits of the existing AVM systems dedicated to real-timemanagement of urban networks. It goes on to propose new functionalities for the TRSS to support the regulators’ work.

2.2.1. The regulation processDifferent studies of regulators’ work usually identify four phases in the regulation process: a monitoring phase, a diag-

nostic phase, a planning phase and a decision phase.

(i) In the first phase, the regulators monitor the network. They collect and aggregate relevant information and analyzewhat it means.

(ii) The diagnostic phase begins with the detection of a disturbance and ends with the assessment of its consequences onthe network activity according to the context. The interface of the AVM system facilitates this phase (see Fig. 1). Thecurrent alarms, which are visually accessible on the screen, save the regulators from having to make computations. Assoon as disturbances are detected by the system, the regulator chose one of them to deal with (see Fig. 2 arrow #1), hehas to complete his knowledge of the problem i.e. he has to find out the delayed bus timetable, network state . . .

2 http://www.transmodel.org.3 http://www.kizoom.com/standards/ifopt/.4 http://www.kizoom.com/standards/siri/.

http://www.transmodel.org

http://www.kizoom.com/standards/ifopt/

http://www.kizoom.com/standards/siri/

Fig. 2. Regulators’ main tasks in urban regulation transportation systems. (For interpretation to colours in this figure, the reader is referred to the webversion of this paper.)


(iii) When a disturbance has been detected and assessed, the planning phase begins. The type of risk (risk of a bus trainand/or risk of a gap) indicates to the regulator what the primary objective is i.e. to increase or decrease the bus supplyin one part of the network (see Fig. 2 arrow #2). He computes the feasible procedures according to the current state ofthe network (see Fig. 2 arrow #3). This complex process involves various information sources including real-timeinformation coming from sensors, theoretical timetables or information coming from drivers and from otherregulators.

(iv) Finally, there is the decision phase, which is made up of two parts: firstly the regulator chooses the regulation proce-dure according to the disturbance assessment (see Fig. 2 arrow #2), then he monitors its execution. Since the planningphase gives a set of feasible procedures (see Fig. 2 arrow #4), the regulator has to choose a given procedure taking intoaccount different contradictory constraints. For example, constraints such as empty runs of buses or failures of theregulation procedures have to be limited. Once the regulator has chosen a procedure, he must check that the requiredresources and the state of the network evolve according to his forecast; if they do not, he may have to make anotherchoice.

Fig. 2 shows the tasks of regulators and how they are interrelated. Because the regulators are involved in all of thesenumerous tasks, their work is very complex; they spend too much of their time monitoring the network rather than focusingon decision-making (based on the results of the diagnosis) and planning tasks, which should be their main activity. As we cansee, they really need an efficient Decision Support System to help them.

2.2.2. Regulation process issuesThe main advantage of the AVM system is to facilitate access to a high number of different information sources, but the

regulator has to use his experience to extract the right information at the right time. This means that the regulation processis strongly dependent on the regulator’s experience. If we want that the regulation process be standardized and in order tobetter support the regulator’s task, we propose a generic regulation process model. Before proposing the model, we highlightfour main problems arising in human regulation processes and that should be taken into account.

The first problem is related to the quality and the completeness of collected data. The AVM system collects informationbut this data may be incomplete, for example if a sensor breaks down. Another problem is that regulators do not have thedata that leads to theoretical timetable creation since they are computed at the beginning of the regulation day and pre-reg-istered in the system. For example, near a school, the regulation process should be adapted but regulators in the STIB net-work do not have this topological information which is essential for a better network regulation. Furthermore, thesetheoretical timetables are computed according to the theory of the domain encompassed in rules (called logics). These logicsare beyond the scope of this paper, but it is sufficient to say that each of them is related to a specific objective. For instance,the aim of the logic of regularity is to minimize the average waiting time of the passengers and spread out the supply of allbuses on the line. On the contrary, the logic of passenger pick up corresponds to a heavy local demand. In this case, the time-tables are computed in order to concentrate resources at the most critical points in the line in order to satisfy the passengers.


Consequently, in case of a disturbance, the regulation process should not be based on the theoretical timetables but on theregulation logics which are only known by experienced regulators thanks to their knowledge of the network. For example, ifall the vehicles are late but the interval between them remains stable, there are no disturbances in the logic of regularity butthere are disturbances in the punctuality logic.

The second problem is related to the distribution of the regulator’s attention. Because of heavy urban traffic, many vehi-cles in the transportation network are late, especially at peak hours, and it is impossible for regulators to take all these delaysinto account. Experienced regulators use their knowledge of the line structure (position in the city, presence of difficultareas) and of the demand structure to determine the most critical lines according to the timetable. For instance, informationindicating that a vehicle is too late to do its next run may be of not interest if this vehicle can make up for lost time on the lastpart of the run. Regulators use the computed alarms as primary indicators of a disturbance and analyze its importanceaccording to the context.

The third problem is related to the assessment of a disturbance by regulators and its evolution on the network. This pro-cess is complex because disturbances evolve independently along three axes. The time axis measures the seriousness of adisturbance according to the timetable; the space axis measures the seriousness of a disturbance according to its positionon the network; and the shape axis measures the consequences a disturbance may have on the network activity. Regulatorshave to determine the importance of disturbances according to these three axes. For example, a vehicle having off-peak hourdifficulties in a suburb (a disturbance that is not normally considered to be critical) may cause a real problem if bus fre-quency is low. Once again, this assessment is based on the regulator’s experience and the estimated state of the network.

The fourth and final problem comes from the lack of a global vision. Given the fact that the monitoring process is done byline, and given the high number of lines to be monitored (for example in the STIB network, each regulator tracks 13 lineswith 5–20 buses running during the day), global management of the network is impossible. Regulators use their experienceto create relation between disturbances without visible links, in order to propose global solutions.

To cope with these problems, we propose in the following section a generic model of the regulation process which doesnot rely on regulators’ experience. It integrates a Decision Support System and the AVM information processing system.

3. Integrating a DSS and an AVM System: a generic model of the regulation process

3.1. Choice of the system architecture

The problem is to find the right place of the DSS in the information processing phase. Two types of cooperation between ahuman operator and a computer system in decision-making processes can be chosen (Rasmussen et al., 1990; Rouse, 1981):(1) a horizontal cooperation and (2) a vertical cooperation. The difference between the two of them depends on the distri-bution of control. In the first case, the operator and the DSS share dynamically the tasks to be performed (Crevits et al., 2002).In the second case, the DSS is used as a guide to support decisions. This section discusses the choice of the cooperation typethat is the best suited to build Transportation Regulation Support Systems. Recall that in transportation domain, the infor-mation processing system is called the AVM system.

In an horizontal architecture, the information processing system feeds the DSS with data and the DSS modifies the infor-mation system data in a loop giving an autonomous system; in this type of architecture, the only task of the operator is tosupervise this information processing. Due to the poor quality of data and the limited formalization of the domain knowl-edge, the choice of an autonomous architecture is not yet realistic in urban network management. Although the generaliza-tion of localization technologies such as GPS (Global Positioning System) will solve this problem, this technical improvementcannot easily replace qualitative information coming from drivers. For example, the management of a disturbance related toa car parked in a wrong place is not the same as that related to an accident involving persons. An autonomous system willdetect that the vehicle is late but will not be able to know why, making difficult to assess the disturbance and consequentlyto automatically propose a solution.

In a vertical architecture, the DSS is considered as a solution server and may base its decisions on a simulation tool or onreal data coming from the AVM information system. In the first case, the simulation tool is used as a what-if model. Forexample, Brezillon et al. (1999) proposes a DSS for the management of the Paris metro based on a simulator which is acti-vated by the regulator in the case of a disturbance. This simulator is fed with data filtered by the regulator who thereforesolves the quality-data problem. In this case, the difficulty is to manage disturbances in real-time in a dynamic environmentbecause the regulator has to perform information filtering. In the second case, a solution is to link the DSS to the data relatedto the network through the AVM system giving a real-time DSS (see Fig. 3). In this proposed solution, the AVM informationsystem keeps its initial function: collecting and displaying data, thus allowing the DSS to access useful data for reasoning.The main advantage of this architecture is that the AVM system may be used as it is, but the limitations are threefold:(1) the regulator has to deal with two interfaces which could be misleading; (2) certain data is duplicated which is alwaysinefficient in term of storage, processing and updating. For example, the positions of buses has to be duplicated because theyare used by the AVM system to display data through the interface and to compute the alarms, and by the DSS in order tocompute a solution. In this architecture, without data duplication, it is impossible to assess the seriousness of a disturbances:how, for example, could the DSS detect whether the delay of a bus is due to a traffic problem or to a failure in the detection ofthe bus position (a sensor has broken down) without a comparison with the previous position of the bus or the previous

Fig. 3. Vertical architecture of a DSS and an AVM system. (For interpretation to colours in this figure, the reader is referred to the web version of this paper.)


records of the sensor? (3) another organizational problem in the vertical architecture is the duplication of tasks. The AVMsystem may already have some of the functionalities that should be incorporated in the DSS, such as alarm computation.

Keeping in mind these limits, we propose a generic model of a transportation regulation system that we call a Transpor-tation Regulation Support System (TRSS). It is based on a global approach that encompasses the four phases of human reg-ulators (c.f. Section 2.2.1): network monitoring through information synthesis, disturbance diagnosis, action planning anddecision-making (see Fig. 4). It takes into account real-time data coming from buses as well as pre-registered theoreticalinformation. Our generic model is built on a multi-agent framework where agents encompass the domain knowledge andwhere the interactions between agents are based on a new model of environment (see Section 3.3). An original feature ofthis model is that the same multi-agent model is used to process data and to find solutions: (1) network monitoring is pro-cessed through dynamic timetable management; (2) distributed diagnosis is based on an original model of disturbance con-sidering the disturbance context and its evolution; (3) feasible solutions are computed taking into account the context andprofiles of vehicles. These three points will be developed in the following sections. One advantage of this generic model isthat it solves the problem of duplication of data. The same data is used to manage the transportation network under normaloperating conditions (monitoring function) and under disturbed conditions (diagnostic and planning functions).

In order to design a multi-agents system, several components have to be defined precisely: the agents, the interactions,the environment and the agent organization. In the next section we justify our choice of the multi-agent paradigm and in thefollowing section we underline the need for a specific environment model.

3.2. Choice of a multi-agent modeling approach

The agent technology is based on the notion of reactive, autonomous, proactive entities that evolve in a dynamic environ-ment. In open systems, agents may appear and disappear over time as it is the case of buses on lines. In our generic model,we propose two categories of agents:

– STOP agents that represent the theoretical structure of the network (organized in lines and routes). They encompass theknowledge of the graph makers: passenger flows and traffic problems used to make up the theoretical timetables. A STOPknows the identifiers of STOP agents situated just before/after it on the same line and route.

– BUS agents that represent the dynamic part of the network. Each BUS agent is the abstract model of an actual vehicle run-ning on the transportation network and reports its movements to the STOP agents.

These agents represent the network information system where the access points are the vehicles and the stops. For in-stance, the identification of a vehicle is used to access the driver timetables.

As we said, the multi-agent paradigm is well adapted to the transportation domain study since it facilitates an approachby analogy in a domain where the objective is the management of distributed entities (the buses). One advantage of thisapproach is the explicit representation of the processes that it models. A multi-agent system makes easier to understand

Fig. 4. A Transportation Regulation Support System architecture. (For interpretation to colours in this figure, the reader is referred to the web version of thispaper.)


a complex reality by the reification of the components of the system to be managed. Underlining the components and theirlinks makes it easier to understand the regulation process which in turn facilitates its formalization.

A significant amount of research has been devoted to multi-agent approach in the transportation domain. Good reviewscan be found in (Davidsson et al., 2005; Schleiffer et al., 2002). In (Davidsson et al., 2005), the authors point out that 63% ofthe research is related to the design of a DSS. Operation research models and Interactive Decision Support Systems have beenwidely used in modeling decision-making processes but these systems are often represented as black boxes that hide thedecision-making process itself. Moreover these systems give synthetic results that have to be analyzed by the regulatorin order to give a final diagnosis. These models compute quantitative data, i.e. statistical indicators of the network operation– service journey run time, passenger demand – enabling transport supply planning but they do not take into account qual-itative data, such as the relative seriousness of a delay according to the position of a bus in the network. Most of all, thesemodels assume that the data is available and reliable but, as said above, this is a too strong hypothesis in urban transpor-tation network management.

Other research works on public transportation network management are based on a multi-agent simulator (Bukkapatnamet al., 2003; Ezzedine et al., 2008; Ezzedine et al., 2006; Jin et al., 2007; Meignan et al., 2007; Ossowski et al., 2005; Ossowskiet al., 2004). Ezzedine et al. (2008, 2006) present an architecture with the integration of the AVM, the DSS and the TravellerInformation System in a Traffic Management System (TMS). However the AVM and the DSS remain two separate sub-mod-ules of the TMS. Ossowski et al. (2005, 2004) introduce an organizational and communicative model of decision supportenvironments applied to transportation management. The drawback of most of these systems is that they are not integratedwithin a single system and they are not directly fed with real-time data coming from vehicle sensors. In that sense they arecloser to the architecture shown in Fig. 3.

3.3. Choice of an interaction model: the environment modeling approach

Interactions between agents through message exchange is an important part of a multi-agents system, and usually a dya-dic approach is proposed. In the domain of urban traffic control, however this approach is not feasible: the sender does notalways know the name of its receivers because the receiver of the message is often identified according to its position. Forexample, when a bus has to contact its closest bus, it does not know its identification. The simplest and more common pro-tocol is a (more or less limited) broadcast protocol. The drawback of this solution is a high communication cost, mainly inreal-time systems like urban transportation systems where the location of buses is updated very frequently (every 40 s inour application, see Section 4). Another traditional solution is the use of acquaintances; the interaction problem is solvedby an increase in the interaction knowledge of the agents. In a TRSS, this solution is inadequate because the problem remainswhen an agent is not able to link its needs to a given agent. A third solution is the use of a middle-agent. This approach, called‘‘capability-based coordination” is a preference/capability matching, used to identify the best provider for a given capability


search. Nor is this solution feasible for TRSS design because each type of agents has the same capabilities, for exampletimetable computation for STOP agents. Because the dyadic interaction solutions are not adapted to the transportation do-main, we propose to base our interaction model on a new principle: the mutual awareness principle. Mutual awarenesscomes from interactions in real-life situations that use other means of communication than direct transmissions (Dugdaleet al., 2000), and is related to a particular state of the participants called awareness.

Although it has long been considered as a passive state, we consider that awareness is an active state and not only theresult of stimuli. Work in the fields of psychology and sociology have discussed whether or not there also has to be an activeparticipation of the ‘‘receiver”. For example, Heath et al. (2002) says that awareness is not only the perceiver’s availability tobe aware of the environment, but also his ability to filter relevant information which is of particular significance. Mutualawareness is based on the sharing of interactions. To be efficient, this principle implies that agents share a common com-munication media. As a consequence, an agent only has to find messages that it is interested in. In the reactive agent com-munity, the environment is already used as a common interaction medium. In the cognitive agent community, we haveproposed the EASI model (Environment as Active Support of Interaction) (Saunier and Balbo, 2009), which enables cognitiveagents to use the environment to exchange messages. More precisely, EASI enables an agent to send messages to anotheragent that is located by the environment, and also enables agents to perceive every message exchanged.

For this purpose, we consider that the environment contains descriptions of messages and agents. The problem is how theagents use these descriptions to locate messages according to the environment state, which implies matching those descrip-tions and the needs of the agents. We therefore propose to represent the components of the environment (agents and mes-sages) as entities. Each entity has a description containing its properties, which are accessible through the environment.Agents have the ability to put filters in the environment and these filters are logical expressions on properties. When a messageis added to the environment, these filters determine by pattern-matching whether the agent is interested in it, in which case itwill receive it. In this way, the filters enable agents to create their communication space where each filter corresponds to a pre-cise communication need. More details on the EASI model can be found in (Balbo and Saunier, 2008; Saunier and Balbo, 2009).

The choice of the environment model as a first order entity is a challenging step in the design of a multi-agents system(Weyns et al., 2005). We made the hypothesis that this type of environment will allows to improve the efficiency of a TRSSand it has been proven true in our application (see Section 5). We thus consider that this type of environment is a usefulcomponent of our generic model of a TRSS. In the light of the research done on this topic (Weyns et al., 2005), we highlightin the remaining, four advantages of this type of environment:

– Due to its intermediary role, this type of environment is able to support spatially and temporally decoupled coordinationmodels and thus it simplifies the modelling phase taking into account dynamic real environments. This role was used inMoujahed et al. (2006) in order to model a facility location problem. It concerns facilities positioning such as train stations.By sharing the demand through the environment and computing what the authors call attractive and repulsive forces, theMAS environment becomes the common referential that enables agents to adapt their behavior according to the dynamicsof the real environment.

– Because shared resources and services are embedded by the environment, this type of environment is able to managetransportation environments which are usually very dynamic, ensuring the coherence of the MAS. In Moujahed et al.(2006), the environment ensures the propagation of the passengers demand. Moreover, the environment ensures serviceseither because theses services cannot be put at the agent level or because this simplifies the agent design. In a traffic lightcontrol system (Bazzan, 2005) based on evolutionary game theory, the environment – which has a global point of view –gives rewards or penalties to self-interested agents according to their local decision.

– Because this type of environment has its own dynamics and thus can control the shared space, it is able to trigger coor-dination rules in a multi-agent transportation system. In Meignan et al. (2006), a bus network simulation example, themain role of the environment is to constraint the perceptions and interactions of agents. In this example, a rule is ‘‘aBus agent and a Traveler agent can interact only when they are located at the same bus stop”. In transportation applica-tions which are characterized by incomplete knowledge of transportation processes, this type of environment modelingsimplifies the design of the MAS by a clear separation between the roles of the agents and their organization. In the coor-dinated monitoring of a traffic jam application (Haesevoets et al., 2007), the environment defines several agents organi-zations, that are dynamically modified by the environment according to the current context. Another characteristic oftransportation applications is the dynamicity. Since this type of environment allows to define laws, it is easy to explicitlyrepresent the way organizations and roles (played by agents) should evolve given the current context.

– Because the environment let its own structure observable and accessible, this type of environment let agents manipulate areification of the communication or coordination MAS components. This feature is important for transportation supportsystems to apply the mutual awareness principle that is based on an active state of the agents that have to filter relevantinformation like the messages or the state of the other agents.

4. A Transportation Regulation Support System prototype

Based on our generic model defined in Section 3, we have built an efficient Transportation Regulation Support Systems inwhich the information required to make informed decisions is available to the regulators and assessed in their context. We


have built on the architecture structure defined in Section 3.1, the multi-agents model defined in Section 3.2 and the inter-action model defined in Section 3.3 to design a TRSS prototype called SATIR. We have applied it to the diagnosis of the STIB(Brussel Intercity Transport Company) bus network. In the following, we present the multi-agent model, the interactionmodel and the system operation.

4.1. Multi-agent modeling of a transportation network

4.1.1. The agentsAccording to the generic model present in Section 3, we have proposed two categories of agents: (1) The STOP agents that

represent the theoretical structure of the network (organized in lines and routes, as defined in the data model in Section2.1.2). They also have topological information on the network: the distance to the next STOP agent (with the identificationof this agent), and more specific information like the physical possibility to do a U-turn. They encompass the knowledge ofthe graph makers (staff that compute the theoretical supply): a data table is associated with each STOP agent for each periodof time giving the quality of traffic (from 0 for low traffic to 2 for heavy traffic) and of passenger flow (also coded from 0 to 2– heavy flow) (Fig. 5, timetable data (a)). This knowledge is used in the dynamic timetable process, in the assessment processand in the search for solutions to a disturbance. (2) The BUS agents that represent the dynamic part of the network. Each BUSagent is the abstract model of an actual vehicle running on the transportation network and reports its movements to theSTOP agents. A BUS agent has a timetable (Fig. 5, timetable data (b)) where a course is modeled by a five-tuple :hDr; St;Dt; L;Riwith Dr the reference of the driver, St the reference of the STOP where the the course begins, Dt the departuretime of the course, L the line number and R the route number. Two more types of agents will be proposed for disturbanceprocessing (see Section 4.2).

4.1.2. A new environment model based on mutual awarenessAs we said in Section 3.3, we propose a new model of environment called EASI (Environment as Active Support of Inter-

action). The active environment is used to dynamically adjust the timetables when a bus, which has not been located at sev-eral stops, reappears. In this type of environment, the agents interact directly, sending and receiving messages throughlogical emission, reception and interception filters.

We have defined three types of filters (Balbo and Pinson, 2001):

– Reception filters: communication is based on a need that is common to two agents. The sender specifies the values of thecharacteristics searched for in the receiver. This description is matched against a communication filter of the category ofthe agents contacted. For example, if the agents have an id property, the filter that enables interaction based on the valueof this property is a reception filter.

– Emission filters: communication is based on a need of the sender which doesn’t match any expectation of the category ofcontacted agents. The consequence of this lack of interest among the receivers is the absence of suitable filterswhich would make it possible for the environment to find the agents concerned. Consequently, the sender must put

Fig. 5. A multi-agent model in a TRSS. (For interpretation to colours in this figure, the reader is referred to the web version of this paper.)


the appropriate filter in the environment at the same time as it sends its message. An emission filter may require the com-parison of the potential receivers and therefore requires the use of predicates. This implies that the filter matches all thepotential receivers of a message and then searches for the exact receiver of that message. For instance, an emission filtercan be used in order to find a particular BUS agent for a BUS agent b: the closest BUS to the sender (BUS agent b) terminus.The sender expresses its need: the receiver position must be less than the position value contained in the message. In thesubset of the potential agents, the exact receiver must have the greatest value of the property ‘‘position”.

– Interception filters: the interception filters allow the agents to receive the messages which are not sent to them but the con-tent of which may be of interest to them. The goal of these filters is to make full use of the environment as a shared workcontext where every communication is potentially available to all of the participants, this is a mutual awareness principle.An example of the use of an interception filter is given in Section 4.1.4 for the management of the false location of vehicles.

4.1.3. Organization modelingAs said earlier, organization modeling has to take into account the hierarchical organization of data and the processing of

this data. The hierarchical organization corresponds to the organization of the agents according to different abstraction levels(see Section 2.1.2). A STOP agent is located on a line and a route. The grouping of the agents according to lines or routes givesthe needed hierarchical level. The monitoring of the network under normal conditions is based on this organization. To man-age the network under disrupted conditions and in order to isolate the information related to a disturbance, we have definedanother organization to group the agents related to the disturbance. The agents are gathered in separate environmentsaccording to their functionalities: the STOP agent environment and the BUS agent environment. In each of these environ-ments, the communication needs are different and the set of communication filters is therefore adapted. Agent communi-cations are allowed inside and between environments. Fig. 5 shows the proposed multi-agent model as well as the firstand second level of the hierarchical data organization. The first level is related to the STOP agents and the second level tothe routes (Fig. 5, topological relation).

4.1.4. The system operationAt the beginning of the process, the theoretical pre-recorded data is distributed between the STOP agents. A data table is

used to define the theoretical state of traffic and of passenger demand and to compute timetables. For instance, if the stop P1is a school, the estimated value of traffic is 1 (the circulation is normal) and the estimated value of passenger flow is 2 at8pm. The timetable-computing protocol is initiated by a BUS agent 10 min before a vehicle departs. The BUS agent send thisevent (departure time) to the STOP agent where the route begins. This STOP agent records in its list of expected vehicles thisBUS agent with its arrival time. The STOP agent adds to this time the time needed to take passengers and to go to the nextSTOP agent taking into account the time of the day and forwards the result to the next STOP agent on the same route. Thiscomputation process goes on until the last stop of the route. Dynamic timetable computing replaces theoretical timetableswhich are no longer a valid reference when the network is disturbed (see Section 2.2.2).

The basic event in the network is reproduced in our MAS by an interaction: when a vehicle passes a stop on the real net-work, a warning message is sent from the BUS agent to the corresponding STOP agent. The STOP agent updates its timetableby removing this vehicle from the list of vehicles due. A STOP agent that does not receive any message detects an anomalyand triggers the disturbance processing presented in Section 4.2. In order to avoid false alarms and to process only seriousdisturbances, the triggering of the disturbance management process depends on the value of a time parameter that may bedifferent for each STOP agent. A default value of 7 min has been used (STIB norm). One of the difficulties of timetable man-agement concerns the management of inconsistencies which arise from the data sent by sensors located in built-up areas.Some vehicles may not be located at a significant number of stops and this may result in the triggering of false alarms. Theincorrect location of a vehicle may lead to inconsistent situations with ‘‘virtual overtakings” (a vehicle is announced beforethe vehicle which precedes it). For the multi-agent network this means that the STOP agents on the bus route have not beeninformed about the passage of this bus, but thanks to an interception filter, they will intercept all new transit announce-ments sent by vehicles not running to timetable. The interceptor agents receive the message and update their timetables.

The mutual awareness principle provides an efficient solution to the problem of the inconsistent data, in that it needs fewinformation: the BUS agents have no information on the topology of the routes; only one warning message is sent for a tran-sit event in the network and the MAS state is updated according to the reaction of agents to this event in their own context.The BUS agents are the intermediaries between theoretical data (STOP agents) and real-time data. The design of two envi-ronments ensure the modularity of our proposal: the BUS environment is the only interface between the multi-agent modeland the data of the transportation network (Fig. 5). Moreover, this separation between the roles of the STOP agents and BUSagents makes easier the development of more evolved functionalities such as diagnosis or planning described in the follow-ing sections.

4.2. Multi-agent modeling of the diagnostic process

Describing a disturbance using the delay of a vehicle is not sufficient. For example, a vehicle may be running late, but thedistance between the previous and the following vehicles is preserved. In this case, a human regulator will not take the dis-turbance into account and will be more interested in a vehicle with a shorter delay, but which leads to an imbalance along


the line. To measure qualitatively the seriousness of a delay, we have taken into account its context and its consequences onthe activity of the network. For this purpose, we have proposed to put together within a specific organization, called the Dis-turbance model (Balbo and Pinson, 2005), all STOP agents and BUS agents related to a given disturbance, according to therepercussions on the network. Unlike other transportation information systems which propose an instantaneous image ofthe network state that regulators have permanently to analyze, we propose the integration of information updates in theDisturbance model, thus allowing a disturbance to be analyzed from beginning to end helping the regulator to do his diag-nostic job.

For this purpose, we have defined three information sets, also called areas:

– The Successor area: this area brings together all the stops waiting for the successor of the late bus; it measures the riskassessment of a bus train (the late vehicle is caught up by the following one).

– The Critical area: this area brings together all the stops where the vehicle is late; it measures the risk assessment of a gap(the late vehicle is left behind by the preceding bus).

– The Predecessor area: this area brings together all the stops where the late vehicle is due but not yet late; it measures therisk assessment of a gap.

By drawing a distinction between the Successor and Critical areas, it is possible to compare disturbances in terms ofseverity. For two disturbances with the same number of stop agents between the late bus and its predecessor, the distur-bance with the greatest number of stops in the Critical area is considered being the most severe. The set of these three areasconstitutes the Disturbance model. Fig. 6 shows three disturbances with their respective areas (Fig. 6A). For each distur-bance, a specific environment, called INCIDENT environment, is created where the organization of the agents is hierarchicalas explained below (Fig. 6B).

To measure qualitatively the seriousness of a delay, we have taken into account its consequences on the activity of thenetwork. We have defined two measures of risk linked to a disturbance, the risk of a bus gap indicator and the risk of abus train indicator. These indicators, detailed in Balbo and Pinson (2005), are based on the study of a priori progression dif-ficulties of vehicles involved in the disturbance and take into account the intrinsic dynamics of a disturbance.

The initial organization of the multi-agent system (in lines and routes) does not enable the disturbance process to bemodeled over time. It must be completed by a hierarchical organization of the agents linked to a disturbance (Fig. 6B). Inorder to aggregate information at each level of the hierarchy and to compute feasible solutions, two new types of agents,STOPAREA agents and the INCIDENT agent are defined. The lowest level of the hierarchy is made up of the elementary enti-ties, the STOP agents. The middle level is composed of the STOPAREA agents that make an initial synthesis of the information.They collect basic information such as theoretical traffic assessment and passenger flow from the STOP agents linked to themand they compute the progression coefficient (an indicator of the seriousness of the disturbance in the area). The INCIDENTagent represents the top of the hierarchy where the risks are computed (Balbo and Pinson, 2005). It displays the aggregatedinformation on the screen and is the interface between regulators and the system.

Fig. 6. Multi-agent dynamic disturbance modeling in a TRSS. (For interpretation to colours in this figure, the reader is referred to the web version of thispaper.)


This organization is dynamic because at each cycle, STOP agents move from one area to the other within the hierarchy,and from and towards the outside of the organization, according to traffic direction (Fig. 6B). When the disturbance disap-pears, this organization stays during some cycles to keep the continuity of the disturbance process. After 5 min, the createdagents related to the disturbance disappear too.

The originality of our approach is the dynamic modeling of a disturbance process from beginning to end and its integra-tion in a multi-agent system. We have defined a model, called the Disturbance model that allows information synthesiswhich is useful for decision-making. Through this model, knowledge relative to the network structure and knowledge rel-ative to the network dynamics (stored in STOP agents and in BUS agents, respectively) are gathered within a single entitycalled ‘‘Disturbance” that is reified in a specific environment (one environment per disturbance) where all the agents in-volved are brought together. This entity allows the follow-up of the disturbance over space and time; it is deleted whenthe disturbance problem is solved.

4.3. Multi-agent modeling of the planning process

When a disturbance has been detected and assessed, the TRSS computes the feasible regulation procedures. Initially, thetransport service matches the theoretical demand to the bus supply. However, when a disturbance appears, there is a dis-crepancy between the service provided and the passenger flow. Thus, the task of the system is to automatically adjustthe initial supply in order to satisfy the needs according to the changes in the network.

4.3.1. Static feasible action planningThanks to predefined procedures, regulators modify the transport service according to the state of the network and to

the possible actions of the buses on the line. They cancel or modify the timetable of the vehicle that is chosen to executethe regulation procedure in order to shift the service to another point on the network. One of the original features of theTRSS is that BUS agents play the role of regulators, enabling a micro-regulation of the network. At the beginning of itsactivity, each BUS agent receives the list of its runs (its timetable) that it may modify dynamically, thus acting as a reg-ulator. When a disturbance is detected, the late BUS agent requests a new run that each BUS agent on the same line triesto insert into its timetable. For each BUS agent, this insertion implies a modification of its timetable. One or more regu-lation procedures can be used, and timetable processing within the multi-agent system is implemented in three steps(Fig. 7):

– Step 1: the availability of the BUS agents is computed in order to eliminate vehicles that are not potential solutions. Usingconditions related to its own characteristics (for example, its size) or to network rules (for example, the last run of a vehi-cle is never changed), a vehicle is eliminated.

– Step 2: the profiles of the available BUS agents are computed. By profile, we mean the characteristics of a group of BUSagents with the same relative position compared to the late BUS agent (i.e. before, after, same direction, etc.). For eachprofile, the regulation procedures are the same. For example, the AlightingOnly procedure may be feasible for all BUSagents located before the late BUS agent but it is useless for the BUS agents located after it. Using the BUS profile maylimit the number of tests: some procedures may be forbidden or limited to specific profiles.

Fig. 7. The planning process. (For interpretation to colours in this figure, the reader is referred to the web version of this paper.)


– Step 3: the feasibility of the regulation procedures is computed. Every regulation procedure has constraints that BUSagents must satisfy in order to be considered as feasible. For example, a vehicle cannot make a U-turn if there is nowhereto do so.

Breaking this timetable processing down into three steps offers several advantages. Since network authorities have theirown regulation rules, they do not apply the same constraints on vehicles and on regulation procedures. The three steps de-scribed above enable a network to adjust the planning process to its own rules and constraints. For example, a networkauthority can decide that the buses belonging to a specific profile are not available. Thanks to the prior breaking down ofthe planning process, this profile is not taken into account (step 2) and the process of the other profiles is not modified.Moreover, the distribution of BUS agents into profiles limits the solution space to just the feasible procedures. From a mi-cro-regulation viewpoint, another advantage of this planning process is that it can be distributed and automatically appliedby BUS agents.

4.3.2. Dynamic feasible action planningIn this section, we propose a general model of the automatic regulation procedure based on the planning process de-

scribed above. The adjustment process begins with an inform message sent by an INCIDENT agent to the late BUS agent. Be-cause computation of the regulation procedures depends on its position and on the information related to its current run, thelate BUS agent looks for the missing information on the network. Firstly, it sends a request message to the STOP agents lo-cated between its own position and the end of its current run to collect the missing data, such as passenger flow and runlength; secondly it forwards this to the BUS agents that are on the same line. The next step is done by the BUS agents thatexecute the steps of the planning process described above. Thanks to its local knowledge of the network and of its own activ-ity, a BUS agent b may propose feasible regulation procedures.

We propose a general model of the regulation procedure as follows (Fig. 8).For each regulation procedure we define three preconditions and a computation process. The preconditions are related to

the steps of the planning process and the computation process computes the insertion of a new run requested by the lateBUS agent into the BUS agent b timetable.

– Let H be the set of hard preconditions related to the characteristics of the BUS agent b (see step 1). Each condition takesinto account the internal state of the BUS agent b and does not require any additional information to be assessed.

– Let P be the conditions related to the profile of the BUS agent b (see step 2). The BUS agent b determines its own profilegiven the current disturbance and computes the regulation procedure that is linked to its profile.

– Let S be the soft conditions that are related to the availability of the procedure (see step 3). If we take the example of theAlightingOnly procedure, it is limited to the profiles of BUS agents that are located at the beginning of the line and beforethe late vehicle. As an example, the soft preconditions, defined by the STIB Belgium bus network, are the following (H = ;which means that AlightingOnly is always possible):� The distance (number of stops) between the BUS agent b and the late BUS agent must be less than 5. This procedure

means that the passengers of the late vehicle have to wait for the following vehicle and that the waiting time must notbe too long.

Fig. 8. The regulation algorithm.


� The position of the terminus of the BUS agent b involved must be superior to the position of the late vehicle terminus. Ifthis is not the case, the procedure will be done only on part of the run of the late vehicle.

� The distance between the late vehicle and its predecessor must be greater than 10 stops. Thanks to this procedure thelate vehicle will make up for lost time. The aim of this precondition is to avoid the creation of a bus train with the pre-vious bus.

In order to compute these conditions, the BUS agent b has to look for the missing information. The number of passengersin the vehicles is not taken into account because the STIB network does not have sensors to give this information. For theAlightingOnly procedure, the BUS agent has to know the position of the predecessor of the late vehicle. The name of this itemof data and that requested for the computation process are recorded in the data set called D (Fig. 8).

Let R1 be the run that will be modified to insert the requested run. R1 is the current run or a run that is chosen accordingto its departure time. As an example, for the AlightingOnly procedure, the chosen run is the current run of the late BUS agentand adjustment consist in computing a new timetable. To do so, the late BUS agent looks for the following data: number ofstops, traffic and passenger flow values between the vehicle and the late one.

In Fig. 8, the algorithm gives the steps to compute the new run of a BUS agent. Only the procedures that belong to theprofile of the BUS agent are taken into account. If the hard conditions are false (hardCondition, line 5), the BUS agent looksfor the missing information (search data, line 6). If the soft conditions are true, then the function procedureValidation returnsthe result of the computation. Using the data on the chosen run R1, the requested run R2 and the collected data Dresult , thisfunction computes the run-time of the regulation procedure.

To close this process, each BUS agent sends the result of its computation to the INCIDENT agent that collects and organizesthis information for the regulator. The result of this process is presented in the next section.

5. Experimentation and results

A prototype has been implemented in C++. In order to study the feasibility of our SATIR system, the prototype was testedusing real data recorded every 40 s from buses on the Brussels Intercity Transport Company network (STIB). In this sectionwe only provide results for the planning process. Results for the assessment process were given in (Balbo and Pinson, 2005).During the test phase, the data was recorded on tape for about 30 buses, on one line, over 8 days for the line number 20 andrepresents more than 43,000 items. This line was chosen because it was a problematic line: it was so disturbed that the timebetween two departures of the same bus was increased to absorb the delays. The SATIR system was run over a period of timethrough cycles on this data representing the move of buses on the network; as expected, it detected and assessed 300 dis-turbances proving the feasibility of our system.

To validate our prototype, an experience has been done with a new line, the line number 54. During this experience, linenumber 54 was studied as follows: (1) Day 1: the monitoring data of the regulator that manages line 54 was recorded. All thedisturbances managed and the solutions chosen were recorded. (2) Day 2: SATIR was tested with the data related to the lineactivity of day 1. A serious disturbance managed by both the regulator and SATIR was identified. (3) Day 3: the solution givenby SATIR was shown to the regulators and the solutions to the same disturbance were compared. This disturbance is relatedto the buses #54806, #54827, #54830 (Fig. 9) that belong to line number 54 and involves a substitute vehicle of line number80. The system was run without the multi-agent component of the TRSS and with this component. The computer run-time is

Fig. 9. Disturbance evolution from 15:38 to 15:56:14.

Fig. 10. Two SATIR proposals with vehicle #54806 to make up a delay.


5 min 36 s (for 5358 network events (bus location, departure, etc.)) without the multi-agent component. This run-time cor-responds to the AVM collecting and displaying processes. The run-time is 6 min 30 s with the multi-agent component. Thisrun-time corresponds to the collecting and displaying processes and the following processes: (1) monitoring (dynamic time-table computation); (2) diagnostic (disturbance assessment); (3) planning (feasible regulation procedure computation). Onthis day, 48 disturbances were detected, assessed and recorded.

The sequence of events from the regulator point of view is the following: the disturbance was caused by a badly parkedvehicle which blocked bus #54806 that therefore ran more and more behind timetable. The regulator was informed by thedriver at 15:33. The regulator chose to call the vehicles near this disturbance in order to organize a diversion, but there was atechnical problem and the driver was unable to get the call. The consequence was a bus train at 15:56:14 (Fig. 9B). This dis-turbance was close to line number 80. The regulator used this proximity to choose a vehicle that was at the end of its runs tosubstitute for the blocked bus. Since this new vehicle did not exist in the AVM system, and a fictitious reference was createdfor it (called vehicle #1). Since vehicle #54806 had a substitute (Fig. 9B) it made a U-turn to make up for lost time. The reg-ulator planned the U-turn where the substitute bus was inserted. When the blocking car had gone, vehicle #54806 continuedits runs to the stop where vehicle #1 was waiting. At this location, vehicle #54806 made a U-turn and the vehicle #1 con-tinued the run of vehicle #54806.

The sequence of events from the SATIR point of view is the following: since the disturbance was detected by the regulatorafter a call, SATIR did not have this information. As a consequence the disturbance was detected by SATIR after 7 min (toavoid a false alarm), at 15:38 (Fig. 9B). This disturbance is also an example of the quality of the data. The green color ofthe vehicle #54806 given by the STIB AVM means that there is no bus delay (Fig. 9A). However SATIR has detected a distur-bance. The vehicle #54806 was at the stop called ARLON and its STOP agent triggered the disturbance assessment process(Fig. 10). Every 3 min, the new INCIDENT agent updates the disturbance assessment. Vehicle #54806 has two proposals thattake into account the next vehicle (#54830). The common objective is to increase the speed of vehicle #54806 by modifyingthe run mode. The first regulation feasible procedure (P, Fig. 8, line (2)) is an empty run (P = Empty run, Fig. 10A); the secondchoice is a run with alighting only (P = AlightingOnly, Fig. 10B). These procedures take into account the next vehicle becausethey are feasible if this vehicle is close enough to pick up the passengers that are not picked up by vehicle #54806. All thisinformation is displayed to the regulators through the interface (Fig. 10A and B). For each procedure, several items of data aredisplayed: the current delay (7 min 4 s), that part of the delay that can be made up, how many stops are necessary to make itup. For instance, in the first procedure (P = Empty run), 15 stops are needed and in the second one (P = AlightingOnly), 23stops are needed.

Each regulator proposes solutions according to his own knowledge, habits and experience. For the situation described inthis paper, the regulator may propose several solutions that are different to those proposed by SATIR. In our example, theregulator has chosen external resources although there are internal solutions that have been proposed by SATIR. In orderto validate our model, the SATIR proposals were shown to the regulators, who approved them as being feasible.

6. Conclusions

In this paper, we have presented an agent-based approach to design a Transportation Regulation Support System (TRSS).Based on a multi-agent model of an urban transportation network, our objective is to combine the functionalities of an exist-ing information system with the functionalities of a Decision Support System. First, we highlighted the main regulation pro-cess problems before going on to propose new functionalities for a TRSS in order to deal with these problems. More precisely,we pointed out the three choices that designers have to make when designing a TRSS: (1) the choice of a TRSS architecturethrough the integration of a DSS and an AVM system, (2) the choice of a multi-agent modeling approach which facilitates therepresentation of distributed entities, their interaction and their processes, (3) the choice of an environment model as a firstorder entity which allows the agents to interact directly, sending and receiving messages through logical emission, reception


and interception filters. Based on these three choices, we have proposed a generic model. Unlike the other approaches thatonly deal with a specific task, the original feature of our generic model is that it proposes a global approach to the regulationfunction. The global approach models the three regulation tasks of human operators in the same system: network monitor-ing through information synthesis, disturbance diagnosis and action planning.

To illustrate our approach, we have presented a TRSS prototype called SATIR (Automatic System for Network Incident Pro-cessing) that we have developed for the STIB bus network. We have shown that, based on our functionalities, the TRSS is ableto monitor the network activity under normal conditions by detecting incoherent data, and under disturbed conditions bydetecting traffic disturbances. Then it is able to automatically adjust to the environment changes by proposing feasible solu-tions in order to optimize the traffic flow.

Since the multi-agent model is integrated in the decision support system, it detects disturbances, manages them andlooks for a solution to reduce the problems resulting from a vehicle being late. For this purpose, we have defined a new mod-el, called the Disturbance model, that allows dynamic information synthesis useful for decision-making.

Using real data coming from the Brussels bus network, our SATIR system has been assessed over several test days in real sizeas the basis of a Decision Support System. The system has been shown to work despite sensor failures. It has detected and as-sessed more disturbances than regulators could have done. It have then proposed solutions that were validated by regulators.

Several key contributions should be mentioned. First, the multi-agent framework provides a good structure to develop adistributed DSS for urban public transportation regulation system management. Second, the original concept of active envi-ronment that we have introduced allows the agents to interact directly, sending and receiving messages through logicalemission, reception and interception filters. The active environment is used to adjust the model when the bus, which hadnot been located at several stops, reappears. This is done by interception logical filters which enable the passed stops to up-date their next bus waiting time. Third, the multi-agent approach allows the dynamic management of the bus timetable, thebus monitoring in real-time. Detection and display of the disturbances enable the operators to quickly identify the malfunc-tioning of the lines of the bus network.

More research has to be done in various directions: (1) full validation and test of the system; (2) the development of userdialog and interfaces to allow human regulators to ask for explanations of the disturbances; (3) the addition of statisticaltools to better understand the network operations; (4) the taking into account of several lines at the same time and inter-change problems. More generally, the use of the multi-agent paradigm for network management opens new perspectives.Since our system reproduces the network activity, we can run it:

(i) To simulate new timetables. When disturbances are detected, the regulators and graph makers solve the problems bychanging the timetables; the simulator could validate the changes by reproducing the network activity using data con-cerning recurrent difficulties. One way to measure the improvement of the service would be to look at the number ofdisturbances.

(ii) To serve as a training tool. The behavior of the vehicles could be simulated in order to create disturbances while thesystem could assess the behavior of the regulator according to the number of disturbances on its test network.

References

Balbo, Flavien, Pinson, Suzanne, 2001. Toward a multi-agent modelling approach for urban public transportation systems. In: Omicini, A., Petta, P., Tolksdorf,R. (Eds.), Engineering Societies in the Agent World II, LNAI, vol. 2203. Springer Verlag, pp. 160–174.

Balbo, Flavien, Pinson, Suzanne, 2005. Dynamic modeling of a disturbance in a multi-agent system for traffic regulation. International Journal of DecisionSupport System 41 (1), 131–146.

Balbo, Flavien, Saunier, Julien, 2008. On the use of symbolic data analysis to model communication environments. In: Klusch, Matthias, Pechoucek, Michal,Polleres, Axel (Eds.), CIA of Lecture Notes in Computer Science, vol. 5180. Springer, pp. 234–248.

Bazzan, Ana L.C., 2005. A distributed approach for coordination of traffic signal agents. Autonomous Agents and Multiagent Systems 10 (1), 131–164.Brezillon, P., Pasquier, L., Saker, I., 1999. Context-based decision making in incident management on a subway line. In: Lenca, P. (Ed.), Proceedings of the

Human Centered Processes Conference, Brest, France, September, pp. 129–134.Crevits, Igor, Debernard, Serge, Denecker, Pascal, 2002. Model building for air-traffic controllers’ workload regulation. European Journal of Operational

Research 136 (2), 324–332.Davidsson, Paul, Henesey, Larry, Ramstedt, Linda, 2005. An analysis of agent-based approaches to transport logistics. Transportation Research Part C:

Emerging Technologies 13 (4), 255–271.Dugdale, J., Pavard, J., Soubie, B., 2000. A pragmatic development of a computer simulation of an emergency call center. Designing Cooperative Systems: The

Use of Theories and Models, 241–256.Ezzedine, Houcine, Trabelsi, Abdelwaheb, Kolski, Christophe, 2006. Modelling of an interactive system with an agent-based architecture using petri nets,

application of the method to the supervision of a transport system. Mathematics and Computers in Simulation 70 (5–6), 358–376.Ezzedine, Houcine, Bonte, Therese, Kolski, Christophe, Tahon, Christian, 2008. Integration of traffic management and traveller information systems: basic

principles and case study in intermodal transport system management. International Journal of Computers, Communications & Control 3 (3), 281–294.Haesevoets, Robrecht, Van Eylen, Bar, Weyns, Danny, Helleboogh, Alexander, Holvoet, Tom, Joosen, Wouter, 2007. Context-driven dynamic organizations

applied to coordinated monitoring of traffic jams. In: Weyns, Danny, Brueckner, Sven, Demazeau, Yves (Eds.), Proceedings of the Workshop EngineeringEnvironment-Mediated Multiagent Systems, pp. 126–143.

Heath, Christian, Svensson, Marcus Sanchez, Hindmarsh, Jon, Luff, Paul, Lehn, Dirk vom, 2002. Configuring awareness. Computer Supported CooperativeWork 11 (3), 317–347.

Jin, Xu, Itmi, Mhamed, Abdulrab, Habib, 2007. A cooperative multi-agent system simulation model for urban traffic intelligent control. In: SCSC: Proceedingsof the 2007 Summer Computer Simulation Conference. Society for Computer Simulation International, San Diego, CA, USA, pp. 953–958.

Meignan, David, Simonin, Olivier, Koukam, Abderrafiaa, 2006. Adaptive traffic control with reinforcement learning. In: 4th Workshop on Agents in Trafficand Transportation (ATT’06), pp. 50–56.


Meignan, David, Simonin, Olivier, Koukam, Abderrafiâa, 2007. Simulation and evaluation of urban bus-networks using a multiagent approach. SimulationModelling Practice and Theory 15 (6), 659–671.

Moujahed, S., Simonin, O., Koukam, A., Ghedira, K., 2006. A reactive agent based approach to facility location: application to transport. In: 4th Workshop onAgents in Traffic and Transportation (ATT’06), pp. 63–69.

Ossowski, Sascha, Hernandez, Josefa Z., Belmonte, Maria-Victoria, Maseda, José-Manuel, Fernández, Alberto, Garcá-Serrano, Ana, 2004. Multi-agent systemsfor decision support: a case study in the transportation management domain. Applied Artificial Intelligence 18 (9-10), 779–795.

Ossowski, Sascha, Hernández, Josefa Z., Belmonte, María-Victoria, Fernández, Alberto, Garcá-Serrano, Ana, Cruz, José-Luís Péres de-la, Serrano, Huan-Manuel, Triguero, Francisco, 2005. Decision support for traffic management based on organisational and communicative multiagent abstractions.Transportation Research Part C: Emerging Technologies 13 (4), 272–298.

Rasmussen, Jens, Brehmer, Berndt, Leplat, Jacques, 1990. Distributed Decision Making: Cognitive Models for Cooperative Work. John Wiley & Sons Ltd.Rouse, William B., 1981. Human–computer interaction in the control of dynamic systems. ACM Computing Surveys 13 (1), 71–99.Saunier, Julien, Balbo, Flavien, 2009. Regulated multi-party communications and context awareness through the environment. International Journal on

Multi-Agent and Grid Systems 5 (1), 75–91.Schleiffer, R., 2002. Intelligent agents in traffic and transportation. Transportation Research Part C: Emerging Technologies 10C (5–6), 325–527.Vahidov, R., Fazlollahi, B., 2004. Pluralistic multi-agent decision support system: a framework and an empirical test. Information and Management 41 (7),

883–898.Weyns, Danny, Parunak, H. Van Dyke, Michel, Fabien, Holvoet, Tom, Ferber, Jacques, 2005. Environments for multiagent systems state-of-the-art and

research challenges. Lecture Notes in Computer Science Series 3374, 2–52.Zhao, J., Bukkapatnam, S., Dessouky, M., 2003. Distributed Architecture for Real-time Coordination in Transit Networks. Technical Report Metrans Project

00-13, University of Southern California.

using intelligent agents for transportation regulation support system design

Documents