Modeling And Simulating Traffic From A Cognitive AndQualitative Reasoning Perspective (*)

Abstract

This paper presents a system for trafficmodeling and simulation based on a cognitivemodel of commonsense reasoning abouteveryday physical situations . The main purposeof the system is to support control activities oncircumscribed traffic situations, in whichseveral and sudden changes occur (e.g .temporary circulation occurring in urbanareas) . The system is able to predict thetemporal evolution of a driving activity and ofthe global traffic density in a circumscribed butnot static topological situation. The system hasbeen developed using an Extended version ofthe Qualitative Process Theory (EQPT) whichallows to represent, simulate and managedynamical behaviour of drivers in presence ofchangeable situations .

1 . Introduction

1i recent and increasing area of interest for theapplication of Artificial Intelligence methods andtechniques in traffic problems (Bielli, et al., 1994)

Stefania Bandini( 1 ), Antonella Carassa(2),Paolo Pegurri( 1), Alessandra Valpiani(3)

Department of Computer Science . University ofMilanvia Comelico . 39 20135 Milan Italy

tel +39 2 55006269 fax +39 2 55006276email : bcntdini@hermes .ntc.dsi.unimi .i t

Department of General Psychology, University ofPaduavia Venezia, 8 351.39 Padua Italy

tel +39 49 8276684 fax +39 49 8276600email : carassa@ipdunidx.unipdit

Centerfor Co,si~nitive Science. University of Turinvia Lagrange, 3 10123 Turin Italy

tel +39 11 549475 fax +39 11 549653email: valpiani@psych .unito.i t

concerns several topics : from the representation ofexpert knowledge for traffic control to theimplementation of qualitative models forsimulating traffic dynamics .Within this framework the investigation ofqualitative and human commonsense reasoningabout everyday physical situations (Weld, etal.,1993) can be extended also to traffic problems(Moreno, et al., 1995 ; Sugimoto, et al., 1992), as inclassical modeling where physical models havebeen usefully employed, e. g., fluid-dynamics(Helbing, et al ., 1995).Several reasons support the use of commonsensereasoning about the physical world in modelingtraffic problems :

traditional analytical models often showlimitations in the description and computationof real life physical and traffic situations (e.g .several parameters are difficult to obtain) ;

drivers' decisions are often determined bycognitive patterns depending on individualcurrent goals and constraints (e.g. to join animportant meeting or to be in a hurry), and by

This research has been partially supported by the ex-4090 MURST project "Knowledge Representation and Automated System". Research Unit of?Milan " Dynamical and computational aspects of intelligent agent." .

The aim of this paper is to present a system fortraffic modeling and simulation, based on acognitive model of commonsense reasoning aboutsituations usually well handled by humans(Carassa, et al., 1995 ; Gemimani, et al. . 1996) .This model has been created in order to give arepresentation and simulation framework forperforming causal reasoning activities . From acognitive perspective, the main features whichcharacterize this model derive from the basicassumption of causality by contact : the physicalcontact between two objects (an agent and atarget) is the necessary condition for determining acausal link between two events involving suchobjects .The creation ofthis cognitiveassumptions and

general cognitive models on the involvedphysical situation (e.g . causal reasoning) .

the explicitcomponents

a computational model based onmodel requires some basic

restrictions to be adopted :

representation of the individualoccurring in the cognitive model

and their relationships;

the possibility of switching from somecondition to another by means of explicitactions activated in the simulation .

The computational environment which implementsthis model is an Extension of the QualitativeProcess Theory - QPT - (Forbus, 1984) hereapplied in the cognitive modeling framework (amore formal description of the Extended QPT(EQPT) is presented in Bandini, et al. (1988)) .EQPT introduces the explicit representation andmanagement of actions during the qualitativesimulation of processes . The causality by contactcognitive model within the EQPT computationalenvironment allows physical situations to besimulated following a psychological approach(Carassa, et al., 1995). In this paper the enhancingof the applicative spectrum of this approach bytraffic control problems is presented in terms ofreproducing some basic principles underlying thehuman reasoning about traffic situations .

The adoption "of traffic computational modelsbased on cognitive models or psychologicalassumptions can be found in the literature . In Espie(1992) a psychological analysis of drivingactivities is reported in order to present some

fundamental principles of human behaviour whichallow realistic traffic situations to be modeled.From a psychological viewpoint, such principleshave been studied analysing behaviours and theirdeterminants in the complex interactions that occurduring driving activities (Saad, 1992). From acomputational standpoint, the developed system isbased on self-organizing dynamics. featuringintelligent agents which use their own knowledge,goals. motivations and strategies to decide what todo in different situations, carrying out theirautonomous tasks.

From a cognitive and qualitative representation andsimulation perspective (within the QualitativeReasoning Artificial Intelligence topic) a trafficcontrol model based on QPT is presented inSugimoto (1992) . With respect to this approach,the proposed system (thanks to the mentionedextension) allows a direct representation andintroduction of actions and objects during thesimulation of a process, and the changing ofcurrent parameters associated to the involvedelements of the simulation (e.g . the introduction ofboth static or moving obstacles on a path, thechangincy of the driver speed or of the road widthdue to an accident) .

In the next section, an overview of the maincomponents of the cognitive model representationwithin the EQPT computational framework isillustrated . Section 3 will present an example of thesimulation results obtained using forwardinference . Finally, some concluding remarks willbe drawn.

2. The Representation of the Model

Following the main characteristics of the causalityby contact cognitive model, the representation ofits basic concepts and the modelling of trafficcomponents will be presented. Then, thesimulation mechanism based on the reasoningmechanism of EQPT will be illustrated .

2.1 The cognitive model

The causality by contact cognitive model is basedon the theory of mental models (Johnson-Laird,1983) which defines the reasoning by building andrevising of models (to build and test a sequence of

It has been experimentally demonstrated that thecontact between objects is the crucial aspect ofphysical causality, not only in perception (Leslie,1984 ; Leslie, et al ., 1987 ; Michotte, 1946 ; Sperber.et al., 1995), but also at the cognitive level ofcausal analysis (Geminiani, et al., 1996) . Whenhuman subjects are reasoning about two events thatthey judge as causally linked, they try to representa physical interaction between objects. The mainconcern is in assigning causal roles to two specificobjects (AGENT and TARGET) and to envisagehow they come into contact.A causal model represents an AGENT. aMEDIUM, a TARGET and a set of MODIFIERS.which are elements able to influence the causalreasoning . The MEDIUM is an element (or a set ofthem) that allows the AGENT to come in contactwith the TARGET in spite of their spatial non-contiguity . For example, if the AGENT is a driverand the TARGET is its goal (e .g . the home to bereached from the office) the MEDIUM is the set ofstreets to be passed through. MODIFIERS areelements influencing the driver's path and distancetime from its starting point to the reached goal .It is possible to match this model with the notion ofprocess introduced in EQPT.

'._' The qualitative processes model

As defined in Forbus (1984), the fundamentalcomponents of a physical phenomenon are theobjects (generic entities) involved, the relevantquantities describing important parameters of theobjects, and the processes operating on or betweenthe objects. Processes act on objects by changingtheir quantitative aspects. QPT has been developedwithin this theoretical assumption .The main components of a process are:

examples as a sequence of progressively morerefined representations of a target situation) .

individual entities (or individuals) : a process isapplied on them;preconditions : represent external conditionsbeing verified to activate a process (statingrelations among individuals) ;quantity conditions : represent relationsbetween the involved parameters and variablesto be verified when a process is activated;relations: represent constraints amongquantities ;

" influences : represent the changes of theinvolved quantities (the essence of a process) .

All the quantities are represented in a qualitativeway by the definition of a quantity space (DeKleer. et al., 1984) .With respect to the causality by contact model, theTARGET, the AGENT and the MEDIUM arerepresented in terms of entities . The MODIFIERScan be represented both by entities interactingwith the AGENT and by all the other componentsof QPT.EQPT introduces another important feature inQPT: the possibility to represent and handlechanges within the execution of a process bymeans of direct actions. Actions are particularprocesses acting on current MODIFIERS .

2.3 The model representation

Within this framework, a traffic situation can berepresented by a model consisting of:

a structural component that represents theentities, their relations and their behaviour .AGENT (driver) ;TARGET (final destination)MEDIUM (road sector with entries and exits)MODIFIERS (accidents, traffic lights, static ormoving obstacles, and so on).In this component the MODIFIERS are objectsable to influence the AGENT's driving from itsstarting point to its final destination .

" a modifiable component that represents thequantitative aspects of entities, their relationsand their behaviour.The quantitative aspects involve :- physical properties (e . g., the position of theAGENT, the width of the road sector);- spatial and temporal physical relations (e . g.the distance between the position of theAGENT and the final destination or the intervalbetween the start time and the arrival time).

The AGENT is represented as a solid particle ; thetraffic flow is considered as a fluid dragging theparticle from a starting point to a final destinationand the road sector as an elastic pipe (modifiable inwidth) . The MODIFIERS cause changes on thetraffic flow, increasing or decreasing its density;for example, if a dynamic obstacle hinders the

AGENT's path, it affects directly the progres.sionof the AGENT towards the exit, slowing it dowel .

2.4 The reasoning mechanism

The system draws two kinds of inferences : forwardand backward .

in forward modality the initial conditions (e . s .position of the AGENT, the final destination)and some enabling conditions (e . g. heavy orlight traffic) are given and the final state (thetime spent by the AGENT to get to destination)has to be established;

in the backward modality the initial and thefinal state are Given and the aim is to definesome enabling conditions evolving the modelin the desired way .

Construction Module

arnvauonorocaaurea

Control Module

maldnu,g7raCeduras -

auannficamnaroceaures

I1

CONCLUSION

PREMISES

5YNTACTIC - SEMANTIC ANALYZERu

an"Calmooel

aynam,e mddeia

IT

re~ernqoroceouns

relaaone oetvreen auanaheve asoeep

s-ulauonprxecures

Fig. l . The inferential system architecture

In accordance with the spirit of mental modelstheory, the 'Inferential processes, drawn by thesystem, consist of three phases :" a construction phase,which takes as input the

premises expressing causal events, and

26

generates mental models of the involvedphysical entities and processes.a matching phase, which receives dynamicmodels as input, carries out a series of

comparison between them and produces asoutput a first conclusion .a revising phase, which evaluates the extent towhich the first conclusion can be consideredacceptable, by searching for counter-examples .

In order to draw such kinds of inferences, thesystem builds models of traffic situations, makinguse. of data drawn from initial information about aspecific situation and from its knowledge . Thesemodels evolve simulating how different influenceshold . In the forward inferences, the systemconstructs a single model of the given trafficsituation and simulates its evolution . In thebackward inferences instead, the system buildsalternative models, by means of revisingprocedures (examples), until a model is reached.which evolves in the desidered way. The cognitivetheory specifies how a model can be revised(Geminiani, et al., 1996), and the use of EQPTallows the computer system to represent andmanage the revision of the models . The systemallows to switch from a state to another by meansof explicit actions activated during the simulation .The revision proceeds by trial and error . Itintroduces in succession only one change . Couplesof models, which differ one from the other in asingle aspect are matched in order to draw theprecise effect of the change . Revision proceeds byselection of chanees that lead the final state of amodel closer to the desired situation .

2.5 The EQPT interpreter

The system builds a model of a specific trafficsituation by representing entities and assigningqualitative values to each of their quantitativeparameters (EQPT allows an accuraterepresentation of them) . The behaviour of theentities and their relations are represented byprocesses and actions. Here the main processes arethe traffic flow and the action of the MODIFIERSon this flow .When conflicting influences act on a singlequantity, EQPT is able to find out the resultingdirection of change for it (Bandini et al., 1988 ;D'Ambrosio, 1989). For example, it allows todetermine how the flow density in a temporal and

spatial step is influenced by joint factors, like an

exit and an obstacle . The actions describe theinstantaneous changes which can occur during aprocess, and they allow the MODIFIERS, notpresent at the start time, to be introduced .

The EQPT interpreter follows two steps .

3. The Simulation

Activation of starting conditions - AGENT andother entities are activated, and values areassigned . Then, the starting processes (thebehaviour of the entities) are activated .

"

Modification of the quantitative aspects - Thefollowing steps are executed, until the steadystate is reached (the AGENT gets to the finaldestination) :

1) Control of activity : the interpreter checks ifan action must be activated (it corresponds tothe introduction of a new entity) or an activeelement must be deactivated (i . e. given theposition of the AGENT some MODIFIERSbecome no more influential) .2) Determination and resolution of influences :the interpreter looks for the influences in theactive processes : these will contribute to theevolution from the current state to the next bychanging the quantitative values of the objectson which the processes act. As an example, theflow of the traffic process contains an influencewhich increases, depending on the current valueof density of the flow, the AGENT's vehicleposition by a spatial step .

3) Propagation of influences : when a changeoccurs in a quantity, all the quantities linked toit by a functional dependency are consequentlychanged.

3_ 1 A simulation example

In this paragraph a simple simulation example willbe presented and commented on . In our intentions,the system here shown suits to small sizescenarios, such as traffic control on a limitedtraffic area with works in progress .The preliminary phase in the preparation of asimulation consists in building a scale-model of thereal traffic situation to analyze.

27

AGENT andTARGET will be modeled as follows :

agent(Name) is_a entity withquantities : position(Name) .

pipe(Name,E)

is_a entity withand quantities :

length(Name) &width(Name)

and relations :q_prop(inv,poli,length(Name),density(E)) .

H,gh H,1.Med Med L-Mnd Lo Zen 1- I-Med Med Hr"dF. P« Pa Pm Po Ne6 N<6 Ne6 Ne6

Exmt

Int

EwRYl

AGENT NNiAOIXNG

LhNANCOUSTACLE ENTNY2

Exn2

Clock: 24Current Interval : 13Active Processes :

Flow-outHinder-progressEntry 2Exit2Nat-rowing

Inactive Processes:Exit]Entrv 1

Fig. 2 . A simulation example

A process called flow out has to be added tothe model, to represent the constant influence ofthe final exit and thus the direction of the trafficflow . This process will be always active .

f1ow_out(flow,A,B,E) isa process withindividuals :

agent(A) &environment(E) &

and s_preconditions :runningand q-preconditions :position(A) &

greater-than infnegand influences :

decrease(flow_out_tot(B),density(E)) .

Referring to the case depicted in Fig. 2, an exampleof the entities, processes and actions definitionsrelated to some MODIFIERS follows:

entry2(in2,E,I) isa process withindividuals :

environment(E) &int4 (I )

and s-preconditions : runningand influences :

increase(flow_in2(E),density(E)) .

exitl(outl,E,I)

is_a process with

individuals :

environment(E) &intl(I)

and s-preconditions : runningand influences :

decrease(flow_outl(E),density(E)) .

narrowingl(rl,E,P,I) is_a action withindividuals :

environment(E) &pipe(P) &int3(I)

and s-preconditions : runningand add-list :

decrease(width(P),normal) .

obstacle(Name) isa entity withquantities : site_obstacle(Name) .

hinder-Progress(hp,A,E,O) is_a process withindividuals :

agent(A) &environment(E) &obstacle(O)

and s-preconditions : runningand q-preconditions : position(A)

greater-than site-obstacle(O)and influences :

decrease(very_weak,site_obstacle(O))&decrease(weak,step(E)) .

Each entry (or exit) can be added to the model bymeans of a process which increases (decreases forthe exits) the traffic density in a particular tract ofthe road, slowing down (or speeding up) theprogression of the AGENT toward the finaldestination .To represent street narrowing (widening) an actionhas to be used, which will be enabled only whenthe AGENT is close . In this case, the action willinstantaneously decrease (or increase) the pipewidth, consequently altering the density in thatparticular road sector and in this way modifyingthe AGENT speed .Dynamic obstacles will be represented with anentity and a process . This process will directlyhinder the progression of the AGENT and will beactive until the AGENT reaches and overtakes theslow-moving obstacle .Obviously, the system gives the user a way for thedefinition of a weight for every MODIFIER, by theassigning of values chosen from the quantity space .

To properly handle the building of a scale-modelof the actual situation to simulate, the MODIFIERSshould have a parameter expressing theirplacement in respect to the final destination .In this way we split the road sector in "subintervalsof influence" of the various MODIFIERS whichwill overlap . According to the AGENT positionsome processes will be on (the ones linked to theMODIFIERS which are, spatially, between theAGENT and the final destination) whereas someothers will be off. The AGENT will thus meet,during its ride, a sequence of not-homogeneous in

28

density subintervals, which will influence itsspeed . and consequently, its progression toward thefinal point.With a structure of this kind, the user can evenrepresent, coherently with the real traffic situation,the length of each subinterval .

Let's focus again on Fig. 2 .The simulation will start by assigning values to thedifferent entities and processes according to achosen scenario .At each temporal step, the influences of the activeprocesses and actions will be considered (e.g . Fig.2depicts the simulation state at clock 24) . By meansof the EQPT interpreter, the system will be able todetermine the resulting direction of change givenconflicting influences that act on a single quantity .The AGENT position will be then advanced by aspatial step in dependency of the value of thecurrent subinterval traffic density (the higher thedensity, the shorter the spatial step) . Thisprogression will result, sooner or later, in thechange of interval in which is located the AGENT,disabling the processes and the actions linked tothe MODIFIERS no more influential, and so onuntil the AGENT gets to the final destination .To control the modification of the AGENTposition, the following code will be executed ateach temporal step :

progression(prog,A,E,O) is_a process withindividuals :

agent(A) &

decrease(step(E),positione(A)) &increase(very_weak,tics(O)) .

In order to make the interaction with the systemeasier, a graphic interface is currently beingdeveloped . It will provide the user with simpletools to build models of specific traffic situationsand with a graphical tool to monitor the simulation .

4 . Conclusions

In this paper a system that models and simulatescircumscribed traffic situations has been presented .The system is based on a cognitive theory ofcommonsense reasoning about physical causality .According to the theory, the main features ofcommonsense reasoning are :

environment(E) &clock(O)

and s-preconditions : runningand q-preconditions :

position(A) greater-than infnegand influences :

people reason by building and testing sample

cases;people are able to create an open set ofdifferent models through revision procedures .

All these aspects are reproduced in the system : on

one side the mental model level conveys thereasoning strategies (as an example the revision ofthe models), on the other side the use of EQPTallows mental models to be realized - as dynamicalrepresentations that involve qualitative knowledgeof the physical world - in computational terms .Moreover, EQPT allows to realize the revisionprocedures by means of the representation and themanagement of actions .The main consequence of this approach is theflexibility of the realized system in representing,simulating and managing the behaviour of driversfacing changeable situations .The model has been developed in order to observeand control the behaviour of a single agent . Tomodel and simulate multiagent situations theforthcoming researches will be headed into twomain directions :

applying simultaneously, in a pseudo-parallelfashion, the same model to several agents :joining the basic model with a distributedartificial intelligence perspective, enriching thesingle agents with an individual model ofcausality and. consequently, increasing theirabilities to decide what to do in differentsituations . This perspective requires thedefinition of a class for every single agent (e .g .tropistic, knowledge based) .

The system has been developed in Prolog at theExpert System Lab at the Department of ComputerScience of the University of Milan.

References

Bandini, S ., M. Bruschi, M.G . Filippini and A .Molesini (1988) . An Interpretation of QPTComputational Model According to LogicProgramming, W.N. of Second Workshop onQualitative Physics, IBM Research Center .Pangi, 1988 . Also in : ComputationalIntelligence (1988) (A . Martelli, G. Valle.(Eds .)), 71-81 . North Holland, Amsterdam.

Bielli, M., G. Ambrosino and M. Boero (Eds .)(1994) . Artificial Intelligence Applications to

29

Traffic Engineering, Utrecht, The Netherlands,VSP.

Carassa A., A. Valpiani, G. Geminiani and S .Bandini (1995) . A cognitive model of causalreasoning about the physical world. In : Topicsin Artificial Intelligence (M. Gori, G. Soda(Eds .)), 217-230. Springer Verlag, LectureNotes in Artificial Intelligence 992, Berlin .

D'Ambrosio, B. (1989) . Extending theMathematics in QPT, International Journal ofIntelligent Systems. 4, 55-80.

De Kleer J . and S. Brown (1984) . A QualitativePhysics based on Confluences . ArtificialIntelligence, 24, 7-83 .

Espie. S. and F. Saad (1992) . ARCHISIM:architecture parallele multi-acteurs pour lasimulation microscopique du traffic et ducomportement des conducteurs. INRETSProject, France .

Forbus, K. D. (1984) . Qualitative Process Theory,ArtiftcialIntelligence, 24 . 85-168 .

Geminiani, G. . A. Carassa and B.G . Bara (1996) .Causality by contact. In : Mental Models inCognitive Science (D . Gamham, J. Oakhill(Eds.)), 275-303 . Lawrence Erlbaum, Hillsdale,NJ.

Helbing, D. (1995) . Improved fluid-dynamic modelfor vehicular traffic, Physical Review E. 51.

Johnson-Laird, P. N. (1983) . Mental models,Cambridge University Press, Cambridge .

Leslie . A. M. (1984) . Spatiotemporal continuityand the perception of causality in infants,Perception, 13, 287-305 .

Leslie, A. M . and S. Keeble (1987) . Do six month-old infants perceive causality?, Cognition, 25,265-288.

Michotte, A. (1946) . La perception de la causality,Louvain, Ed. de l'Institut Supeneur dePhilosophie (Etudes de Psychologie), VIII .Trad . (1963) . The perception of causality,Methuen, London.

Moreno, S., F. Toledo, F. Rosich and G. Martin(1990) . Qualitative Simulation for TemporalReasoning in Urban Traffic Control,Proceedings of the 10th InternationalConference on Expert Systems and theirApplications . EC2, Avignon, France .

Saad, F. (1992) . In-depth analysis of interactionsbetween drivers and the road environment:contribution of on-board observations andsubsequent verbal report . Proceedings of the4th Workshop of ICTCT, University of Lund,Bulletin 110.

Sperber, D., D. Premack and A. J. Premack (1995) .Causal cognition . A Multidisciplinary Debate .Clarendon Press, Oxford .

Sugimoto, T., S . Nonaka and H. Oohama (1992) .Traffic Prediction and Qualitative Reasoning,Proceedings of the International Conference onArtificial Intelligence Applications inTrasportation Engineering, Institute ofTransportation Studies, University ofCalifornia, Irvine, Ca.

Weld, D. and J. De Kleer (1993) . Readings onQualitative Physics, Morgan Kaufmann .