From: AAAI Technical Report WS-93-04. Compilation copyright © 1993, AAAI (www.aaai.org). All rights reserved. Traffic ManagementApplications of Neural Networks JohnF. Gilmore andKhalidJ. Elibiary Computer Scienceand InformationTechnology Lab GeorgiaTechResearch Institute GeorgiaInstitute of Technology Atlanta, Georgia30332 NaohikoAbe MitsubishiHeavyIndustries,LTD NagoyaGuidance& PropulsionSystems Aichi-Ken,Japan Abstract usefulnessof all transportation modes.Theintelligent, adaptivecontrol aspectsof this problem are attunedto the features of neural network systems. Neural networks [Andersonet al., 1988]are computational structures that modelsimplebiological processes usually associatedwith the humanbrain. Adaptable and trainable, they are massively parallel systemscapable of learning from positive andnegativereinforcement. AdvanceTraffic Management Systems(ATMS)must be able to respondto existing andpredictedtraffic conditions if they are to addressthe demands of the 1990’s. Artificial intelligence andneural networkare promising technologiesthat provideintelligent, adaptiveperformancein a variety of application domains.This paper describestwo separateneural networksystemsthat have beendevelopedfor integration into a ATMS blackboard Thebasic elementin a neural networkis the neuron architecture[Gilmoreet al., 1993a].Thefirst system is an (Figure 1). Neurons receiveinput pulses(1+) frominteradaptivetraffic signal light controller baseduponthe Hopfield neural networkmodel,while the secondsystem connectionswith other neuronsin the network. These interconnections are weighted(W+) basedupon their is a backpropagation modeltrained to predict urbantraffic contribution to the neuron.Weighted interconnectionsare congestion.Eachof thesemodelsare presentedin detail with results attainedutilizing a discretetraffic simulation summed internally to the neuron and comparedto a shown to illustrate their performance. threshold value. If the threshold value is exceeded,a binary output pulse (O+) is transmitted, otherwisethe 1 Introduction neuronexhibits no output value. A variety of specialized It hasbeenestimatedthat Americandrivers annually neural network modelsbasedupon this simple neuron spendovertwobillion hoursin traffic jamsresulting in an structure havebeendeveloped. estimated$80billion loss to the U.S. economy. Withthis total expected to growfive fold by the year2010andannual traffic fatalities in the UnitedStatesexceeding 35,000,the 12 Ol 11~ W3 developmentof advancedtraffic management systems I~02 hasbecome a majorinitiative in transportation agencies aroundthe world. Thegoal of an ATMS is to optimally manage existing transportation resourcesthrough the use of adaptive control systemsin order to maximizethe efficiency and 85 Thispaperdescribesthe application of neural network to two ATMS functions. First, an intelligent neural network-basesignal light control systemcapable of adaptivelyoptimizingtraffic flow in urbanareasis presented. Second, a neural network model capable of learning howto accuratelypredict traffic congestionis discussed. This neural networkmodelis able to learn basedupon data from previous congestionoccurencesand has producedencouragingresults in forecasting congestionon surfacestreets. Current signal control systems, such as the Los AngelesAutomatedTraffic Surveillance and Control (ATSAC),use responsive control [Rowe, 1991]. comparingactual surveillance data to available model data, a timing plan is selected. This approachis an improvement over the time-of-daytiming plan in instances wheretraffic varies each day. The proposedneural networkapproachextendsthe ATSAC philosophy further by applyinga neuralnetworkoptimizationmodelto actual traffic flow datato determine the signallight settingsthat will produce the mostefficient traffic flow in a specialevent area. Utilizing informationon street segment capacities, traffic flow rates, andpotentialflow capacities,the network modelexaminesthe effects of signal light settings in relation to the traffic flow awayfroma designatedarea. Thesystem"settles" on control settings that maximize the flow and adaptively changesthe signal settings based uponchanges in street segment traffic density. 2 Traffic SignalLight Controller Manyresearchers have demonstratedthat neural networkmethodsbasedon a back-propagationalgorithm are able to deal with complicatednonlinearforecasting tasksin stockprices, electricity demand, andwatersupply [Canuet al., 1990].Byspecifyinginput valuesrepresenting importanttraffic congestionattributes, a neural network architecture capableof capturingthe underlyingcharacteristics of the transportationdomain hasbeendeveloped. workof neuronstodescendonto an energyfunction. Since a discrete time simulationis beingused,a discrete-time modelwasadopted. Thedynamicsof the discrete-time HopfieldNet are given by: At first glance,a traffic flow systemappearsto bean interwovenand connectedarray of road sections whose traffic flow is determined by a seriesof traffic lights. The controlof thesetraffic lightsis vital in orderto allowtraffic to flow throughout the systemwith minimaldelay. A neural networkapproachto the control of signals at traffic intersections is proposed.Becausethe optimal traffic signalconfigurationis not availablea priori, a supervised learning architectu re suchas a back-propagation network is not suitable. Theneural net architecture should do unsupervisedleaming in an optimization network. The Hopfield neural networkmodelwaschosenbecauseof its matchto the traffic signal control problem. Themainparameter in Hopfieldis the energyfunction whichis distributively definedby the connectionarchitecture amongthe neuronsand the weights assignedto each connection. The Hopfield model and the Current developmentsin advancedtraffic control traffic-derived energyfunctionit utilizes for intelligent techniques are giving rise to an increasingrequirement for signal light control are describedin the following subreliable near-futureforecastsof traffic flow. Thesepredictions are required in order to attain the background sections. informationfor solvingtraffic congestion beforeit develops 2.1 The Hopfield Model using methodssuchas "gating" or "dynamicroute guidTheHopfield modelis an additive neural network ance". Existing systemssuch as SCOOT [Robertson et ~ that the individual weightedinputs to al., 1991]only react to presenttraffic patternsand, by model.This means themselves,do not prevent congestionfrom occurring. a neuron are added together to determine the total Conventionaltraffic modelingandsimulation procedures activation of neuron. This activation is then passed can be applied to this problembut have a numberof throughan output function to determinean output value. shortcomings, particularly in real-timeapplications. TheHopfield modelusesa fully interconnectednet- 86 iV Ui = .i__~1 Ti,jV j d- v, = g(u,) Ii modeled as an individual neuron(V0 in a Hopfield neural network.If a light is on, the traffic will flow in the E-W directionat that intersection,otherwise traffic flowsin the N-Sdirection. I, canbe viewedas the input potential into where T(I,j) are the interconnection weights, I(i) arethe inputbiases, U(i) are the internalstates, V(i) are the neuronoutputs, and g(x) is a nonlinearactivation function whichcan taken as a nodeof the networkwith T~,i viewedas the connections between traffic lights. Thefirst stepin anyapplicationof the Hopfieldnetwork is the construction of an energyfunction. Thesystem’s primaryobjectiveis to dispersetraffic awayfroma special of time. In anabstractsense, whichapproaches a hard limiter as Xo tends to zero. An eventin the shortestamount onemethod of dispersingtraffic is to route vehiclesfrom asynchronousupdate rule was used which meansthat road segments containinga large numberof vehicles and neurons are randomly chosen to be updated. This asynchronousupdating scheme tends to greatly reduce a high capacity percentage(vehicle/capacity) onto road with a smaller numberof vehicles anda lower oscillatory or wandering behaviortypical of synchronous segments capacity percentage. This can be achieved by [a] updating schemes. changingsignal lights basedon the potential of a given Hopfield and Tank[Hopfield et al., 1984and 1985] traffic light to increaseflow, and[b] synchronizing signals showthat the dynamicsof this modelfavors state transwith adjacenttraffic lights to maximizeoverall system itions that minimizethe energyfunction throughput. N 1 N N E=-~- )-’. Y, T~j.V~Vj-Z IiV/ ~i=lj=l ’ i=l where Ti, j are the interconnection weights, Vsare the neuronoutputs, Ii are the input biases,and N is the number of signals, so that the networkgraduallysettles into a minimaof this function. Thedifficulty to anyapplicationusinga Hopfield Net is to determinea suitable energyfunction that the networkwill descendon. 2.2 Neural Signal Controller Thegoal of traffic management is to maximize the flow of traffic whileminimizingincidentsanddelaysin aregion. Controllingthe timingof signalsto increasetraffic flow is one methodof supporting this goal. Themanagement of traffic flow by a systemutilizing the control of signals appearsto directly mapinto a Hopfield network. Since traffic lights havetwo states, eachtraffic signal canbe 87 Initial attemptsat creating an energyfunction that addressesthese criteria concentratedon examiningthe numberof vehicles on roadwaysto determine optimal signal light configurations. Experimentationsoonindicatedthat the percentage of the capacity of an incoming roadway wasa better indicator for determiningthe state of anygiventraffic signal. Research into the integrationof road capacity percentagesinto the energyequationwas then undertaken. First, a roadsegment nearingits capacitywill trigger a signal to turn on. Second,optimal flow tendsto favor nearcapacityroadsectionsdirecting their vehiclesonto road sections with a lower percentageof capacity. Because of this, a look at the difference of percentages wasadopted. Thepotentialof a givensignalfor traffic flow could be determinedby addingthe differences between the traffic capacity percentages into the signal andthe percentages of capacity of the flow awayfrom the intersection. Thus,the networktendsto turn traffic lights on whentheyhavea higherpotential for large traffic flow. -% 1 Dbj,c II II C 8 P! l÷tI Priority TimeOn Figure 2. Traffic Signal ControlEnergyFunctionDerivation Thenext step in developingthe energyequationwas to addresssignal light synchronization. If a signal is on, adjacentsignalsshouldtendto turn onto further increase traffic flow. A term wasaddedto the energyfunction to reflect this tendency of anadjacent traffic light to turn green basedon the potential of adjacent traffic. With this addition, the final energyfunction specifyingthe optimal controlof the traffic lights in the simulation wasdefinedas: N (" "~(" p. ,:~ Ll+t~)t, M s:~ L ’J ,:l , J intersections.Thenext termof the energyequationfavors lights whichhavehighpotentialflow out of the intersection andinto the next light. Thefinal portion of the energy equation addressesthe connection betweensecondary signals by measuringtheir contribution to the overall primarysignalspotential traffic flow. If onesignal is on, the equationtendsadjust signals to also settle into on states. A symbolicrepresentation of energy function components is illustrated in Figure2. 2.3 Results where Thegraphical interface displays the current traffic flows, capacities, andpotentials for eachstreet segment ona highresolutioncolor monitorfor viewingby the user. Figure3. is anexample of the neuralnetwork traffic signal controller display for the GeorgiaDome andOmnisports arenas.North, south andeast of the Dome and the Omni are the parkinglots for spectatorsattendingarenaevents. Eachlot hasa numberindicating the numberof vehicles currentlyin the lot. For example, the lot north of the Dome currently contains757vehicles. Nis the number of traffic lights, M& L are the # of outgoingroadways, P(/)is thepriorityof a traffic light t(i) is the timethat traffic light i hasbeenon, C(b) is the capacitypercentof roadsection D(a)(b(j)) is the differencecap(a)-cap(b(j)), Ill is theoutputof neuron i representing traffic light i, r/ is the %of vehiclesturningfroma ontob(j), Skis the %of vehiclesturningfromb(j) ontoc(k). Thenumerator of the termP,/(1 +t,) providesa prioritization weightof individualtraffic lights in the simulation, whilethe denominator negatively weightstraffic lights that have beenon for a longtime. Theenergyfunctiontendsto favor signal lights with large capacitiesflowing throughtheir 88 Traffic data wasgeneratedusing a discrete traffic simulator [Homburger,1982] of downtown Atlanta. This area wasselected becauseit is the site for the 1996 Summer Olympics and presents the most challenging traffic management problemthe UnitedStateswill witness ~.~iiiiiiiiiiiiiiiiiiill iiiiiiii!i!i iiii!i 89 in the 1990’s.Thetotal number of vehicleson eachstreet segment is also indicatedby a numericvalue with the top number indicating traffic flow to the right andthe bottom. numbertraffic flowingto the left. A singlenumber indicates a onewaystreet. Streetlights are designated by a dashin anintersection with the flow of traffic (e.g., the greenlight) moving in the samedirection as the dash.For example, the signal light at the intersection of International and Techwood is currently greenon International, but red on International at its intersectionwith SpringStreet. Theclockin the top right handcorner displays the actual travel time the simulation has beenrunning (e.g. one minutesin this example). employed to learn the special case traffic classesand adaptthe weightsto the presentsituation. In the adaptive learningphase,the error function is computed by placing a restriction on the weightchanges so that the knowledge learnedthroughthe initial learningphaseis retained.The prototypesystemis tested throughcomputersimulations, with results indicating that the applicationof the neural networksto traffic congestionforecastingis promising. 3.1 Neural NetworkFor Forecasting A multi-layerednetworkconsistingof three completely connected layers(i.e. the input layer, the hiddenlayer, and the outer layer) wasdevelopedto addressthe traffic forecasting problem.Thelearning wasachievedusing a Figure4. indicatesthe flow of traffic after ten minutes back-propagation algorithmwith a sigmoidtransfer funcof driving time. This example has assumed that the only tion. Thisfunctionwasverysuitableto the traffic problem traffic flow in the areawasfromthe arenaparkinglots. as traffic flowsalwayspossess a saturationcharacteristic. This wasdoneto illustrate multiple venuetraffic flow Thenumberof input neuronsusedin the initial prointeraction. Thefigure shows that within ten minutesafter totype network was 48. These were composedof the the Dome andOmnieventshaveendedtraffic is efficiently target flow (T in Figure 6) andthree inflows into the beingdisbursed. simulatedarea(11, 12, andIs) whichare closelyrelevant Thegraphical user interface has beendesignedto the target flow. Eachvariable wasnormalizedto [0,1] by providetransportation engineerswith several levels of using the capacity percentage of the road for the target abstractionduringsystemoperation.Figure5. is a display flow andthe maximum valuesof observed data for inflows, of a larger areaof downtown Atlanta, but with only primary respectively.It shouldbe notedthat all data wassampled streets indicated. TheFulton CountyStadiumarea preevery5 minutes, thoughtime samplingin the systemsis viously shown canbeseenin less detail in the bottomright variable. Theoutputsaretarget flowsin the next30 minute protion of this figure. ATMS operators mayselect the period, thus 6 neuronsare required. Althoughthere is granularity of display they wishto viewbasedupontheir muchdifficulty in determiningthe numberof neuronsfor currentinterests. the hidden layer, 12 neuronswere chosenwithout any 3 Traffic Congestion Forecasting effort for optimization in this study. AnATMS mustnot only control currenttraffic, but also Theforecastingalgorithmcontainstwo phases.First, predict wherecongestionwill occur. Predictingcongestion the learning phaseis usedto compute the optimalweights so that preventiveactions maybe takenin advancewill of the neuralnetworkfor a typical pattern. Thesecondis greatly alleviate traffic gridlocks.This sectiondescribes an adaptiveforecastingphaseto adaptthe weightsto the the results achievedutilizing a back-propagation neural ¯networkalgorithmto predict the traffic flow on surface presenttraffic flows andforecastfuture congestion. Thetrainingdataconsistedof traffic flow historiesfrom street in metropolitanareas.Theneuralnetworkis trained in twophases.First, an initial learningphasedetermines a typical business day, and the network connecting the mostappropriate connectingweights for data on a weights werecomputedby following the standardbacktypical business day. Second,adaptive learning is propagationlearning rule describedbelow: 90 0 -r" 0 3.2 Results where 1 ~(gt - 20,) Y,= t’hdesiredoutput Ot = t ’h actualoutput Oncethe initial learningis completed,the networkis ready for the adaptive and forecasting process. The connectingweightsare correctedaccordingto the forecasting error of the last few data through backpropagation.In order to retain the knowledge learnedin the initial learningphase,an error functionwith a termto suppresschangesof weights wasimplemented: The data describedabovewasused for the initial learning. In order to generatetest data, random fluctuations (+20%)wereaddedto the inflow traffic patterns shownin Figure 7, and randomfluctuations werealso added to splitting rates of all flows. Figure8 shows traffic patterns as the simulation results. Threeforecasting criteria, PAAE,standarddeviation of absoluteerror, and PITP, were used to evaluate the system performance [Srirengan et el., 1991]. PAAE(Percentage Average AbsoluteError) andstandarddeviation of absoluteerror are calculatedas PAAE = (l/N) Y. I Yt -O, I /(target range)*lO0 t STDEV =’~/(1/N)~( IYt-O, I-(1/N)Y,, I Yt-O, I )2 wherenote that target rangeis equalto 1 andtherefore PAAE is equivalent to O/N)T. I y,- o, I. E ~ =Eo+kE where t 2Eo =I ~(y,-o,) PITP(Percentageof Incorrect Turning Points) is the percentageof times the prediction of the systemis an increase(or decrease) in a periodwhenin the actualresult is the opposite. 21 -(Wii-W°,’~ W°i= initial weight Threecases(as shownin Figure 9.) weretested for comparison: the casewithout the adaptivelearning (Case AW~,~ = maximum valueof I Wii - W°Ji I 1 ), the casewith the adaptivelearningusingthe standard error function (Case2), andthe casewith the adaptive In addition, a shifting learning methodwasused to take error function(Case3). In all the latest available data into account.Thenetworkwas learningusingthe proposed instances,the correct trend wasforecastedin morethan adaptedto data for the past 2 hours, and wouldthen forecasttrafficflowforthenext hour. Inthe learningperiod, 85%of the time for the test set. Althoughthe forecast accuracybetweenCases1 and2 did not alwaysimprove, all of outputscorresponding to input data(up to the last it wasclearly alwaysimprovedin Case3. This indicates 30 minutes)are available for training data. A portion of the outputs corresponding to input data for the last 30 that the error function wassuccessful in guiding the minutes, however, are not available. This approach network’sadaptivelearning. In addition, it appearsthat of the next5 minutesis superiorand adapted all the weightsto the training data(exceptfor the the forecastaccuracy last 30 minutes)andonly the weightsbetween the availforecastof the next30minutes is poorestin all cases.This is because the data for the next 5 minuteshasa stronger ableoutputsandthe hiddenlayer to the training data from the last 30 minutes.After the adaptationprocess,a traffic correlationwith the input data, as compared to datafor the flow forecastfor the next 30 minutesis generated. next 30 minutesdata. 93 0 r-n, ,r-- 44- C7~ or-, I-._J >. U~ U~ (..er~ t/) or.., q-o r"-, o ol,-, U) (1) 6 (lP el,-. 0, ~3 I ] lil I I c~ n, I 1pro 0 ~oil 1~8 poz~ou 3~ o ._~ tn o 0 (It -F-) 0 er-~ .t~ ;i -R q ’e" I"" u~ m ¯r > 0 z ’" i .A 0 | I ,~-.°o .... V e .... t~ .,.°~lP o ~ u d °°°.~ ,r’quS’O "’ ’ Ol~g’O OllWS"0 Ol~’O 94 0 ,,~ e,,, ¢x: e,*rm Z= o qc-.,- d 4 Summary Twoapplicationsof neuralnetworksin advance traffic management havebeenpresented.Theintelligent traffic signal control hasbeendeveloped andapplied to several specialcasetraffic situationsincludingthe multiplevenue traffic congestion anticipatedduringthe 1996Olympics in Atlanta. Thetraffic congestionforecasting systemhas shownpromise in predicting congestion basedupon learning the factors that contribute to traffic jamsand gridlocks. Gilmore,J. F., andAbe,N., "A NeuralNetworkSystem For Traffic Congestion Forecasting",submittedto International Neural NetworkSociety Conference,Nagoya, Japan, October1993b. Hayer,D.A., andTamoff,P.J., "FutureDirectionsfor Traffic Management Systems", IEEE Transactions on VehicularTechnology,February1991, Vol. 40, No.I. Homburger, W.S.,TransportationandTraffi(;: Engineering Handbook. editor, Institute of Transportation Engineers, Washington,D.C., 1982 Developed independently,researchis continuing to integrate these systemsinto an ATMS blackboardarchitecture containing additional subsystems for incident detection, emergency vehicle management,ramp metering,and traffic monitoring. In this configuration results of predictedtraffic congestion wouldbe postedto a blackboarddata structure. This action wouldactivate the traffic signal control systemwhichwouldattemptto divert traffic fromthe predictedcongestionarea. A more detailed interaction of the ATMS blackboardknowledge sourcescanbe foundin [Gilmoreet al., 1993a] Hopfield, J.J., "NeuralNetworks andPhysicalSystems with Emergent Collective Computational Abilities", Proceedingsof the National Academy of Sciences, Volume 79, p. 2518-2577,1982. 5 References Robertson,D.J., and Bretherton, R.D., "Optimizing Networksof Traffic Signals in Real Time- TheSCOOT Method", IEEETransactions on Vehicular Technology, February1991, Vol. 40, No.1. Anderson,J.A. and Rosenfeld, E., Editors, NeurocomDutina:Foundationsof Research,MIT Press, Cambridge, Massachusetts, 1988. Canu,S., Sobral,R., andLengelle,R., "FormalNeural Network as an Adaptive Model for Water Demand", International Neural NetworkConference-Paris,July 1990, Volume1, p. 131-136 Hopfield, J.J., andTank, D.W., "Neural Computation of Decisions in Optimization Problems", Biological Cybernetics, Volume52, 1985. Kornhauser,A.L., "NeuralNetworkLateral Control of Autonomous HighwayVehicles", Vehicle Navigation and Information SystemsConferenceProceedings,Dearborn, Michigan, October1991. Rowe,E., "The Los AngelesAutomatedTraffic Surveillance and Control (ATSAC) System", IEEETransactions on VehicularTechnology,February1991, Vol. 40, No.1. Srirengan, S., and Looi, C. K., "On Using BackGilmore,J.F., Elibiary, K.J., andPeterson,R.J., "A propagationfor Prediction: An EmpiricalStudy", InterNeural NetworkSystemFor Traffic Flow Management", national Joint Conference on NeuralNetworks-Singapore, SPIEApplicationsof Artificial Intelligence, Orlando,Flo1991,Vol. 2., p. 1284-1289 rida, April 1992. Tank,D.W.,andHopfield, J. J., "SimpleNeuralOptiGilmore,J.F., andElibiary, K.J., "AI in Advanced Traffic mization Networks:An A/D Converter, Signal Decision Management Systems",AAAI93Workshopon Intelligent Circuit, and a Linear Programming", IEEETransactions on Circuits and Systems, v.33 n.5, May 1986. HighwayVehicle Systems,Washington,DC,July 1993a. 95