From: AAAI Technical Report WS-93-04. Compilation copyright © 1993, AAAI (www.aaai.org). All rights reserved.
Improving the Cost-Performance
Tradeoff
Control using Autonomous Intelligent
in Traffic
Agents
WAYNE IBA
RECOMTechnologies
AI Research Branch, Mail Stop 269-2
NASA Ames Research Center
Moffett Field, CA 94035 USA
IBA~PTOLEMY.ARC.NASA.GOV
1
Introduction
Rapid and concentrated growth in many
urban areas has led to a dramatic increase in the number of commuters and
a subsequent strain on manycities’ traffic signal facilities. Althoughtraffic delays
have almost become an accepted part of
our culture, significant efforts (and funding) are focussed on improving the overall
performance of urban traffic-control systems. Unfortunately, responses to congestion have typically focussed on centralized
control systems that are extremely expensive.
Recent developments in overhead infrared traffic sensors enable inexpensive
access to more detailed information on
traffic conditions [1]. In conjunction with
more intelligent controllers that can take
advantage of the better information, this
new sensor technology can lead to greater
efficiency at muchlower cost.
The task that we pose is to improve the
overall efficiency of traffic flow in urban
55
settings. Morespecifically, we want to find
an ideal tradeoff between the efficiency of
traffic flow and the costs of development,
acquisition, installation, and maintenance
of the necessary physical upgrades. In
order to select the appropriate tradeoff,
we must identify the significant sources of
improvement (over current traffic-control
systems), quantify the resulting improvements, and determine the fiscal impact
of implementing these improvements. We
recognize that current safety requirements
must be strictly maintained and assume
that they are not subject to any tradeoff.
2
Motivation
ground
and
Back-
There are three very obvious benefits provided by a satisfactory solution to the
tradeoff selection problem that we have
undertaken. First, by reducing congestion, a more efficient and responsive signal
control system would save time for every-
one whowould otherwise be sitting in traf- The prediction is a key to efficient traffic
fic. Second, reducing the number of starts control and without it, many useful apand stops would reduce the frequency of proaches will appear inferior.
individual vehicle accelerations, which is
In our current work, we address a very
responsible for most of the fuel consump- modest instantiation of the problem outtion and auto pollution associated with lined above. Using techniques for tempocity driving [2]. We would further ar- ral prediction, we hypothesize that intelgue that reducing the number of starts
ligent local control with detailed sensory
and stops would reduce the number of mi- information can approach or exceed the
nor accidents - a major source of traffic
performance of systems using coordinated
congestion. Third, a low-cost approach timing. That is, by improving responsuch as we suggest would save considerable siveness (via better sensory data and maamounts of money as compared to many chine learning techniques) we can achieve
current initiatives intended to improve ur- much of the same improvements provided
ban traffic flow. As an example of a large through coordination but at significantly
centralized project, San Jose’s Traffic Sig- reduced costs.
nal Management Program currently has a
Naturally, this hypothesis is based on a
projected total budget over 26 million dol- number of assumptions. First, and most
lars [3]. Muchof the cost involves com- obviously, it requires the availability of
munication infrastructure,
which would be low-cost sensors capable of providing rich
unnecessary in our framework.
data about individual vehicles in a timely
Considerable work has been performed fashion. This may require some advances
towards achieving these benefits. Much in image processing as well as sensor techof this work has explored the issue of nologies. We also assume (although the
controller coordination or integration. A hypothesis does is not depend on this) that
second, controller responsiveness, is often providing coordination can improve overconfoundedwith the first. Yet another is- all performance. That is, our approach
sue of more recent interest is that of adapt- does not require coordination but should
ability. These three issues can be thought makeadvantageous use of it whenit is proof as dimensions defining a matrix of pos- vided. Weenvision a hierarchical control
sible control approaches.
architecture similar in spirit to that deWe propose that much of the negative scribed in Cuena, et al [4].
results obtained from experiments with
on-line adaptive approaches is a consequence of poor responsiveness and adapt- 3
Our Traffic Simulator
ability. A responsive controller should not
only react to the current situation, but Because our line of work depends upon
should operate according to predictions
detailed sensory information, we need to
about situations in the very near future. use a simulator that models individual
67
cars.The outputsof our assumedsen- pies a single cell of a grid, and travels
sorsareindividual
vehicle
locations,
veloc-along the road according to stochastic opities,
andperhaps
sizes.
In addition,
our erators. An intersection consists of two
research
interest
is in flexible
controllers
crossing
streams
of cars;thecommon
gridthatutilize
theavailable
sensory
informa-cellcanonlybe occupied
by a single
car.
tionin arbitrary
ways.Thatis,thecon- Thecontroller
mitigates
thecontention
for
troller
shouldnotbe limited
to adjust-thecommon
cellby stopping
traffic
on one
ing a timeperiodby somefactorof the stream, while permitting the other stream
numberof carsdetected
passing
or wait- to pass. The sensors provide the controller
ing.Thesimulator
shouldaUowthecon- with information about the state of the
trollerto performmorecomplexreason-tile world. In the current implementation,
ingthattakesintoaccount
current
sen- the controller has access to (but is not resorvalues
andpredicts
likely
values
inthe quired to use) complete information about
nearfuture.
So,ourneedis fora micro- the simulated world.
simulator
thatdecouples
thesimulation
of With several extensions and modificatraffic
flowfromthebehavior
of thecon- tions that are currently underway, we can
troller;
thiscanbe accomplished
throughprovide a framework enabling traffic ena cleaninterface
between
thesensorval- gineers to experiment with more sophisuesandtheagent’s
control
actions.
The ticated sensors and more intelligent consimulator
onlyprovides
thesensory
infor- trollers than is possible with currently
mation
andresponds
to thecontrol
actionsavailable simulation frameworks. Furtherof thecontrol
agent.
more, this can facilitate the rapid protoCurrentmethodseitherhelpoptimizetyping of advanced traffic control agents,
thetimingof coordinated
signals
within which could then be tested on similar, but
a coarse
model
of vehicular
traffic,
orpro- finer grained, simulators that have yet to
videa fine-grained
evaluation
ofa fixed
co- be developed.
ordination
strategy
(e.g.,TRANSIT7F
and
TRAF-NETSIM
respectively
[5]).However,
A Design for a Flexible
existing
systems
assume
theonlysource
of 4
sensory
information
is froma limited
numTraffic Control Agent
berof loopdetectors
andthecontrol
algorithmis constrained
to a fixedsequence
of Our proposed solution to the tradeoff bemodes.
tween efficiency and cost is based on a simSincea satisfactory
simulator
wasnot ple testable hypothesis - intelligent local
available
offtheshelf,
we developed
our control with detailed sensory information
ownsimple
simulator
thatsatisfies
ourre- can approach or exceed the performance
quirements.
We modified
the NASATile- of systems using coordinated timing. This
worldsimulator
[6]to createa two di- combination of enhanced sensors and armensional
trafficworld.Eachcar occu- tificial intelligence techniques represents a
68
fresh logic applied to the standard traffic- information and the possible changes to
control problem. If the sensory informa- the state of the traffic light. Based on
tion can include a limited level of com- the system’s knowledge about world states
munication between local signals, perfor- and the effects of its actions, it can reason
mance can be further improved. If com- about the consequences of various choices.
munication with a central coordinator is For example, if a platoon of cars is deprovided, even further improvements can tected approaching the intersection, the
be gained. The goal is to provide as much system can reason that changing the light
improvement as possible by combining the at this point will result in greater conmore detailed sensory information and an gestion. However,if other features of the
intelligent controller, while taking advan- world are also true (such as cross traffic altage of existing infrastructure.
ready waiting a certain amount of time),
Theintelligent controller is based on de- then it may need to change states anyway.
velopments in integrated cognitive archiThe system’s knowledge about the
tectures [?]. Our approach views the prob- world and about the consequences of its
lem as that of mapping available sensory actions can be programmed directly into
data onto allowable signal modes, and as- memory at the beginning. Through expesumes that more sensor information en- rience, this knowledgecan be fine-tuned to
ables more efficient control. In order to more closely approximate the actual situaperform the correct mapping, the system tions encountered at a particular intersectreats the sensory data at, and approach- tion. However, even when the knowledge
ing, an intersection as a temporal signa- base is accurate, the system can search for
ture. Previously, we have successfully rep- improvements by evaluating various measures of effectiveness that are continuously
resented and recognized temporal information as probabilistic concepts that are available to the system. These measures
organized within a polythetic decision tree are the basis for determining how effi[8]. The classification of the temporal sig- ciently the system is controlling the flow
nature determines a current world state of traffic.
that prescribes a certain mode of operAlthough the knowledge base is not
ation for the signal. Often, no response static, strict constraints can be imposed
by the controUer is needed, but when the to ensure safety and prevent wild divermodechanges the controller will cycle the gence from intended or acceptable behavsignal through appropriate states in re- ior. For example, the system may discover
sponse to the new conditions.
a modification that would yield a signifiThe particular effectiveness of the ap- cant overall improvementin efficiency but
proach stems from the ability to predict would prove to be psychologically unacthe state of the world a reasonable dis- ceptable to individual drivers. These contance into the future. These predictions
straints, like the initial rules and strateare a function both of the current sensory gies, can easily be specified ahead of time.
69
5
Current Status
ture Work
and Fu-
over the coordinated scheme. These results must be taken with a grain of salt;
they represent a single set of traffic conditions and measure only distance and travel
Wehave implemented the preliminary vertime (speed). Clearly, more extensive exsion of our simulator and it is running
periments are necessary (and are underon a Sun Sparc station under Allegro
way). However, we can interpret the reCommon-Lisp. A graphical interface ussuits as weak evidence in support of our
ing X-windows provides a visual display
primary hypothesis and as an encourageof the movementof traffic in the simulated
ment to continue this line of work.
world and the behavior of the intersection
Above, we outlined a possible design
controllers. Traffic streams, controllers,
for a flexible intelligent controller that is
graphics display features, and the evaluabased on earlier work with autonomous
tion executive all run as separate processes
agents. Most notably, we need to ground
within lisp using the multi-process capathis design in an implementation that can
bilities provided by Allegro Common-Lisp.
be tested against more conventional methTo date, we have implemented two con- ods. This work is currently underway. We
trollers. A pre-timed (non-actuated) con- are also extending the simulator to handle
troller and a simple "rule-based" con- more sophisticated and realistic driving
troller. The rule-based controller consid- behavior in individual cars. In conjuncers the demandon both approaching traf- tion with these developments, we are exfic streams and switches between phases perimenting with other independent variwhen the demand on the stopped stream ables, such as numberof starts and stops,
exceeds a specified factor of the demand and expanding the number of dependent
on the permissive stream. Even this prim- measures we can use to evaluate a conitive controller shows improved perfor- troller’s performance.
manceover the pre-timed one. For the deThese research directions are straightfault traffic conditions that wehave tested, forward extensions of the work described
the rule-based controller improves average in this paper. However, we also intend
speed by 20.6%.
to characterize the improvementsthat are
In a second experiment with similar
possible in order to better assess the utiltraffic conditions but with two traffic
ity of expenditures towards new and more
streams crossing a main street, we com- complex traffic control systems. In some
pared a pair of pretimed controllers in co- cases, cities are spending manymillions of
ordination to two independent rule-based dollars to improve traffic flow. Although
controllers. The rule-based controllers had an improvement may be measurable, it
no information about the actions or states maynot be subjectively significant to indiof the other controller. The independent vidual drivers. Traffic improvement must
controllers achieved a 26.2% improvement be measurable and significant,
but must
70
also be individually relevant; an individual
driver should notice the difference. This
essentially introduces another dimension
or dependent variable along which proposed (and implemented) control systems
should be evaluated.
igate traffic contention, and more importantly, have provided supporting evidence
to our hypothesis that intelligent local
control can compete favorably with centralized control. Weare currently designing and running more extensive experiments that test our hypotheses over a
broader range of conditions and evaluate
6 Conclusions
a more complete range of measures of effectiveness. The primary focus of our conThis research is strongly motivated by at tinuing work is the design and implemenleast three factors. First, growth in urban tation of flexible intelligent controllers.
areas has caused considerable strain on existing systems, causing significant congestion and delays. Second, progress in in- Acknowledgments
flared detectors suggests that a low-cost
We thank A1 Harbury, Rajeev Saraf, and
traffic detector with sufficient resolution
Cindy Mason for lengthy and helpful discan be produced in the near term. These
cussions on the ideas presented in this pasensors will provide more extensive inforper. Mark Drummond helped focus the
mation (e.g., traffic locations and velociideas for the traffic simulator. Peter Friedties over a significant range of distances)
land, Kevin Thompson, and John Allen
at a lower initial cost, and will operate
all provided helpful commentson earlier
with much less maintenance than loop dedrafts of this paper.
tectors. Finally, advances in autonomous
agents and adaptive control mechanisms
have paved the way for more responsive
References
and safer traffic signals. Webelieve our experience with autonomous agents can pro- [1] Wood, R. A., et al. (1992). Uncooled
vide a solid foundation on which to explore Monolithic Silicon Focal Plane Developsignificant improvementsto existing traffic ment. Presented at Infrared Information
Symposium, Detector Specialty Group.
control systems.
Currently, we have implemented and [2] Fuel Efficient Transportation Simulatested an initial version of a micro traf- tion contract. Institute of Transportation
fic simulator that does not constrain the Studies, University of California, Berkeley
research to a particular sensor or control [3] Ron Northouse (personal communicaparadigm. Preliminary experiments with tion), City of San Jose’s Traffic Signal
the simulator and a very simple rule-based Management Program.
agent have demonstrated the expected advantages of using actuated control to mit-
71
[4] Cuena, J., et al. (1992). A general knowledge-basedarchitecture for traffic control: The KITS approach. Proceedings of the Second OECDWorkshop
on Knowledge-based Expert Systems in
Transportation.
[5] McTrans, Center for Microcomputers
in Transportation, Gainsville, FL.
[6] Phillips, A. B., & Bresina, J. L. (1991).
NASA TileWorld Manual. Artificial
Intelligence Research Branch Technical Report, FIA-91-11. NASAAmes Research
Center.
[7] Langley, P., McKusick, K. B., Allen,
J. A., Iba, W. F., & Thompson,K. (1991).
A design for the ICARUSarchitecture.
SIGART Bulletin,
Vol. 2, No. 4, 104109.
[8] l’ba, W.(1991). Learningto classify observed motor behavior. Proceedings of the
Twelfth International Joint Conference on
Artificial Intelligence (pp. 732-738). Sydney: Morgan Kaufmann.
72