A MULTI-AGENT DEMAND MODEL - UFRGS

A MULTI-AGENT DEMAND MODEL
Rosaldo J. F. Rossetti 1 , Ronghui Liu 2 , Helena B. B. Cybis 3 , and Sergio Bampi 1
1 Informatics Institute, PPGC – UFRGS, Brazil
{rossetti, bampi}@inf.ufrgs.br
2 Institute for Transport Studies – Univ. of Leeds, UK
rliu@its.leeds.ac.uk
3 School of Engineering, LASTRAN – UFRGS, Brazil
helenabc@vortex.ufrgs.br
1 INTRODUCTION
The rapid growth of urban areas has deserved special attention from the technical and
scientific community over the last decades. Traffic and transportation systems have been a
subject of special concern as they play an important and indispensable role in today’s
society. Heterogeneity and uncertainty have become even more evident, and researchers
have been challenged by how to cope with modelling and simulating new measures brought
about by the adoption of new information technologies. According to Schleiffer (2000),
modelling heterogeneity in a microscopic level is a key step toward the understanding of a
macroscopic behaviour, and the use of artificial agents to represent individual mechanisms
is the tool to better comprehend highly complex and dynamic collective behaviour of traffic.
In this work, we use agent-based techniques to represent traffic demand as a community of
autonomous driver agents. Our approach relies on an extension to the DRACULA
microsimulation suite, as proposed in (Rossetti et al., 2000). Drivers are modelled by way of
the BDI architecture (Belief, Desire, and Intention) and our focus is given to route and
departure time choices.
2 MULTI-AGENT SYSTEMS AND THE BDI APPROACH
In general terms, Artificial Intelligence (AI) is especially concerned with the development
of computational models that mimic human intelligent and rational behaviour, and
encompasses a wide range of multidisciplinary knowledge areas. Multi-agent systems
(MAS) is a sub-field of the distributed AI whose abstraction approach basically consists of
representing the application domain by means of multiple entities, coined agents. An agent
is able to perceive facts through sensors and to act upon the environment by way of its
effectors (Russell and Norvig, 1995). Autonomy, social ability, reactivity, adaptability, and
pro-activity are some features that characterise intelligent agents, and the overall system
performance is expected to emerge from the particular behaviour of each individual.

In this work, we approach the task of modelling demand and representing driver behaviour
by the adoption of agent-based techniques. Each driver is modelled as a rational agent
featuring a BDI reasoning model, which was conceived on the basis of the framework
proposed by Rao and Georgeff (1991). The underlining idea of their theory is the
understanding of practical reasoning as the process of deciding, moment by moment, which
action to perform in the furtherance of the individual’s goals (Wooldridge, 1999). As briefly
described by Wooldridge (1999), a BDI agent is composed of base beliefs that represent the
information about its environment (beliefs are updated through perceptions). On the basis of
current beliefs and current intentions, an option generation function evaluates the options
available to the agent (its desires). Current options represent possible courses of action, and
a deliberation process determines the agent’s new intentions on the basis of its current
beliefs, desires, and intentions (states of affair that it has committed to trying to bring
about).
In order to turn theory into practice, Rao (1996) specified the AgentSpeak(L) language to
support the construction of BDI agents. Albeit it has not, as yet, been actually implemented,
Machado and Bordini (2001) suggest AgentSpeak(L) also serves as a specification language
for representing BDI agent models, which we have experienced in previous work (Rossetti
et al., 2001). In practice, modelling a BDI agent in AgentSpeak(L) is accomplished simply
by means of identifying its base beliefs and a set of non-instantiated plans. Intentions are
dynamically generated during the reasoning process. Readers are referred to (Rao, 1996) for
a detailed explanation and formal specification of the BDI logics and AgentSpeak(L).
We have used JAM (Huber, 1999) to implement our demand model, which was specified in
AgentSpeak(L). JAM is a Java-based architecture that combines the aspects of several
leading-edge agent theories and intelligent agent frameworks, including the BDI logics. We
rely on the operational semantics and on the constructors of JAM to implement and test our
multi-agent demand approach. Another way for practical implementation of AgentSpeak(L)
models is presented in (Machado and Bordini, 2001).
3 MADAM: THE DEMAND MODEL
MADAM (Multi-Agent DemAnd Model) is an agent-based model aimed at representing
variability in traffic demand by means of a population of BDI driver agents. Our approach
relies on an extension to DRACULA (Dynamic Route Assignment Combining User
Learning and microsimulAtion), which is a microscopic traffic simulation suite that has
been developed in the Institute for Transport Studies, University of Leeds, UK (Liu et al.,
1995).
A key ingredient in the DRACULA model is variability, which rises from two centrally
important concepts, namely the day-to-day dynamics and the within-day decision making
process. The former is concerned with modelling how the state of the network changes over

time, as the spatial knowledge of a driver is constantly evolving in response to trips made
through the network. The latter deals with choices regarding the journey, made by each
individual on a daily basis.
As in conventional models, the approach used in DRACULA begins with the concept of
demand and supply (Liu et al., 1995). In the demand model, trip makers are represented in
terms of their individual choices, whereas the microscopic simulation of vehicles moving
through the network is represented in the supply side. MADAM replaces the demand model;
rather than representing travel choices through variable values in a simple data structure,
demand results from the cognition process of each single BDI driver agent.
We focus on the departure time and route choices, following the decision-making
approaches currently implemented in DRACULA, as presented in (Liu et al., 1995).
Departure time is chosen in response to traveller’s previous experiences and preferred
arrival time. The absolute delay for a driver m travelling from origin i to destination j on day
k is given in Equation 1, where d ijm (k) is the departure time, t ijm (k) is the travel time, and a
ijm (k) is the desired arrival time. Drivers are also assumed to be indifferent to a delay of e mt
ijm (k) (relative to the travel time experienced). Equation 2 represents the lateness perceived by
individuals.
δ ijm (k) = d ijm (k) + t ijm (k) – a ijm (k) (1)
∆ (k) ijm = δ (k) ijm – e mt ijm (k) (2)
As suggested in (Mahmassani et al., 1991), drivers are likely to be indifferent to early
arrivals. Bearing this in mind, we consider that users only adjust their departure time for a
future journey in the case of ∆ (k) ijm > 0, otherwise they will keep the same departure time.
The adjustment is made as in Equation 3.
d ijm (k+1) = d ijm (k) (k)
– ∆ ijm (3)
The route choice also follows the model implemented in DRACULA (Liu et al., 1995),
which is the “boundly rational choice” based on the work presented by Mahmassani and
Jayakrishnan (1991). Drivers are assumed to use their habit routes as on the last day, unless
the cost expected for the minimum cost route is significantly better. Thus, a driver will use
the same route unless C p1 – C p2 > max(η × C p1 , τ), where C p1 and C p2 are the costs along
the habit and the minimum cost routes, respectively, and η and τ are global parameters
representing the relative and the absolute cost improvement required for a route switch. This
model is the basis for the BDI driver agent, which is described by means of a set of base
beliefs and a set of plans.
4 A SIMULATION FRAMEWORK
MADAM is implemented in Java to support the JAM-based driver agents, and replaces the
original demand sub-model of DRACULA, as shown in Figure 1. On the demand side, day-

to-day variability is represented in terms of individuals that have decided to travel on a
certain day. Decision-making is driven by mental attitudes of each of our BDI agents. MA
Initialisation synthesises a population of agent drivers from an OD matrix and sets the
initial base beliefs for each individual including possible routes for each origin-destination
pair (possible routes are identified at the beginning of the simulation). The demand on each
day is built up from the population and is formed by drivers who have decided to travel on
that day. The Input MA file gathers drivers’ route and departure time choices, so that they
can be launched onto the network to perform their journeys. The Output MA file returns the
travel costs (given in terms of the travel time experienced) for each driver in the demand
yielding an update of the base beliefs. In the supply module, dracprep is responsible for
setting supply day variability whereas dracsim simulates the microscopic movement.
yes: stop simulation
Routes OD Matrix
MA Initialisation
population
of
BDI drivers
demand
End?
Input MA
Output MA
no
DRACPREP
DRACSIM
MADAM
(demand)
DRACULA
(supply)
Figure 1. The simulation framework
We have carried out a simple simulation in order to test our approach. The network used as
the testing environment consists of 11 origin-destination pairs, 12 priority and two
signalised junctions. A total of 2323 driver agents were generated and simulated over a
period of 50 days (30 min spreading period on each day). Departure time, arrival time,
desired arrival time, and travel time for a driver of the population are plotted on the graph of
Figure 2, illustrating the approach adopted for the departure time choice.

A BDI Driver Agent
40
time (min)
30
20
10
Departure Time
Arrival Time
Desired Arrival
Travel Time
0
0 5 10 15 20 25 30 35 40 45 50
days
Figure 2. Departure time decisions of a BDI driver agent over a period of 50 days
5 CONCLUSIONS
We have devised an agent model on the basis of the driver behaviour currently implemented
in DRACULA. In such an approach, demand is built up by means of BDI drivers that reason
about departure time and route options, as well as on their needs for travelling on a certain
day. The use of mental attitudes, such as beliefs, desires, and intentions allowed for an
appropriate representation of drivers’ cognition on trip parameters. The MADAM structure
is also flexible to support different behaviour models by way of simply specifying new sets
of beliefs and plans. The very next steps in this project are the validation and the
improvement of our BDI driver model and the simulation of more realistic scenarios with
the analysis of aggregate parameters. Further developments also include the assessment of
drivers’ response to exogenous information, such as those provided by Advanced Traveller
Information Systems (ATIS) and Variable Message Signs, as suggested in (Rossetti et al.,
2001). We also intend to formalise and develop a more elaborated learning model as a way
of improving decision-making.
ACKNOWLEDGEMENTS
We gratefully thank Dr. Dirck Van Vliet and Dr. Rafael H. Bordini for their invaluable
comments and suggestions. This project has been financially supported by CNPq and
CAPES, the Brazilian agencies for R&D.

REFERENCES
Huber, M.J. (1999) JAM: a BDI-theoretic mobile agent architecture. Proc. of the 3rd International Conference
on Autonomous Agents, Agents’99. Seattle, May 1999. pp.236-243.
Liu, R., D. Van Vliet, and D.P. Watling (1995) DRACULA: dynamic route assignment combining user
learning and microsimulation. Proc. of the 23rd European Transport Forum, PTRC. Vol. E, pp.143-152.
Machado, R. and R.H. Bordini (2001) Running AgentSpeak(L) agents on SIM-AGENT. Proc. of the
International Workshop on Agents Theories, Architectures, and Languages (ATAL’01). Seattle, 1-3 Aug. 2001.
Mahmassani, H.S. and R. Jayakrishnan (1991) System performance and user response under real-time
information in a congested traffic corridor. Transportation Research 25A, 293-308.
Mahmassani, H.S., S.G. Hatcher, and C.G. Caplice (1991) Daily variation of trip chaining, scheduling, and
path selection behaviour of work commuters. Proc. of the 6th International Conference on Travel Behavior.
Quebec City, 22-24 May 1991.
Rao, A.S. (1996) BDI agents speak out in a logical computable language. Proc. of the 7th European Workshop
on Modelling Autonomous Agents in a Multi-Agent World. Heidelberg: Springer-Verlag (LNAI 1038). pp.42-
55.
Rao, A.S. and M.P. Georgeff (1991) Modeling rational agents within a BDI architecture. Proc. of the
International Conference on Principles of Knowledge Representation and Reasoning (KR-91). San Mateo:
Morgan Kaufmann. pp.473-484.
Rossetti, R.J.F, S. Bampi, R. Liu, D. Van Vliet, and H.B.B. Cybis (2000) An agent-based framework for the
assessment of drivers’ decision-making. Proc. of the 3rd Annual IEEE Conference on Intelligent
Transportation Systems. Dearborn, 1-3 Oct. 2000. pp.387-392.
Rossetti, R.J.F., R.H. Bordini, A.L.C. Bazzan, S. Bampi, R. Liu, and D. Van Vliet (2001) Using BDI agents to
improve driver modelling in a commuter scenario. Transportation Research Part C. (forthcoming in a special
issue on Intelligent Agents in Traffic and Transportation).
Russell, S.J. and P. Norvig (1995) Artificial Intelligence: a modern approach. Englewood Cliffs: Prentice-
Hall, Inc. 1995.
Schleiffer, R. (2000) Traffic itself is simple – just analyzing it is not. Proc. of the 33th Annual Hawaii
International Conference on System Sciences (HICSS-33). Maui, 4-7 Jan. 2000.
Wooldridge, M. (1999) Intelligent agents. In: G. Weiss (editor) Multiagent Systems: a modern approach to
distributed artificial intelligence. Cambridge: The MIT Press. pp.27-77.