From: AAAI Technical Report WS-93-04. Compilation copyright 1993, AAAI (www.aaai.org). All rights reserved.

Traffic

ManagementApplications

of Neural Networks

John Gilmore KhalidJ. Elibiary F. and Computer Scienceand InformationTechnology Lab Georgia TechResearch Institute Georgia Institute of Technology Atlanta, Georgia 30332 NaohikoAbe Mitsubishi Heavy Industries,LTD Nagoya Guidance PropulsionSystems & Aichi-Ken,Japan Abstract
usefulness all transportation modes. intelligent, of The adaptive control aspects this problem attuned the of are to features of neural network systems. Neural networks [Anderson al., 1988]are computational et structures that model simplebiological processes usually associated with the human brain. Adaptable and trainable, they are massively parallel systemscapable of learning from positive andnegativereinforcement.

AdvanceTraffic Management Systems(ATMS)must be able to respond existing andpredictedtraffic conto ditions if they are to address demands the 1990s. the of Artificial intelligence andneural network promising are technologies that provideintelligent, adaptiveperformancein a variety of application domains.This paper describestwo separateneural network systems that have beendeveloped integration into a ATMS for blackboard Thebasic elementin a neural networkis the neuron architecture [Gilmore al., 1993a]. first system an et The is (Figure 1). Neurons receiveinput pulses(1+) frominteradaptivetraffic signal light controller basedupon the Hopfield neural networkmodel,while the second system connectionswith other neuronsin the network. These interconnections are weighted(W+) basedupon their is a backpropagation trained to predict urban model traffic contribution to the neuron. Weighted interconnections are congestion. Each thesemodels presented detail of are in with results attainedutilizing a discretetraffic simulation summed internally to the neuron and compared a to shown illustrate their performance. to threshold value. If the threshold value is exceeded, a binary output pulse (O+) is transmitted, otherwisethe 1 Introduction neuron exhibits no output value. A variety of specialized It hasbeenestimated that American drivers annually neural network modelsbasedupon this simple neuron spend twobillion hours traffic jams over in resulting in an structure havebeendeveloped. estimated billion loss to the U.S. economy. this $80 With total expected grow to five fold by the year2010 annual and traffic fatalities in the United Statesexceeding 35,000, the 12 Ol 11~ W3 development advancedtraffic management of systems I~02 hasbecome majorinitiative in transportation agencies a around world. the Thegoal of an ATMS to optimally manage is existing transportation resourcesthrough the use of adaptive control systems order to maximize efficiency and in the Thispaperdescribes application of neural network the to two ATMS functions. First, an intelligent neural network-basesignal light control systemcapable of

85

adaptivelyoptimizingtraffic flow in urbanareasis presented. Second, a neural network model capable of learning howto accuratelypredict traffic congestion is discussed. Current signal control systems, such as the Los AngelesAutomated Traffic Surveillance and Control (ATSAC),use responsive control [Rowe, 1991]. comparing actual surveillance data to available model data, a timing plan is selected. This approachis an improvement the time-of-daytiming plan in instances over wheretraffic varies each day. The proposedneural networkapproach extendsthe ATSAC philosophy further by applyinga neuralnetwork optimizationmodel actual to traffic flow datato determine signallight settingsthat the will produce most the efficient traffic flow in a special event area. Utilizing informationon street segment capacities, traffic flow rates, and potentialflow capacities, network the modelexamines effects of signal light settings in the relation to the traffic flow away froma designated area. Thesystem "settles" on control settings that maximize the flow and adaptively changes signal settings based the upon changes street segment in traffic density.

This neural networkmodelis able to learn basedupon data from previous congestionoccurences has proand ducedencouraging results in forecasting congestionon surfacestreets.

2 Traffic SignalLight Controller

At first glance,a traffic flow system appears bean to interwovenand connected array of road sections whose traffic flow is determined a seriesof traffic lights. The by controlof these traffic lightsis vital in order allowtraffic to to flow throughout system the with minimal delay. A neural networkapproach the control of signals at traffic to intersections is proposed. Because optimal traffic the signalconfiguration not availablea priori, a supervised is learning architectu re suchas a back-propagation network is not suitable. Theneural net architecture should do unsupervised leaming in an optimization network. The Hopfield neural network modelwaschosen because its of match the traffic signal control problem. to

Themainparameter Hopfieldis the energy in function whichis distributively definedby the connection architecture among neuronsand the weights assignedto the each connection. The Hopfield model and the Current developments advanced in traffic control traffic-derived energy functionit utilizes for intelligent techniques giving rise to an increasing are requirement for signal light control are described the following subin reliable near-futureforecastsof traffic flow. These predictions are required in order to attain the background sections. information solvingtraffic congestion for beforeit develops 2.1 The Hopfield Model using methods suchas "gating" or "dynamic route guidTheHopfield modelis an additive neural network ance". Existing systemssuch as SCOOT [Robertson et ~ 1991]only react to presenttraffic patternsand, by model.This means that the individual weighted inputs to al., themselves,do not prevent congestionfrom occurring. a neuron are added together to determine the total Conventional traffic modeling simulation procedures activation of neuron. This activation is then passed and can be applied to this problembut have a number of throughan output function to determine output value. an shortcomings, particularly in real-timeapplications. TheHopfield modelusesa fully interconnected netManyresearchers have demonstratedthat neural networkmethods basedon a back-propagation algorithm are able to deal with complicated nonlinearforecasting tasksin stockprices, electricity demand, watersupply and [Canu al., 1990].Byspecifying et input values representing importanttraffic congestion attributes, a neural network architecture capable capturingthe underlyingcharacof teristics of the transportation domain been has developed. workof neuronstodescendonto energyfunction. Since an a discrete time simulationis beingused,a discrete-time modelwasadopted. Thedynamics the discrete-time of HopfieldNet are given by:
iV U = .i__~1 Ti,jV j di

I i

v, = g(u,)

86

where T(I,j) are the interconnection weights, I(i) arethe inputbiases, U(i) are the internalstates, V(i) are the neuron outputs, and g(x) is a nonlinearactivation function whichcan taken as

modeled an individual neuron in a Hopfield neural as (V0 network. a light is on, the traffic will flow in the E-W If directionat that intersection, otherwise traffic flowsin the N-Sdirection. I, canbe viewed the input potential into as a node the network of with T~, viewed the connections as i between traffic lights.

The first stepin anyapplication the Hopfield of network is the construction of an energyfunction. Thesystems primary objectiveis to disperse traffic away a special from of whichapproaches hard limiter as Xo tends to zero. An eventin the shortestamount time. In anabstractsense, a onemethod dispersingtraffic is to route vehiclesfrom of asynchronous update rule was used which means that road segments containinga large number vehicles and of neurons are randomly chosen to be updated. This (vehicle/capacity) onto road asynchronous updating scheme tends to greatly reduce a high capacity percentage with a smaller number vehicles anda lower of oscillatory or wandering behaviortypical of synchronous segments capacity percentage. This can be achieved by [a] updating schemes. changing signal lights based the potential of a given on Hopfield and Tank[Hopfield et al., 1984and 1985] traffic light to increase flow, and synchronizing [b] signals show that the dynamics this model of favors state transwith adjacenttraffic lights to maximize overall system itions that minimize energy the function throughput. E=-~- )-. Y, T~j.V~Vj- IiV/ Z
~i=lj=l i=l

1 N N

where Ti, j are the interconnection weights, Vsare the neuron outputs, Ii are the input biases,and N is the number signals, of so that the network graduallysettles into a minima this of function. Thedifficulty to anyapplication usinga Hopfield Net is to determine suitable energyfunction that the a network will descend on. 2.2 Neural Signal Controller Thegoal of traffic managementto maximize flow is the of traffic whileminimizing incidents delaysin aregion. and Controlling timingof signalsto increase the traffic flow is one method supporting this goal. Themanagement of of traffic flow by a system utilizing the control of signals appears directly map to into a Hopfield network. Since traffic lights have states, eachtraffic signal canbe two

Initial attemptsat creating an energyfunction that addresses these criteria concentrated examining on the number vehicles on roadwaysto determine optimal of signal light configurations. Experimentation soonindicatedthat the percentage the capacity of an incoming of roadway a better indicator for determining state was the of anygiventraffic signal. Research the integration into of road capacity percentages into the energyequationwas then undertaken. First, a roadsegment nearing capacitywill trigger its a signal to turn on. Second, optimal flow tendsto favor nearcapacityroadsectionsdirecting their vehiclesonto road sections with a lower percentageof capacity. Because this, a look at the difference of percentages of was adopted. potentialof a givensignalfor traffic flow The could be determinedby addingthe differences between the traffic capacity percentages the signal andthe into percentages capacity of the flow away of from the intersection. Thus,the network tendsto turn traffic lights on when havea higherpotential for large traffic flow. they

87

-% 1 Dbj,c

II
C 8

II
P! lt I Priority TimeOn

Figure 2. Traffic Signal ControlEnergy Function Derivation Thenext step in developingthe energyequationwas to address signal light synchronization. a signal is on, If adjacent signalsshould tendto turn onto further increase traffic flow. A term wasadded the energy to function to reflect this tendency anadjacent of traffic light to turn green basedon the potential of adjacent traffic. With this addition, the final energy function specifyingthe optimal controlof the traffic lights in the simulation defined was as:
N (" "~(" p. M L

intersections.Thenext termof the energy equationfavors lights which have highpotentialflow out of the intersection andinto the next light. Thefinal portion of the energy equation addressesthe connection betweensecondary signals by measuring their contribution to the overall primary signalspotential traffic flow. If one signal is on, the equationtendsadjust signals to also settle into on states. A symbolicrepresentation of energy function componentsillustrated in Figure2. is 2.3 Results Thegraphical interface displays the current traffic flows, capacities, andpotentials for eachstreet segment ona highresolutioncolor monitor viewing the user. for by Figure3. is anexample the neuralnetwork of traffic signal controller display for the GeorgiaDome Omni and sports arenas.North, south andeast of the Dome the Omni and are the parkinglots for spectators attending arenaevents. Each hasa number lot indicating the number vehicles of currentlyin the lot. For example, lot north of the Dome the currently contains757vehicles. Traffic data wasgenerated using a discrete traffic simulator [Homburger, 1982] of downtown Atlanta. This area wasselected because is the site for the 1996 it Summer Olympics and presents the most challenging traffic management problem UnitedStateswill witness the

,:~ Ll+t~)t, where

s:~

,:l

, J

Nis the number traffic lights, of M& L are the # of outgoingroadways, P(/)is thepriorityof a traffic light t(i) is the timethat traffic light i hasbeen on, C(b) is the capacitypercent roadsection of D(a)(b(j)) is the difference cap(a)-cap(b(j)), Ill is theoutput neuron representing of i traffic light i, r/ is the %of vehicles turningfroma ontob(j), Skis the %of vehiclesturningfromb(j) ontoc(k). The numerator the termP,/(1 +t,) provides prioritization of a weight individual of traffic lights in the simulation, whilethe denominator negatively weightstraffic lights that have been for a longtime. Theenergy on functiontendsto favor signal lights with large capacitiesflowing throughtheir

88

~.~iiiiiiiiiiiiiiiiiiill

iiiiiiii!i!i

iiii!i

89

in the 1990s.Thetotal number vehicleson eachstreet of segment also indicatedby a numeric is value with the top number indicating traffic flow to the right andthe bottom. number traffic flowing the left. A singlenumber to indicates a onewaystreet. Streetlights are designated a dash anintersection by in with the flow of traffic (e.g., the green light) moving the in same direction as the dash.For example, signal light the at the intersection of International and Techwood is currently greenon International, but red on International at its intersection with SpringStreet. Theclockin the top right handcorner displays the actual travel time the simulation has beenrunning (e.g. one minutesin this example).

employed learn the special case traffic classesand to adapt weights the present the to situation. In the adaptive learningphase,the error function is computed placing by a restriction on the weightchanges that the knowledge so learned through initial learningphase retained.The the is prototypesystem tested throughcomputer is simulations, with results indicating that the applicationof the neural networks traffic congestion to forecastingis promising.

3.1 Neural Network For Forecasting

A multi-layerednetwork consistingof three completely connected layers(i.e. the input layer, the hidden layer, and the outer layer) wasdeveloped addressthe traffic to forecasting problem.Thelearning wasachievedusing a Figure4. indicatesthe flow of traffic after ten minutes back-propagation algorithmwith a sigmoidtransfer funcof driving time. This example assumed the only tion. Thisfunctionwas has that verysuitableto the traffic problem traffic flow in the areawasfromthe arenaparkinglots. as traffic flowsalways possess saturation a characteristic. This wasdoneto illustrate multiple venuetraffic flow Thenumber input neurons of usedin the initial prointeraction. Thefigure shows within ten minutes that after totype network was 48. These were composed the of the Dome Omni and events haveended traffic is efficiently target flow (T in Figure 6) andthree inflows into the beingdisbursed. simulated (11, 12, andIs) which closelyrelevant area are Thegraphical user interface has beendesignedto the target flow. Each variable wasnormalized [0,1] by to providetransportation engineers with several levels of using the capacity percentage the road for the target of abstraction duringsystem operation.Figure5. is a display flow andthe maximum valuesof observed for inflows, data of a larger areaof downtown Atlanta, but with only primary respectively.It should notedthat all data was be sampled streets indicated. TheFulton CountyStadium area preevery5 minutes, thoughtime samplingin the systems is viously shown beseen less detail in the bottom can in right 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 viewbased upon their much difficulty in determining number neurons the of for currentinterests. the hidden layer, 12 neuronswere chosenwithout any 3 Traffic Congestion Forecasting effort for optimization this study. in AnATMS not only control currenttraffic, but also must Theforecastingalgorithmcontainstwo phases. First, predict where congestion occur. Predictingcongestion the learning phase usedto compute optimalweights will is the so that preventiveactions may takenin advance be will of the neuralnetwork a typical pattern. Thesecond for is greatly alleviate traffic gridlocks.This sectiondescribes an adaptiveforecastingphase adaptthe weights the to to the results achieved utilizing a back-propagation neural present traffic flows andforecastfuture congestion. network algorithmto predict the traffic flow on surface The trainingdataconsisted traffic flow historiesfrom of street in metropolitan areas.Theneuralnetwork trained is in twophases. First, an initial learningphase determines a typical business day, and the network connecting the mostappropriate connectingweights for data on a weights werecomputed following the standardbackby typical business day. Second,adaptive learning is propagation learning rule described below:

90

-r"

3.2 Results The data describedabovewasused for the initial learning. In order to generate test data, random fluctuations (+20%) wereadded the inflow traffic patterns to shown Figure 7, and random in fluctuations werealso added splitting rates of all flows. Figure shows to 8 traffic patterns as the simulation results. Threeforecasting criteria, PAAE, standard deviation of absoluteerror, and PITP, were used to evaluate the system performance [Srirengan et el., 1991]. PAAE (Percentage Average Absolute Error) andstandard deviation of absoluteerror are calculatedas PAAE(l/N) Y. I Yt -O, I /(target range)*lO0 =
t

where 1 ~(gt - 2 0,) h Y,= t desired output Ot = t h actualoutput Once initial learningis completed, network the the is ready for the adaptive and forecasting process. The connecting weightsare correctedaccordingto the forecasting error of the last few data through backpropagation. order to retain the knowledge In learnedin the initial learningphase, error functionwith a termto an suppresschanges weights wasimplemented: of E =Eo+kE ~ where 2 =I ~(y,-o,) Eo 2 1 -(Wii-W,~

STDEV =~/(1/N)~( IYt-O, I-(1/N)Y,, I Yt-O, I )2 where note that target rangeis equalto 1 andtherefore PAAE equivalent to is O/N) I y,- o, I. T.
t

PITP(Percentageof Incorrect Turning Points) is the percentage times the prediction of the systemis an of increase decrease) a periodwhen the actualresult (or in in is the opposite.

Threecases(as shown Figure 9.) weretested for in comparison: casewithout the adaptivelearning (Case the AW~,~ maximum of I Wii - WJ I = value i 1 ), the case with the adaptive learningusingthe standard error function (Case2), andthe casewith the adaptive In addition, a shifting learning method used to take was error function(Case In all 3). the latest available data into account.Thenetworkwas learningusingthe proposed instances,the correct trend wasforecasted more in than adaptedto data for the past 2 hours, and wouldthen of the forecast forecasttrafficflowforthenext hour. Inthe learningperiod, 85% the time for the test set. Although accuracybetween Cases and2 did not alwaysimprove, 1 all of outputscorresponding input data(up to the last to it wasclearly always improved Case This indicates in 3. 30 minutes)are available for training data. A portion of the outputs corresponding input data for the last 30 that the error function wassuccessful in guiding the to adaptivelearning. In addition, it appears that minutes, however, are not available. This approach networks of is adapted the weights the training data(exceptfor the the forecastaccuracy the next5 minutes superiorand all to last 30 minutes)andonly the weightsbetween availthe forecastof the next30minutes poorest all cases. is in This is because data for the next 5 minutes a stronger the has ableoutputsandthe hidden layer to the training data from the last 30 minutes. After the adaptation process, traffic a correlationwith the input data, as compared datafor the to flow forecastfor the next 30 minutes generated. is next 30 minutes data.

Wi= initial weight

93

r-n, ,r--

44-

C7~
or-,

I-._J
U~ U~ (..er~ t/) or..,

>.

r"-, ol,-,

q-o

U) 6

(1)

(lP
el,-.

0,

~3
I
1pro 0

c~ n,

] lil
I

~oil 1~8 poz~ou

3~ o ._~ o (It -F-) 0 er-~ .t~ tn 0

-R
u~ mr 0 z

;i

q
e" I""

"
i
,~-.o

>

.A
0

.... .,.~lP

V
.... t~

I
o ~ u

0 ,,~ e,,, x: e,*rm

Z= o qc-.,-

d
.~

,rquSO Ol~gO OllWS "0 Ol~O

"

94

4 Summary Two applicationsof neuralnetworks advance in traffic management beenpresented.Theintelligent traffic have signal control hasbeen developed applied to several and specialcase traffic situationsincludingthe multiplevenue traffic congestion anticipatedduringthe 1996 Olympics in Atlanta. Thetraffic congestionforecasting systemhas shownpromise in predicting congestion basedupon learning the factors that contribute to traffic jamsand gridlocks. Developed independently,researchis continuing to integrate these systems into an ATMS blackboard architecture containing additional subsystems incident for detection, emergency vehicle management,ramp metering,and traffic monitoring. In this configuration results of predicted traffic congestion would postedto be a blackboard data structure. This action wouldactivate the traffic signal control system whichwouldattemptto divert traffic fromthe predictedcongestion area. A more detailed interaction of the ATMS blackboardknowledge sources be foundin [Gilmoreet al., 1993a] can 5 References Anderson,J.A. and Rosenfeld, E., Editors, NeurocomDutina:Foundationsof Research,MIT Press, Cambridge, Massachusetts, 1988. Canu, Sobral,R., andLengelle,R., "Formal S., Neural Network as an Adaptive Model for Water Demand", International Neural NetworkConference-Paris,July 1990, Volume p. 131-136 1,

Gilmore,J. F., andAbe,N., "A NeuralNetwork System For Traffic Congestion Forecasting",submitted Interto national Neural NetworkSociety Conference,Nagoya, Japan, October1993b. Hayer,D.A., andTamoff, P.J., "FutureDirectionsfor Traffic Management Systems", IEEE Transactions on VehicularTechnology, February 1991, Vol. 40, No.I. Homburger, W.S.,TransportationandTraffi(;: Engineering Handbook. editor, Institute of Transportation Engineers, Washington, D.C., 1982 Hopfield, J.J., "NeuralNetworks PhysicalSystems and with Emergent Collective Computational Abilities", Proceedingsof the National Academy Sciences, Volume of 79, p. 2518-2577, 1982. Hopfield, J.J., andTank, D.W., "Neural Computation of Decisions in Optimization Problems", Biological Cybernetics, Volume 1985. 52, Kornhauser, A.L., "NeuralNetwork Lateral Control of Autonomous Highway Vehicles", Vehicle Navigation and Information Systems Conference Proceedings, Dearborn, Michigan, October1991. Robertson, D.J., and Bretherton, R.D., "Optimizing Networks Traffic Signals in Real Time- TheSCOOT of Method", IEEETransactions on Vehicular Technology, February 1991, Vol. 40, No.1. Rowe,E., "The Los AngelesAutomated Traffic 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 propagation Prediction: An EmpiricalStudy", Interfor Neural NetworkSystemFor Traffic Flow Management", national Joint Conference NeuralNetworks-Singapore, on SPIE Applications Artificial Intelligence, Orlando, of Flo1991,Vol. 2., p. 1284-1289 rida, April 1992. Tank,D.W.,andHopfield, J. J., "SimpleNeuralOptiGilmore, J.F., and Elibiary, K.J., "AI in Advanced Traffic mization Networks:An A/D Converter, Signal Decision Management Systems",AAAI93 Workshop Intelligent on Circuit, and a Linear Programming", IEEETransactions on Circuits and Systems, v.33 n.5, May1986. Highway Vehicle Systems,Washington, July 1993a. DC,

95