Transportation Research Part C 9 (2001) 415±432

www.elsevier.com/locate/trc

A real-time trac signal control system: architecture,

algorithms, and analysisq
Pitu Mirchandani a, Larry Head b,*
a
Department of Systems and Industrial Engineering, University of Arizona, Tucson, AZ 85721, USA
b
Gardner Transportation Systems Inc., 6375 East Tanque Verde #170, Tucson, AZ 85715, USA
Received 1 April 1998; accepted 26 June 2000

Abstract
The paper discusses a real-time trac-adaptive signal control system referred to as RHODES. The
system takes as input detector data for real-time measurement of trac ¯ow, and ``optimally'' controls the
¯ow through the network. The system utilizes a control architecture that (1) decomposes the trac control
problem into several subproblems that are interconnected in an hierarchical fashion, (2) predicts trac
¯ows at appropriate resolution levels (individual vehicles and platoons) to enable pro-active control,
(3) allows various optimization modules for solving the hierarchical subproblems, and (4) utilizes a data
structure and computer/communication approaches that allow for fast solution of the subproblems, so that
each decision can be downloaded in the ®eld appropriately within the given rolling time horizon of the
corresponding subproblem. The RHODES architecture, algorithms, and its analysis are presented.
Laboratory test results, based on implementation of RHODES on simulation models of actual scenarios,
illustrate the e€ectiveness of the system. Ó 2001 Elsevier Science Ltd. All rights reserved.

Keywords: Trac controls; Trac optimization; Real time systems; Adaptive controls

1. The system

Over the last six years, the authors, with some students, faculty and researchers have proposed
and developed a real-time trac-adaptive signal control system, referred to as RHODES.1 The

q
An earlier version of this manuscript was presented at TRISTAN III (Triennial Symposium on Transportation
Analysis), June 17±23, 1998 in San Juan, Puerto Rico.
*
Corresponding author. Tel.: +1-520-290-8006; fax: +1-520-290-8178.
E-mail address: lhead@gardnersys.com (L. Head).
1
The authors also acknowledge the signi®cant support received from the Arizona Department of Transportation,
City of Tucson and other local agencies in Arizona, and the Federal Highway Administration, without which the
authors would not have been able to develop RHODES up to the ®eld testing phases.

0968-090X/01/$ - see front matter Ó 2001 Elsevier Science Ltd. All rights reserved.
PII: S 0 9 6 8 - 0 9 0 X ( 0 0 ) 0 0 0 4 7 - 4
416 P. Mirchandani, L. Head / Transportation Research Part C 9 (2001) 415±432

system takes input from the surface street detectors (allowing whatever technology that is being
utilized: induction loops, video, etc.), predicts the future trac streams at various hierarchical
levels of aggregation, both spatially and temporally, and outputs ``optimal'' signal control settings
that respond to these predictions. The optimization criterion can be any that is provided by the
jurisdiction using the system but must be based on trac measures of e€ectiveness (average de-
lays, stops, throughput, etc.).
The RHODES architecture for surface streets is depicted in Fig. 1 (Head et al., 1992). At the
highest level of RHODES is a ``dynamic network loading'' model that captures the slow-varying
characteristics of trac. These characteristics pertain to the network geometry (available routes
including road closures, construction, etc.) and the typical route selection of travelers. Based on
the slow-varying characteristics of the network trac loads, estimates of the load on each par-
ticular link, in terms of vehicles per hour, can be calculated. The load estimates then allow
RHODES to allocate ``green time'' for each di€erent demand pattern and each phase (North±
South through movement, North±South left turn, East±West left turn, and so on). These decisions
are made at the middle level of the hierarchy, referred to as ``network ¯ow control''. Trac ¯ow
characteristics at this level are measured in terms of platoons of vehicles and their speeds. Given
the approximate green times, the ``intersection control'' at the third level selects the appropriate
phase change epochs based on observed and predicted arrivals of individual vehicles at each in-
tersection.
Essentially, at each level of the hierarchy there is an estimation/prediction component and a
control component. These components are discussed in Sections 2 and 3, respectively. Currently,
we have not done much model development at the highest level (dynamic network loading) and so
we will not include much discussion of level 1 of the RHODES hierarchy.

Fig. 1. The RHODES hierarchical architecture.

 P. Mirchandani, L. Head / Transportation Research Part C 9 (2001) 415±432 417

There are three aspects of the RHODES philosophy that make it a viable and e€ective systems
to adaptively control trac signals. First, it recognizes that recent technological advances in
communication, control, and computation (a) make it possible to move data quickly from the
street to the computing processors (even now most current systems have communication capa-
bilities that are not utilized to their potential), (b) make processing of this data to algorithmically
select optimal signal timings fast, and (c) allow the ¯exibility to implement through modern
controllers a wide variety of control strategies. Second, RHODES recognizes that there are
natural stochastic variations in the trac ¯ow and therefore one must expect the data to stoch-
astically vary (by simply smoothing the data and working with mean values does not make the
actual trac that the system sees smooth and average as assumed by some real-time trac control
schemes). And third, RHODES pro-actively responds to these variations by explicitly predicting
individual vehicle arrivals, platoon arrivals and trac ¯ow rates, for the three corresponding
levels of hierarchies described above.

2. The prediction methods

For pro-active trac control, it is important to predict vehicle arrivals, turning probabilities
and queues at intersections, in order to compute phase timings that optimize a given measure of
e€ectiveness (e.g., average delay). To emphasize this importance, consider an intersection with
several approaches. Associated with each approach are several possible trac movements: left
turn, right turn and a through movement. Any non-con¯icting combination of movements that
can share the intersection at any one time can be assigned a signal phase that allows those
movements protected use of the intersection. Now consider the signal-timing problem given two
possible perfect predictions of arrivals during the planning horizon as depicted in Fig. 2. Each
arrival pattern represents the number of vehicles to arrive at the intersection in ®xed time in-
tervals. Both arrival patterns are identical until time t0 when the signal control has to decide
whether to serve this approach or to serve another approach. In the top case, the demand occurs

Fig. 2. Graphical depiction of the e€ect of future arrivals on scheduling phase sequences and durations.
418 P. Mirchandani, L. Head / Transportation Research Part C 9 (2001) 415±432

immediately following t0 , whereas in the bottom case there is little demand immediately following
t0 and greater demand in the future. In each case, the total number of vehicle arrivals is equal.
However, the optimal signal timings could be signi®cantly di€erent. It is of fundamental im-
portance to know the temporal arrival distribution to build a truly real-time trac-adaptive signal
control logic.
Two issues are important to predict trac ¯ow: (1) the length of the time horizon, and (2) the
number of prediction points per time horizon (called the prediction frequency). The prediction
time horizon provides the real-time trac-adaptive signal-timing control logic with the ability to
plan future signal-timing decisions. If the prediction horizon is short, perhaps several seconds,
then the signal-timing decisions are restricted. For example, if the predictions are made over a 10 s
horizon, the signal-timing logic can only make timing decision that extend or shorten the current
phase. On the other hand, if the predictions are made over a longer horizon, the signal-timing
decisions can include decisions on phase sequencing and phase duration.
The prediction frequency provides information about the distribution of vehicle arrivals over
time. If the predictions are made at a frequency of only one prediction for the decision time
horizon, which, say, is 30 s, then the signal-timing logic must assume that the vehicles arrive
uniformly during that 30 s. If the predictions are made more frequently, say every second over the
prediction horizon, then the signal-timing logic will have a more accurate representation of the
distribution of vehicle arrivals over time.

2.1. The PREDICT approach

The PREDICT algorithm (Head, 1995) uses the output of the detectors on the approach of
each upstream intersection, together with information on the trac state and planned phase
timings for the upstream signals, to predict future arrivals at the intersection under RHODES
control. This approach allows a longer prediction time horizon since the travel distance to the
intersection is longer and the delays at the upstream signal are considered. A bene®t of this ap-
proach is that it includes the e€ects of the upstream trac signals in the intersection control
optimization problem.
To understand how this approach works, consider the scenario shown in Fig. 3. It is desired to
predict the ¯ow approaching intersection A at detector dA . Making the prediction for the point dA
is important because it is a point on link AB where the actual ¯ow can be measured, hence the
quality of the prediction can be assessed in real-time.

Fig. 3. Prediction scenario based on detectors on the approaches to the upstream intersection (B).
 P. Mirchandani, L. Head / Transportation Research Part C 9 (2001) 415±432 419

Trac contributing to the ¯ow at dA originates from the approaches to intersection B and can
be measured at detectors dl , dt and dr representing the ¯ows that will turn left, pass through and
turn right, respectively, onto link AB. It is possible to have other trac that originates at sources
between intersections A and B, but this can be considered as unmeasurable ``noise''. Also, it is
possible that vehicles passing over dl , dt , and dr will terminate their trip before arriving at dA . This
can also be considered as ``noise'' in the prediction. Signi®cant ¯ow volumes that enter or exit the
network midblock can be modeled in the predictions as constant bias terms.
When a vehicle passes a detection point, say di where i 2 fl; t; rg, several factors a€ect when it
will arrive at dA including (1) the travel time from di to the stop bar at intersection B, (2) the delay
due to an existing queue at B, (3) the delay due to the trac signal at B, and (4) the travel time
between B and dA .
Fig. 4(a)±(d) depicts the delay associated with each of these factors. In Fig. 4(a), the vehicle
arrives at detector di and passes freely to detector dA . In Fig. 4(b), the vehicle arrives at detector di
and is delayed by the signal at intersection B. Hence the travel time from di to dA must account for
the travel time from di to the stop bar, the delay due to the signal and the travel time from the stop
bar to dA . In Fig. 4(c) the arrival at di encounters delay for the signal as well as a standing queue,
and has to travel from di to the stop bar at B, and from the stop bar to dA . Fig. 4(d) depicts the
case when the arrival at di occurs after the signal has begun serving the desired phase, but a
standing queue is present. This case is similar to the above, except that the delay due to the

Fig. 4. Delays associated with the prediction of arrivals at the detector dA : (a) detected vehicle passes freely through
intersection; (b) detected vehicle arrives during red signal ± signal delay; (c) detected vehicle arrives during red signal
and a queue exists ± signal and queue delay; (d)detected vehicle arrives during green signal and a queue exists ± signal
and queue delay.
420 P. Mirchandani, L. Head / Transportation Research Part C 9 (2001) 415±432

standing queue must be adjusted based on the amount of time that has elapsed between the onset
of the signal and the arrival of the vehicle at di and the travel time to the back of the queue.
Once the arrival time at dA is predicted, the PREDICT model adds a fraction to the current
estimate of the expected number of arrivals at that time. For example, if 15% of the vehicles that
pass over di continue on to dA , then for each actuation of di , 0.15 is added to the current estimate
of the expected number of arrivals at the predicted arrival time.

2.2. Estimation of parameters

Observe that to use the PREDICT model, several parameters (given in bold) need to be pro-
vided: (1) travel times on links (detector to detector) which depends on the link free-¯ow speed and
current trac volumes, (2) queue discharge rates which also depends on volumes (as well as on
queue spillbacks and opposing- and cross-trac volumes), and (3) turning probabilities. In ad-
dition to these parameters, to estimate arrivals and demand for various phases, we also need to
have estimates of queues at the intersections and the ramps.
Link free-¯ow speed can be estimated from historical data and capacity analysis. Link free-¯ow
speeds are needed even in traditional o€-line models to optimize ®xed phase timings (cycle times,
o€sets, splits), so to obtain these should pose no major problem.
Through-trac queue discharge rates are e€ected by downstream through-trac volumes,
which can be easily measured. Likewise, left-turn queue discharge rates depend on opposing trac
volumes, and right-turn queue discharge rates depend on cross-trac in that direction. These
three discharge rates are initially given from calculated default functions ± functions of trac
volumes, but are then adjusted based on how well they predict remaining queues at the stop-bar
presence detectors. For example, if the left-turn queue estimate tends to be non-zero when in fact
it is zero, then the left-turn discharge rate is adjusted upwards; this procedure will be explained
shortly.
TURN algorithm. An assumption for RHODES (as well as current o€-line methods to set signal
timings) is that some estimates for turn probabilities at the intersection are given. Even the
CORSIM2 model needs information about the turning probabilities when modeling trac
movement. However, from the real-time trac control perspective, these probabilities are not
deterministic; they change stochastically over time. For example, suppose PWN is the probability
that a vehicle arriving from the West to some intersection will turn left (North), then it is clear that
PWN will depend on the time of the day, the volume of trac, and the particular mix of the origins/
destinations in the group of arrivals being modeled. In other words, PWN is described by a random
process.
Our assumptions for PWN are (1) a prior estimate is available whose uncertainty is modeled with
a normal distribution with known mean and variance; (2) at any given time, we have measured the
percentage of vehicles that have turned left in the last, say, 5 min, as well as percentages that
turned right and driven straight through the intersection; and (3) we know the error distributions
for these measurements. We had a choice of three turning probabilities models: (1) information
minimization/entropy maximization (Mekky, 1979; van Zuylen, 1979; Hauer et al., 1981);

2
CORSIM is a software package for modeling/simulating trac on a network; it has been developed by FHWA.
 P. Mirchandani, L. Head / Transportation Research Part C 9 (2001) 415±432 421

(2) Bayesian (Maher, 1984); and (3) maximum likelihood (Maher, 1984; Nihan and Davis, 1989).
The Bayesian model was picked for implementation since the other two models involved a non-
deterministic number of iterations based on an error tolerance whereas Bayesian method con-
sisted of exactly seven iterations. In this method, prior variances for the turning volume errors are
used along with the prior means. The covariance's of the turning volume errors are assumed to be
zero since the trac detectors are assumed to operate independent from one another.
QUEUE algorithm. There have been a few algorithms that have been reported in the literature
that address the problem of estimating queues at an intersection using detector information, most
notably that of Baras et al. (1979). However, these are not applicable here because they require
excessive computational e€ort and time that is not available in our real-time prediction scenario.
Instead, for our purpose, we developed a simple estimation procedure of accounting for arrivals
and estimated departures based on queue discharge rates. Suppose at the beginning of a green
phase, say at time t0 , our initial queue estimate at some stop-bar is q…t0 †. At the end of the green
phase, say at time t1 , the remaining queue q…t1 † is given by
q…t1 † ˆ q…t0 † ‡ a…t1 ; t0 † d…t1 ; t0 †;
where a…t1 ; t0 † is the number of predicted arrivals between t1 and t0 , and d…t1 ; t0 † the predicted
number of departures (using a given queue discharge rate). What allows us to keep biases from
creeping into the estimates is that at some epochs we are certain that the queue length is zero,
speci®cally when there is no queue at the stop bar as con®rmed by the stop-bar presence detector.
If the estimated queue is positive while the stop-bar presence detectors indicate no queue, then we
e€ectively decrease our estimate and make it zero. If the estimated queue is zero while the stop-bar
indicates a positive queue, then the estimated queue is increased by some positive number such as
one.
When the queue discharge rate at the stop bar is estimated well, it would be expected that, on
the average, half the time the estimated queues will be greater than the actual queues and half the
time less than actual queues. If the estimated queues more often than not tend to be higher than
the actual queues (i.e., when there were no vehicles while the queue estimate was non-zero), then
we adjust the queue discharge rate upwards by a small amount. If it tends to be less (i.e., when
there were vehicles while the estimates were zero), then we adjust it downwards by a small
amount. We note that the adjustment of queue discharge rate is only possible when there is no
queue spillback into the intersection; in this case queue discharge rate is zero because of blockage
and does not depend on trac volumes. The exact adjustment step size can be con®gured by the
system operator, but typically step sizes of 0.1 vehicles/cycle are sucient.
It is important to note that the PREDICT model is based on processing arrival data as it
becomes available. At any point in time, the predicted arrival ¯ow pattern at dA accounts for
vehicles that have already passed the detectors dl , dt and dr . The bene®t of this vehicle-additive
process of the predictor is that it constantly provides, for a given prediction horizon, (1) nearly
complete information of anticipated vehicle arrivals in the very near future (of those vehicles that
have already passed the upstream intersections) and (2) partial information of anticipated vehicles
in remaining part of the prediction time horizon (of those vehicles that have not passed the up-
stream intersections, since some new vehicles may still arrive that will e€ect the delays in the
prediction time horizon). Results of an evaluation study of the PREDICT algorithm for arrivals
at an intersection have been reported by Head (1995).
422 P. Mirchandani, L. Head / Transportation Research Part C 9 (2001) 415±432

2.3. Network ¯ow prediction

The resolution of trac at the network ¯ow control level (i.e., level 2 of the RHODES hier-
archy) is in platoons. The scope of the prediction is a subnetwork of several intersections (the
number of intersections depends on the computational power available but we envision that nine
intersections can be controlled by RHODES using only an Intel Pentium processor) with a larger
decision time horizon. Typically, RHODES will use a 20±40 s rolling horizon to predict arrivals
and queues at each intersection, based on upstream detector data; at the network ¯ow control
level, RHODES will use a 200±300 s rolling horizon.
At the subnetwork level, the APRES-NET model (Dell'Olmo and Mirchandani, 1996) is a
simpli®ed trac simulation model based on the same principles as the PREDICT model described
above, but instead of propagating a single vehicle at a time from upstream intersections, it
propagates platoons of vehicles through a subnetwork of intersections. It is necessarily a sim-
pli®ed model because it is used as an objective function evaluator, or as a network wide per-
formance estimator, for the network control logic.

3. The control algorithms

Fixed control strategies are based on a signal-timing plan de®ned in terms of operating pa-
rameters for traditional signal control, namely cycle time, splits, and o€sets. These parameters are
generally developed based on trac studies and standard procedures, such as the Highway Ca-
pacity Manual, or signal-timing software such as TRANSYT and PASSER. The trac studies
result in estimates of trac conditions, link volumes and turning percentages, for speci®ed time
periods. Signal-timing parameters are developed for each of these time periods and, typically,
implemented on a time-of-day basis with no consideration of current actual trac conditions. In
many cases, even the use of standard procedures for the development of signal-timing plans is
abandoned and trac engineers operate in a judgment-based fashion with moderate levels of
success. None of these approaches is truly trac-adaptive or even attempt to actually minimize
some measure of trac performance such as average vehicle-delay.
Currently available trac responsive systems attempt to address the problem of responding to
actual trac conditions by switching these parametric signal-timing plans based on current wide-
area trac conditions rather than time of day. This requires that signal-timing parameters be
developed for a variety of possible trac conditions. Nevertheless, implicit in the usage of
parametric timing plans is the assumption that for the next several minutes, or even hours, the
trac in the network can be well characterized by the measured average ¯ows and parameters. No
account is taken of the fact that the second-by-second and minute-by-minute variability of trac
are signi®cant and plans based on averages produce unnecessary delays for some trac move-
ments when the trac on con¯icting movements is absent, or very small, during some periods.
The RHODES approach is to predict both the short-term and the medium-term ¯uctuations of
the trac (in terms of individual vehicle arrivals and platoon movements, respectively), and ex-
plicitly set phases that maximize a given trac performance measure. Note that we do not set
timing plans in terms of cycle times, splits and o€sets, but rather in terms of phase duration for
any given phase sequence. (RHODES does not necessarily require a pre-speci®ed phase sequence,
 P. Mirchandani, L. Head / Transportation Research Part C 9 (2001) 415±432 423

but since many trac engineers prefer a pre-speci®ed sequence, RHODES has been developed to
allow the trac engineer to specify a desired sequence.) In other words, in the RHODES control
strategy, the emphasis shifts from changing timing parameters in reacting to trac conditions just
observed to pro-actively setting phase duration for predicted trac conditions.

3.1. Intersection control

At the lowest level of the RHODES hierarchy for a surface street network, i.e., at the inter-
section control level, RHODES uses a dynamic-programming based algorithm COP (Sen and
Head, 1997). There are other signal-timing schemes, which have been experimented, that do not
provide parametric timing plans but instead provide phase duration, notably OPAC (Gartner,
1983; Gartner et al., 1991) and PRODYN (Khoudour et al., 1991) and UTOPIA (Mauro and
DiTaranto, 1990). In some ways these too use dynamic programming (DP) or related optimiza-
tion schemes, but, in their current implementations, the underlying models are more approximate
and the methods are not as ecient.
Fig. 5 depicts the states of the DP formulation. A rolling horizon approach is used to allow the
optimization to take advantage of the most recent predictions and observations. An optimization
is started at some time t0 and considers a time horizon of T s, say 45±60 s. Each stage of the DP is
associated with a signal phase. A phase order is provided to COP so that a stage corresponds to a
phase. The DP state variable sj is the amount of time that has been allocated to all past phases
1; 2; . . . ; j. The decision in stage j is to allocate xj time units to phase j. It should be noted that in
general there are more stages in the DP's planning horizon than the number of phases used for
control. If there are P phases and N stages …N > P † some of the phases may be repeated as stages.
If the trac engineer does not restrict the phases to be in a particular sequence, then this ¯exibility
allows for variable phase sequencing through phase skipping (by e€ectively allocating zero time
for the corresponding stage).
Each decision xj has an associated value based on a performance measure such as stops or
delay. This value is determined by using the predicted vehicle arrivals, the current and prior
decisions, and an imbedded trac ¯ow model that accounts for estimated queues, startup lost
time, queue discharge and arrivals, as well as other trac dynamics that relate the decision to the
performance measure.
The DP is completed when each possible decision for each stage has been evaluated in a for-
ward recursion. Then a backward recursion is used to determine the sequence of phases and phase
duration that will result in the lowest value of the performance measure over the optimization
horizon.

Fig. 5. Stages and states of the COP model.

424 P. Mirchandani, L. Head / Transportation Research Part C 9 (2001) 415±432

The decision for the ®rst stage of the optimization is implemented as the desired signal control.
Just prior to the end of this ®rst phase, the optimization problem is solved again in a rolling
horizon approach. The sequence of phases in the second optimization begins with the current
phase, which allow for the phase to be terminated early or extended, based on the re-evaluation
with more recent observations and predictions.

3.2. Network ¯ow control

The network ¯ow control logic is based on a model called REALBAND (Dell'Olmo and
Mirchandani, 1995) which optimizes the movement of observed platoons in the subnetwork. If
minimizing total stops was the measure of network performance, then REALBAND attempts to
form progression bands based on actual observed platoons in the network. In general, any delay
and/or stops based measure of performance may be optimized.
The basic idea of REALBAND can be understood by considering the following example. Fig. 6
shows a small network with several platoons of vehicles traveling in di€erent directions. Platoons
are de®ned from observed detector data as a ¯ow density above a pre-speci®ed level for some
length of time. Each platoon is characterized in terms of size (number of vehicles) and speed.
When two (or more) platoons are predicted to arrive at an intersection and request opposing
signal phases, a con¯ict is said to occur. Fig. 7 depicts the type of con¯icts that may occur. A
decision tree is built where each branch of the tree represents one possible resolution of a con¯ict.
The decision tree developed is based on the predicted platoon movement over some pre-de®ned
horizon, such as 200±300 s, with node and two out-links for each con¯ict resolution. Fig. 8 shows
the decision tree for the example.

Fig. 6. An example to illustrate REALBAND.

P. Mirchandani, L. Head / Transportation Research Part C 9 (2001) 415±432 425

Fig. 7. Realbands for the example of Fig. 6.

Fig. 8. REALBAND decision tree for the example of Fig. 6

426 P. Mirchandani, L. Head / Transportation Research Part C 9 (2001) 415±432

REALBAND evaluates, using APRES-NET, the performance for each branch of the decision
tree. When all branches have been explored, a path on the tree (corresponding to a set of con¯ict
resolutions) is chosen with best-estimated performance. For example, in Fig. 8, the path with 550 s
of total delay would be selected.
The REALBAND decisions are used as constraints to the intersection control logic (COP).
When COP begins it's rolling horizon optimization, a set of decisions on phase durations in the
phase order is required to accommodate any constraints that REALBAND con¯ict resolutions
impose, with a relaxation that COP may adjust the phase start and end times based on recent, and
more accurate observations of the vehicles in each platoon.

4. The RHODES prototype

The current version of the RHODES prototype logic is depicted in Fig. 9. The prototype
consists of ®ve modules: (1) intersection optimization logic, (2) link ¯ow prediction logic, (3)
network ¯ow optimization logic, (4) platoon ¯ow prediction logic, and (5) parameter and state
estimation logic. The intersection optimization logic and the link ¯ow prediction logic together
form the intersection control logic. The network ¯ow optimization logic and the platoon ¯ow
prediction logic together form the network control logic.
The RHODES prototype is currently under evaluation by FHWA as one strategy within the
real-time trac-adaptive signal control system (RT-TRACS) development program. This pro-
totype has been implemented in software and evaluated using FHWA's CORSIM simulation
model. This implementation is described below.

Fig. 9. Components of the RHODES prototype.

 P. Mirchandani, L. Head / Transportation Research Part C 9 (2001) 415±432 427

4.1. Prototype software design

This software prototype has been developed over a ®ve-year period and was designed as a
modular and portable software tool to support research and development of a wide variety of
trac management and control strategies. It should be emphasized that this software is not in-
tended to be an implementable or deployable system, but is a research tool. The software has been
developed to interface to the latest version of CORSIM trac simulation model as is supported
by Kaman Sciences Corporation for FHWA.
Fig. 10 presents an overview of the software architecture. The design centers around a simple
¯at (or blackboard) database and an executive event controller. The database contains all relevant
network and control information. The executive event controller schedules di€erent types of
control related events such as updating surveillance information, setting signal control at desired
intersections, running network coordination algorithms, running intersection optimization algo-
rithms, and running trac ¯ow prediction algorithms.
The database contains three types of information: dynamic data, model parameters, and static
data. Dynamic data refers primarily to data that changes on a second-by-second basis, such as
vehicle detector information, past and planned signal control states, and trac ¯ow predictions.
The time trajectory of dynamic data is important to the control algorithms. Model parameter data
refers to information that is either constant or changes slowly over time such that only the current
value is relevant and the time-trajectory (past and future) is less relevant. Examples of model
parameters include turning percentages, queue departure rates, and average link travel speeds.
These values may vary slowly over time and can be estimated by an appropriate algorithm, but for
the purpose of modeling the trac dynamics for intersection control, they are treated as constant
parameters. The model parameter data also includes information related to the interface between
control algorithm components.
Examples of interface information include signal-timing coordination and signal-timing con-
straints (e.g., a minimum green time for a phase, or a constraint that requires a phase be served at
a given point in time to ensure coordination). Static data includes values that are assumed to

Fig. 10. Software architecture.

428 P. Mirchandani, L. Head / Transportation Research Part C 9 (2001) 415±432

remain constant. Network geometrics (node identi®cation numbers, number of lanes on each
approach, link lengths, location of detectors, etc.) are the primary types of static data.
When the prototype is initialized at run time, for example, when CORSIM calls the prototype,
a disk ®le is read that initiates and controls the creation of the necessary data structures for the
database. The disk ®le format is very similar to a CORSIM *.trf ®le. Each record in the ®le has a
card type identi®er that speci®es the information according to a pre-de®ned format. Examples of
card types include link geometrics, node de®nitions, detector locations, signal-timing parameters,
and optimization logic parameters.

4.2. Analysis using simulation models

It is clear that any type of trac control algorithm needs to be tested in the ``laboratory'' before
it is implemented and evaluated in the ®eld. The most appropriate method to do this ``laboratory''
testing is to (1) have a realistic simulation model of trac ¯ow at an interchange, (2) emulate the
detection of the trac ¯ow, and (3) observe the resulting changes that would come about if the
algorithm was implemented in place of the current control system. CORSIM was used for lab-
oratory testing the real-time trac control algorithms. An actual set of intersections was selected
for the simulation model.
There are many factors that must be considered in the evaluation, including
1. the type of network, including issues such as number of trac signals, spacing between signals,
2. the trac demand, including time varying demand, vehicle mixtures, bus stations and head-
ways, and pedestrians,
3. statistical issues such as how to characterize the demand, what measures should be examined,
how these measures are de®ned, as well as the number of simulation replications that must be
made to support statistical conclusions or statements.
In this discussion, we will present results based on an FHWA test case used by ITT Systems and
Sciences Corporation as part of the RT-TRACS prototype evaluation process. The network is
based on a section of Tara Blvd. in Atlanta, Georgia. The network consists of nine intersections
along a 17.7-km (11-mile) arterial. The trac volume represents a peak ¯ow that gradually in-
creases and then gradually decreases over a period of 2 h.
The results presented below are based on comparing average delay per vehicle on the sub-
network of links that are directly e€ected by the RHODES prototype and, on the same set of
links, as controlled be semi-actuated logic using timings derived from several optimization
packages, including PASSER and TRANSYT. In reporting these results, we draw a useful
analogy with trac on communication networks, as described in Fig. 11.

Fig. 11. O€ered load, throughput and delay in a communication network.

 P. Mirchandani, L. Head / Transportation Research Part C 9 (2001) 415±432 429

The evaluation of communication networks, as a system of links and nodes, has received
considerable attention in the literature (Bertsekas and Gallager, 1992). Fig. 12 shows the rela-
tionship between the o€ered load, throughput, and delay in the context of routing and ¯ow
control. The analogy in trac control is that the o€ered load represents the demand, ¯ow control
corresponds to signalization (and e€ective capacity), and routing corresponds the signal timing
(which is related to the capacity). The throughput in the communication system is analogous to
the trips in a transportation network. The rejected load can be thought of as the unserved demand
(spillback queues of vehicles).
Fig. 12 illustrates how good signal control (or good routing) can e€ect system performance. In
periods when the demand does not exceed the e€ective capacity, the throughput is equal to the
o€ered load. If poor signal control is used, then the e€ective capacity, in terms of served load, is
reduced. Then the throughput is less than the o€ered load and delay increases considerably. When
good signal control is used, the throughput is increased and the delay performance is improved.
This analogy with a communication system suggests that we need to measure both the o€ered
load, the throughput, and the delay when considering the performance of a new signal control
strategy. It also suggests the need to consider the performance at various loading levels. The
o€ered load is measured as the number of trips on each link entering the network under study.
The throughput is measured as the number of trips on each link exiting the network.

Fig. 12. Relationship between o€ered load, throughput and delay.

430 P. Mirchandani, L. Head / Transportation Research Part C 9 (2001) 415±432

Fig. 13 shows the measured throughput and delay at each of the o€ered load levels for
RHODES and the semi-actuated control for Tara Blvd. in Atlanta. Fig. 13(a) shows, as expected
for the congestion levels simulated, throughput is equal to the o€ered load, implying that the
steady state has been achieved. Fig. 13(b) shows a behavior that is consistent with the notion that
improved signal control provides less delay. Thus, this simulation study suggests that the
RHODES prototype is more ecient in utilizing the capacity of the network.
Observe that average vehicle delays decrease in the range of 50% (for low loads) to 30% (for
high loads). In the high load case, not only are the average delays smaller, but also the variance of

Fig. 13. (a) O€ered load versus delay and; (b) throughput.
 P. Mirchandani, L. Head / Transportation Research Part C 9 (2001) 415±432 431

the delay is signi®cantly reduced, making the movement through the network more predictable for
the driver.
We have implemented RHODES on several CORSIM models of actual transportation net-
works. Most of these networks have consisted of one single arterial with several cross arterial that
carry considerable ¯ow. We have also implemented RHODES at a diamond interchange where
two signals are closely spaced and must operate in a highly coordinated manner. The results of all
of these tests were remarkably consistent. The RHODES logic appears to take advantage of the
natural stochastic variations in trac ¯ow ± as it was designed ± in decreasing trac delay.

5. Conclusions

In this paper, we have presented the RHODES real-time trac-adaptive signal control system,
the software architecture that has been developed, and a simulation-based analysis of the system.
The simulation experiments show promising results that encourage future experimentation. Based
on these results, three ®eld tests are being planned, one for a 10-intersection arterial segment in
Tucson, AZ, (funded by the City of Tucson and FHWA), a 9-intersection arterial segment in
Seattle, WA, (funded by FHWA), and a diamond interchange in Tempe, AZ, (funded by ADOT
and the City of Tempe).

Acknowledgements

The Federal Highway Administration, the Arizona Department of Transportation, and the
Cities of Tucson and Tempe have supported this research. The authors would like to thank and
acknowledge Debbie Curtis and Raj Ghaman of FHWA; Jim Decker of the City of Tempe, AZ;
Dr. Richard Nassi of the City of Tucson, AZ; Steve Owen and Tim Wolfe of ADOT; Dr. Sarath
Joshua of Maricopa Association of Governments, Phoenix, AZ; Charlie Stallard and Larry
Owens of ITT Systems and Sciences Corporation; Paolo Dell'Olmo of the University of Roma,
``Tor Vergata'', Italy; and Suvrajeet Sen, David Lucas, Steve Shelby and Doug Gettman of The
University of Arizona.

References

Baras, J.S., Levine, W.S., Lin, T.L., 1979. Discrete-time point processes in urban trac queue estimation. IEEE Trans.
Autom. Control AC-24, 12±27.
Bertsekas, D., Gallager, R., 1992. Data Networks. Prentice-Hall, Upper Saddle River, NJ.
DellÕOlmo, P., Mirchandani, P.B., 1995. REALBAND: an approach for real-time coordination of trac ¯ows on a
network. Transport. Res. Record 1494, 106±116.
DellÕOlmo, P., Mirchandani, P.B., 1996. A model for real-time trac coordination using simulation based
optimization. In: Bianco, L., Toth, P. (Eds.), Advanced Methods in Transportation Analysis. Springer, Germany,
pp. 525±546.
Gartner, N.H., 1983. OPAC: a demand-responsive strategy for trac signal control. Transport. Res. Record 906, 75±
81.
432 P. Mirchandani, L. Head / Transportation Research Part C 9 (2001) 415±432

Gartner, N.H., Tarno€, P.J., Andrews, C.M., 1991. Evaluation of the optimized policies for adaptive control (OPAC)
strategy. Transport. Res. Record 1324, 105±114.
Hauer, E., Pagitsas, E., Shin, B.T., 1981. Estimation of turning ¯ows from automatic counts. Transport. Res. Record
795, 1±7.
Head, K.L., 1995. An event-based short-term trac ¯ow prediction model. Transport. Res. Record 1510, 45±52.
Head, K.L., Mirchandani, P.B., Sheppard, D., 1992. Hierarchical framework for real-time trac control. Transport.
Res. Record 1360, 82±88.
Khoudour, L., Lesort, J.-B., Farges, J.-L, 1991 PRODYN ± three years of trials in the ZELT experimental zone.
Recherche-Transports-Securite (English issue: Special Trac Management), pp. 89±98.
Maher, M.J., 1984. Estimating the turning ¯ows at a junction: a comparison of three models. Trac Eng. Control 25,
19±22.
Mauro, V., Di Taranto, D., 1990. UTOPIA. In: Proceedings of the Sixth IFAC/IFIP/IFORS Symposium on Control
and Communication in Transportation, Paris, France.
Mekky, A., 1979. On estimating turning ¯ows at road junctions. Trac Eng. Control 20, 486±487.
Nihan, N.L., Davis, G.A., 1989. Application of prediction-error minimization and maximum likelihood to estimate
intersection O±D matrices from trac counts. Transport. Sci. 23, 77±90.
Sen, S., Head, K.L., 1997. Controlled optimization of phases at an intersection. Transport. Sci. 31, 5±17.
van Zuylen, H.J., 1979. The estimation of turning ¯ows on a junction. Trac Eng. Control 20, 539±541.