You are on page 1of 23

EUROPEAN

JOURNAL
OF OPERATIONAL
RESEARCH
ELSEVIER

European Joumal of Operational Research 92 (1996) 550-572

New Operations Research and Artificial Intelligence approaches


to traffic engineering problems
M a u r i z i o Bielli a, *, P i e r f r a n c e s c o R e v e r b e r i b
a Istituto di Analisi dei Sistemi ed Informatica, Consiglio Nazionale delle Ricerche, Viale Manzoni 30, O0185 Rome, Italy
b Department of Production, Systems and Computer University of Rome "Tor Vergata", Via della Ricerca Scientifica, O0 133 Rome, Italy

Abstract

The purpose of this paper is to review some of the main Operations Research (OR) and Artificial Intelligence (AI)
research achievements in the field of traffic engineering, with particular reference to road traffic control. Therefore, the
performances of the discussed approaches are illustrated and compared with respect to the features and the requirements
which characterize some of the most relevant problems in the area of interest. The potential for a combined use of typical
OR and AI methods and techniques when solving complex real-world, large-size traffic control problems is also emphasized.
Finally, some concluding remarks are drawn that highlight the new challenges awaiting for the scientific community efforts.

Keywords: Mathematical programming;Simulation; Artificial Intelligence;Traffic engineering;Road traffic control

1. Introduction
The perception of the key role played by the
transport sector in the development of a world economy, together with the significant and continuous
growth of the demand of mobility, have led to the
activation of several research programs (such as
DRIVE and PROMETHEUS in Europe, IVHS in the
United States of America, and RACS, AMTICS,
VICS and SVSS in Japan) that aim at finding new
solutions to the main problems originating from the
widespread use of traffic networks, based on the
application of information technology and telecommunications (Transportation Research Circular,
1993).

* Corresponding author.

Due to the large amount of expected benefits, the


majority of cities are currently developing system
architectures which incorporate functions such as
on-line traffic control, public transport and parking
management, traffic and travel information, demand
management, route guidance and pollution monitoring. Technical advances in computer science offer
traffic engineers an entire set of programming options and control tools to implement the various
strategies required to cope with the broad range of
situations originating from pattern changes in traffic
behaviour. Presently, advanced telematic systems
provide a data-rich, real-time environment which
gives rise to a new challenge awaiting for the scientific community efforts (Bell, 1992).
In this framework, new Operations Research (OR)
models are currently being developed, together with
practical on-line heuristic algorithms for traffic engi-

0377-2217/96/$15.00 1996Elsevier Science B.V. All fights reserved


PII $0377-2217(96)00010-0

M. Bielli, P. Reverberi / European Journal of Operational Research 92 (1996)550-572

neering. If the application of such techniques to


transportation network planning and design is not
new (see, e.g. Florian, 1994), there is a growing
interest in their use in the related areas of monitoring, management and control, in view of improving
the safety and efficiency as well as reducing the
environmental impact of the transportation system
(Papageorgiou, 1991a). At the same time, traffic
engineers have begun to recognize the significant
progress made in the development and validation of
Artificial Intelligence (AI) and knowledge-based approaches to transportation problems, both for the
high flexibility, extensibility and interactivity degrees shown by these methods in solving complex
real-world tasks and for the computational burden
that typical OR optimization procedures often impose on the available hardware (Bielli, Ambrosino
and Boero, 1994).
Scientific research has promoted the application
of OR and A/techniques in the study of any possible
transportation mode operation; for instance, air traffic control problems are addressed by means of
several OR models in Bianco and Bielli (1993),
while AI methods are discussed by Goslin (1987),
Nevertheless, road traffic has been recognized not
only as the transport sector where benefits expected
from the application o f OR methods and A / t e c h niques are potentially t h e most relevant, but also:as
the most challenging sector for assessing feasibility,
effectiveness and usefulness of s u c h approaches
(Ambrosino et al., 1991; Sussman, 1992),
At present, conventional software systems based
on OR models and mathematical programming techniques are largely in use to provide support in a
variety of tasks including traffic analysis, signal
timing; simulation and evaluation:of alternative traffic management strategies (McDonald and Hounsell,
1991). On the other hand, A/techniques can provide
significant contributions in two complementary directions (Ambrosino et al., 1991):
(a) to overcome some of the major limitations of
current traffic control technology and extend the
range of situations the control system is able to deal
with (e.g. traffic congestion, incidents e t c . ) ;
(b) to provide traffic control operators with better
support for managing all the system facilities and
coping with the increasing complexity and flexibility
of available technology.

551

The focus of this paper is on the road traffic


control problem, which requires to select the best set
of controllable inputs (e.g. signal timings) from an
admissible region, so as to achieve a desired process
output (e.g. minimum travel delay). If the task of the
control strategy is performed continuously during
process operation, by using current measurements,
the system performs a real-time (or traffic-responsive) control. In particular, this is feedback (closed
loop) if it makes use of process variable measurements, or feedforward (open loop) if it makes use of
disturbance measurements, both nonpredictable (e.g.
incidents) and predictable (e.g. demand, Origin-Destination patterns). On the other hand, if control
input specification is made before operation, by using historical data, the system performs a fixed-time
control. Public transport and fleet management needs
are often taken into account, by allowing priorities at
signalized intersections.
In face of the huge amount of articles and books
dealing with OR a n d / o r AI applications to traffic
engineering problems, the authors felt to have no
other choice than narrowing the scope of their presentation to a proper subset. Therefore, this survey
does not claim to be exhaustive, although it aims not
to come up against too many relevant omissions. In
any case, these are not intended to down play the
significance of the proposed approaches; rather, it is
to be hoped that they are possibly due to space
limitations and not to the authors' negligence.
Based on the foregoing remarks, this paper is
organized as follows. Section 2 presents some of the
main OR mathematical and computer simulation
models for road: traffic management, with particular
reference to Origin-Destination (O-D) trip matrix
estimation, traffic assignment and prediction. Section
3 analyses the features and the role of the major AI
techniques - namely, Qualitative Reasoning (QR),
Artificial Neural Networks (ANN), constraint programming and :others - in traffic engineering probautomatic incident
most relevant OR
junction, arterial or
discussed in detail.
Then, the rationale for AI Knowledge-Based (KB)
approaches and their combined use with classic optimization procedures is remarked. Finally, in Section
5, some concluding remarks are drawn.

M. Bielli, P. Reverberi / European Journal of Operational Research 92 (1996) 550-572

552

2. Operations Research models and algorithms


Mathematical programming and computer simulation models have received great attention for many
years in the field of road traffic engineering, both in
urban traffic management (see, e.g. Bianco, Bielli
and Speranza, 1993) and freeway traffic control (see,
e.g. Papageorgiou and Schimdt, 1991; Papageorgiou,
1991b; Allsop, 1991a). In particular, this section
reviews the most common approaches to traffic analysis, assignment and forecasting problems.

2.1. Origin-Destination (O-D) trip matrix estimation


Knowledge of O-D flow patterns is essential for
traffic control to reduce congestion. Static O-D estimation methods use traffic counts on a subset of
observed links and additional a priori information
provided by a target O-D matrix in order to set up a
system of equations which in general takes the following form:
Min

Fl( t, t ) + F2( v, ~)

s.t.

v = M(t),

/), t

where t is a target matrix, represented as the column


vector of trips between each O-D pair, ~ is the
vector of observed link flows, t and v are vectors
representing the O-D matrix and link flows to be
estimated respectively, Fl(t,-t) and F2(v, ~) are
generalized distance measurement functions and
M(t) is the assignment map describing the relationship between the predicted link flows and O-D matrix. Common methods include entropy maximization (Van Zuylen and Willumsen, 1980), maximum
likelihood (Spiess, 1987), generalized least squares
(Cascetta, 1984) and Bayesian inference (Maher,
1983) (for a comprehensive review, see Cascetta and
Nguyen, 1988).
In the uncongested network case, a flow-independent, proportional assignment can be performed.
Moreover, when traffic counts and the assignment
proportions are assumed to be error free, the reference model family turns out to be greatly simplified,
as follows:
Min

Fl(

t, i~)

/.',t

s.t.

~ = Pt,

where P is the assignment proportion matrix. Such


models have computational advantages, since they
represent single convex optimization problems.
However, inaccurate estimates are produced in the
general case when their underlying assumptions are
not fulfilled and several methods have been devised
to cope with more realistic practical situations (see,
e.g. the bilevel programming formulation for congested networks given by Yang et al., 1992).
Static methods cannot track time-varying O-D
patterns, as traffic-responsive on-line control systems
would require. Therefore, dynamic estimation methods have been defined which take into account the
short-time variation of traffic measurements, while
providing a unique and bias-free solution for O-D
flows. Thus, such methods avoid the waste of information typical of static methods, that consider only
accumulated traffic counts (Cremer, 1991).
Dynamic O-D estimation procedures are based on
causal relationships between time records of traffic
network entry and exit flows. Therefore, it seems
that they can be applied to small-scale systems (such
as intersections), where these records may be suitably assumed to be correlated. On the other hand, it
is reasonable to believe that the same causal relationships get weaker in large networks, because of various effects (e.g., the impact of signals on traffic
flow/time functional relationship). Hence, the rationale for dynamic estimation procedures will also
diminish and a combination of dynamic and static
methods represents the most promising approach in
such cases (Keller and Ploss, 1987).
Since international research activities aim at
equipping vehicles with communication devices for
route guidance and driver information systems for
traffic control purposes, O-D trip matrix estimation
procedures ought to be proposed in the near future
that make effective use of the additional information
both provided to and gathered by individual cars
(Cremer, 1991).

2.2. Traffic assignment


Traffic assignment models are mathematical programming schemes intended to simulate the interaction between transportation supply (that is, the set of

M. Bielli, P. Reverberi / European Journal of Operational Research 92 (1996) 550-572

facilities and means available to the users of the


network) and travel demand (expressed by the number of users of the network at a given time of the
day), producing as a result a flow pattern on the links
of the network. Travel demand depends on travel
costs on links and paths, which in turn depends on
flow patterns. Therefore, taking the costs as a measure of supply, a mutual interaction is generated
between demand and costs that defines simultaneously the demand level for each O-D pair and the
flow pattern resulting from the optimal distribution
of demand on available paths. An equilibrium is
reached when the flow pattern produces a set of
costs which induce a choice of paths generating the
same flow pattern (Cantarella and Sforza, 1991).
Depending on the behavioral assumptions concerning individual route choice, it is possible to
conceive either a user optimal traffic assignment,
when users attempt to minimize individual travel
cost, or a system optimal traffic assignment, when
users cooperate in minimizing total transportation
cost over the road network (Wie, 1991). Costs are
usually defined as an increasing function of travel
time and expected delay with respect to desired
destination arrival time. According to the Wardrop's
principles, at equilibrium each commuter going from
a given origin to a given destination cannot reduce
his total travel cost by changing departure time
a n d / o r route, given all other commuters' decisions.
Deterministic models, assuming that travelers are
perfectly informed about network attributes and conditions, have to be distinguished from stochastic
models, allowing the more realistic hypotheses of
limited information, perception errors or specific
habit for road users' behaviour. Stochastic approaches rely on discrete choice models which are
built based on the main principles of random utility
theory. Although stochastic models are the most
appealing ones from a theoretical viewpoint, they
a greater computational effort in searching
require
for an equilibrium solution (Cascetta and Nguyen,
1988; Cantarella and Sforza, 1991).
Dynamic traffic assignment models have also been
devised based on mathematical programming, computer simulation or optimal control theory (de Palma,
1991). These models attempt to predict the temporal
evolution of traffic flows on congested networks,
where travel demands and costs vary over time and

553

space, not necessarily seeking an equilibrium condition. Thus, they are concerned with the way travelers
adjust their route and departure time decisions, either
within day or from day to day (Cascetta and
Cantarella, 1991). Adjustment processes that describe traffic pattern dynamics are characterized formally in order to give account of minor changes
affecting constantly traffic networks, at both the
system and the individual traveler level.
In this framework, there is an urgent need for
software programs that may take into account the
potentialities of new information technologies, together with emerging socio-economical requirements
related to energy and the environment, In fact, theoretical and empirical models have been developed in
order to include route guidance and information system impacts on travelers' behaviour in optimal assignment procedures (Bell et al., 1991; Ben-Akiva,
Koutsopoulos and Mukundan, 1992; Mahmassani,
Hu and Jayakrishnan, 1992; Barcelo' and Martin,
1994; Ben-Akiva, de Palma and Kaysi, 1994). Other
models try to evaluate changes in pollution levels
with respect to alternative driving cycles, through
microscopic recording of time and speed in floating
car measurements (Bruno and Improta, 1992).
Due
the great complexity o f the considered
problems, no satisfactory model o f general u s e has
been produced so far. Indeed, it is very difficult and
costly to evaluate dynamic processes, s o that Mahmassani and Chang (1986) prefer to conduct microsimulation procedures where road users make
daily decisions according to some exogenously given
rules. Then, a special purpose traffic simulation prog r ~ gives the resulting congestion levels: and provides commuters with the actual travel times, that are
the most impo~ant decision variables: regulating
travelers' choices relative t o the next day. According
to the so-called bounded-rationality behaviour,
~ v e r s are assUmed to have anindifference band that
includes the:range of acceptable arrival times at their
destination, s0 that they will not modify: their subsequent choices as long as such band is met (Bellei and
Bielli,: 1992):
Computer simulation models a n d tools are well
suited for traffic studies, either when analysing and
understanding traffic behaviour, or when testing and
evaluating alternative management and control
strategies. Several simulation approaches and labora-

554

M. Bielli, P. Reverberi / European Journal of Operational Research 92 (1996) 550-572

tory experiments have been carded out that show


great potentialities in the field of dynamic approaches to behavioral models in transportation, although they should necessarily be calibrated on real
data (Barcelo', 1991a). At a macroscopic level, models of multimodal traffic flow assignment on large
transportation networks (such as EMME-2 and SATURN) are currently used; a mesoscopic approach
(such as CONTRAM) is suitable for medium-size
networks; finally, fully microscopic simulation models (such as SITRA-B, MISSION, AIMSUN) are
necessary to work at a local level, in particular at
intersections (Bielli, 1992).
Macroscopic simulation models consider groupings of vehicles and apply flow relationships to
determine successive traffic states. The main parameters to be calibrated, apart from the physical description of the roadway, concern flow-density relationships, time-slice length, mean vehicle length, occupancy rate and speed thresholds between free flow
and congestion (Barcelo', 1991b). On the other hand,
microscopic simulation models deal with movement
of single vehicles. Generally, they include a car-following and a lane-changing model that govern the
traffic process and are programmed in a modular
framework, allowing the user to modify easily all the
underlying submodels in an interactive way
(Barcelo', 1991a).
For instance, AIMSUN performs microsimulations on a discrete time sample basis by splitting the
reference horizon into short fixed intervals called
'simulation cycles', in which all the elements of the
system (namely: vehicles, traffic signal settings, etc.)
are successively updated. The outputs of AIMSUN
are an animated graphical representation of traffic
network state as well as link flow and occupancy
rate data collected by simulated detectors. Moreover,
a set of statistics are produced about traffic flow,
speed, vehicle delay and number of stops that could
not be obtained in on-street trials (Barcelo', Ferrer
and Montero, 1989). System improvements have
taken to the development of a new version of the
model, called AIMSUN 2, that includes the possibility of modifying signal plans and other parameters
(such as external flows and turning movement proportions at junctions) during the course of a simulation run.

2.3. Traffic forecasting


One of the major concerns in traffic engineering
is the development, implementation and evaluation
of mathematical models providing real-time shortterm forecasts of traffic flows and travel times along
all links of urban and interurban networks. In fact,
traffic management and control centers need accurate
information about traffic parameters and variables in
order to maintain stable traffic patterns and reduce
expected near-future congestions. Three types of
traffic volume predictions can be distinguished, depending on time scale (Lesort, 1991):
(a) long term predictions (a few months), that can
be used to forecast the required capacity of transportation facilities, or else compute fixed-time coordination signal plans;
(b) short term predictions (a few minutes), that
can be used in adaptive regulation systems, where
the implemented coordination plan is successively
selected from a pre-stored package;
(c) very short term predictions (a few seconds to a
minute), that can be used by real time regulation
algorithms, which are highly responsive to traffic
variations.
Several kinds of predictive models based on techniques of time series analysis are used for short term
prediction: these include linear models using
smoothed information (McShane, Lieberman and
Goldblatt, 1976), spectral analysis (Nicholson and
Swarm, 1974), autoregressive integrated moving average (ARIMA) models (Eldor, 1977) and filtering
methods (Baras et al., 1979). All such models are
endogenous, namely, their input is represented by
previous values of the parameter to be predicted. On
the other hand, very short term prediction models use
previous measurements at upstream locations, rather
than at investigated sites. Given that the relationships
between flow volume and travel time cannot be
identified precisely in an urban network, the most
relevant problem becomes to estimate travel time
from upstream to downstream points on the basis of
flow volume information. As Lesort (1991) states,
linear or piecewise linear models are usually built for
this purpose. These models are first fitted to a representative data set and then corrected on-line.
Prediction models are also crucial to the develop-

M. Bielli, P. Reverberi / European Journal of Operational Research 92 (1996) 550-572

ment and application of advanced transport telematics strategies, because they are related to within-day
decisions on mode, route, time and even parking
destination choice. Several prediction models have
been studied and applied for specific purposes, particularly within the DRIVE project, but they generally need to be cross validated (McDonald, 1994).

3. Artificial intelligence methods and techniques


Artificial Intelligence (AI) research field has been
developed with the aim of emulating human problem-solving behaviour in complex real-world tasks.
In recent years, the potentialities of expert systems
and AI techniques in transport applications have
received considerable and increasing attention within
the traffic engineering community (Bonsall and
Kirby, 1986; Bielli, Ambrosino and Boero, 1994).
Essentially AI means knowledge processing. A
knowledge b a s e contains a set of facts and rules
relevant to a specific domain, while the inference
engine provides the methods for using the knowledge base through heuristic, largely nonalgorithmic
procedures.
3.1. Qualitative reasoning and applications

Qualitative Reasoning (QR) represents a viable


approach to tackle real-world traffic problems, where
one is often dealing with uncertain, missing and
erroneous data and has to incorporate knowledge
which is present only as human expertise or heuristics and expressed in non-numerical or symbolic
form. In such cases, it is not possible or even not
necessary to know the exact value of the different
parameters, while plausible confidence intervals for
them are enough in light of proper decisions to be
made. These may be related to analysis, interpretation, diagnosis, prediction, advising, p l ~ n g and
design tasks and, in particular, to da~ completion,
congestion monitoring, traffic assignment and forecasting problems (Wild, 1994).
Qualitative techniques greatly reduce some of the
major disadvantages of quantitative approaches, Actually, they are simpler than physical models based
on the mathematics of continuous variables and differential equations, which characterize quantitative

555

methods. Moreover, they generate causal explanations of observed phenomena based on mechanisms
that can be easily understood by humans, thus providing the basis for common sense modeling (Martin,
Toledo and Moreno, 1994).
A qualitative representation of the fundamental
diagram of traffic theory is needed in order to apply
this method. The diagram has to relate the macroscopic traffic variables -flow, density and speedthrough the definition of a proper 'quantity space',
that identifies the range of traffic states (free flow,
undersaturation, saturation, oversaturation, congestion). QR implies thinking about problems in terms
of the involved concepts and then building the semantic network of interrelationships among variables. As a consequence, analyses can be carried out
based on propagation mechanisms that are required
to satisfy some locally and empirically defined constraints (according to the constraint propagation approach). Some advanced qualitative traffic control
systems have been developed and partially implemented; the most remarkable among these include
AURA and TRYS in Spain (Cuena, Molina and
Martin, 1992; Cuena, 1989).
AURA is an expert system for traffic control on
urban motorways by ramp metering and variable
message signs. Incident and congestion detection, as
well as traffic prediction, are achieved through a QR
approach based on the theoretical method developed
by de Kleer and Brown (1984). Thus, free-flow
traffic dynamics is modeled by the continuity and
dynamic differential equations, that can be transformed into corresponding qualitative relationships
called confluences. These in turn are represented by
sets of rules that generate possible state variable
changes and alternative hypotheses about future traffic states. Since the confluence approach lacks an
explicit representation of time and events, AURA is
suitable for motorways, where continuous flow and
unfrequent stops are prevailing features, but not so
much for urban road networks (Cuena and Molina,
1994).
In order to overcome some of the main limitations
of currently used QR approaches when applied to
urban traffic control, an alternative (k, t)-formalism
that is able to deal with temporal reasoning and
qualitative representation of multidimensional parameters has been developed (Martin, Toledo and

556

M. Bielli, P. Reverberi // European Journal of Operational Research 92 (1996) 550-572

Moreno, 1994). According to such formalism, an


adequate representation of urban traffic behaviour
can be achieved only if traffic density values are
known for each instant and each point in the network. The proposed model consists in a knowledge
representation hierarchical method and an inference
engine capable of managing time-space parameters,
thus extending usual qualitative paradigms dealing
with unidimensional parameters, such as those proposed by de Kleer and Brown (1984), Forbus (1984)
and Kuipers (1986).
3.2. Neural networks and applications
Artificial Neural Networks (ANN) is another attempt to model associative reasoning and pattern
matching typical of human brain. At present, these
networks only model the process that connects input
data with output data by exploiting computer ability
to perform an iterative series of fast numerical computations (Hajek and Hurdal, 1993). Unlike rulebased systems, ANN do not require detailed encoding of causal relationships and existing expertise.
Neural networks are parallel distributed information
processing architectures suited to hardware implementation and real-time operation. They consist of
the following elements (Dougherty, Kirby and Boyle,
1994):
a set of nodes that receive a vector of inputs and
compute an analogue output according to a transfer
function;
- connection links of various strength - measured
by proper weights - joining the different nodes;
- a transfer function and an associated threshold
level for the weighted sum of each node inputs, that
possibly triggers its activation (that is, a switch
between the two node states ' o n / o f f ' ) ;
a set of layers, that gives rise to a topological
arrangement of nodes such that all nodes in adjacent
layers are connected to each other. A neural network
thus has an input layer, an output layer and possibly
one or more hidden (or internal) layers.
One of the most used learning technique to train
ANN is the back-propagation paradigm (Hecht-Nielsen, 1990). It'presents input data to the input layer of
the network and computes an outcome that emerges
from the output layer based on current connection
weights. This output is compared with the one that
was expected for the given input data by means of a
-

global error function. This in tum is 'back-propagated' in the network via a set of partial derivatives
that smoothly update the connection weights, to displace the output towards the desired level. If the
training is successful, the value of the function reduces over time as input data are repeatedly presented. Various methods exist that improve the rate
of convergence, such as 'variable momentum', that
regulate step sizes in successive updating iterations.
Neural networks represent a valuable methodological tool in the field of transportation research, particularly in the areas of traffic pattern recognition,
classification and prediction, congestion and incident
detection, driver route choice modeling (Dougherty,
Kirby and Boyle, 1994). For instance, as far as
traffic distribution is concerned, conventional systems are based on time interval-dependent actions. In
other words, owing to the difficulty of dynamic
traffic assignment modelling, it is generally assumed
that network traffic state (i.e., the traffic pattern) is
static in a given time interval. The Adaptive Resonance Theory (ART) is an ANN model able to
provide more reasonable traffic pattern recognition
results than the most common methods, by allowing
parallel processing and tolerance adjustability (Faghri
and Hua, 1992).
Moreover, ANN have been used in signal timing
control, through a network-based pattern classification and evaluation procedure of monitoring and
control strategies. Back-propagation has been used
also in O-D matrix forecasting, where trip generation
and attraction zones can be assumed as external
input, while the neural network output gives future
O-D distribution. Noise present in past O-D patterns
can be removed through proper training (Hua and
Faghri, 1993).
Although large data sets are needed for ANN
effective training, it can be assumed that road monitoring systems are currently in position to meet this
requirement. However, several points need to be
further investigated i n order to assess the performance of ANN models, among which there are the
extensibility of ANN traffic predictions made at
specific sites (that is heavily connected with implementation costs of ANN systems), as well as the
sensitivity of the achieved results on the representativeness of input data and traffic pattern changes
over time.

M. BieUi, P. Reverberi / European Journal of Operational Research 92 (1996) 550-572

557

3.3. Other methods and applications

3.4. The case o f automatic incident detection

Rule-based programming is the best method developed so far in order to capture expert knowledge
about some specific domain. In particular, production systems, which represent the most common
rule-based approach, consist of:
(a) a knowledge base comprising a set of rules
formed by a precondition and a set of actions that
capture some heuristics of the expert;
(b) one or more data bases that contain information about the domain and the problem-solving state
(the so-called working memory);
(c) a control strategy, that is, a way of controlling
and directing the reasoning steps performed by a rule
interpreter in a 'recognize-act cycle'.
Production systems are normally employed when
relevant knowledge can be given in terms of situation-action pairs. Thus, they are suitable for expressing the heuristic knowledge that is necessary for
interpreting a set of sensor readings or assessing the
effectiveness of a given signal plan (Bielli, 1992).
Model-based reasoning is a deeper form of encoding
and manipulating knowledge about the domain, that
implies the use of models describing the underlying
system structure and the interactions among its components. It can be employed successfully for reasoning on hypothetical consequences of actions or making diagnoses with respect to the observed conditions
(Bobrow, 1985).
Constraint programming is a recent AI paradigm.
It is based on techniques for representing functional
relationships among variables as normative rules that
regulate the propagation of values over the reference
network, according to constraint satisfaction mechanisms (McDermott and Chamiak, 1985). Given some
initial assignment of values to a subset of elements, a
constraint satisfaction mechanism tries to achieve an
assignment of values to all elements that is consistent with the constraints. Depending on the kind of
propagated values, one can have symbolic or numeric constraint propagation. This method can be
suitably used for data completion problems, where
one is asked to derive traffic flows on nonequipped
links and intersections on the basis of the information provided by sensor-detected links. In such a
case, constraints could be inherent to the structure of
the network, acquired traffic data, operating signal
plans and so on (Bielli et al., 1991).

Automatic Incident Detection (AID) has been developed mainly for freeway traffic, due to the expected high safety improvements deriving from incident prevention a n d / o r management. So-called California AID algorithms are among the most widely
used tools that enable to perform such a complex
task (Payne, Helfenbein and Knobel, 1976). However, since congestion is a widespread phenomenon
in the majority of large urban areas, it is highly
desirable to detect early an incident also in such
contexts, in order to take effective actions that aim at
alleviating its direct and induced impacts on traffic
flows.
The effect of an incident is an immediate change
in the quantitative relationship of the macroscopic
traffic variables (volume, density and speed) in the
neighborhood. Hence, as long as the inconveniences
last, a different fundamental traffic diagram holds in
the affected area. Classical AID algorithms are based
on filtering techniques and stochastic process analysis, but also pattern recognition, cluster analysis and
probabilistic approaches have been used for the case.
Although these algorithms are generally conceived to be integrated into traffic control systems,
very simple procedures are employed in practice
which do not reflect the state of the art. Moreover,
the confidence in such methods is relatively low, so
that human control and supervision is still considered
necessary (Busch, 1991). At present, no method has
proved to be clearly superior to the others. Due to
the whole possible specla'um of traffic conditions, a
single algorithm can scarcely find the optimal solution; therefore, a mixture of various algorithms in a
multimodei approach has been proposed and investigated (Morelio and Sala, 1993).
An Incident Management Expert System (IMES)
has been developed for freeways and motorways,
where unexpected congestion may occur even when
surveillance, communication and control systems are
in operation (Chang and Huarng, 1993). It aims at
evaluating off-line control strategies and assisting the
overall decision making process, which includes incident detection, confirmation, prediction, management and response. It operates with AID algorithms,
a production rule editor, a rule-based reasoning
mechanism that perform forward-chaining to formulate responses as advice to users.

M. Bielli, P. Reverberi / European Journal of Operational Research 92 (1996) 550-572

558

The same features are addressed by the set of


real-time expert systems which have been integrated
in a distributed problem-solving blackboard architecture called FRED (Freeway Real-time Expert system
Demonstration) (Ritchie, 1990), that aims at providing an effective support to detecting and verifying
traffic incidents, as well as identifying proper response strategies. Further developments have taken
to a more comprehensive real-time KBES prototype,
that works as a decision support system assisting
traffic engineers in designing and selecting strategies
for variable message signs, highway advisory radio
and adaptive signal control. The system, called
ARTIST (Arterial Real-time Traffic Incident ReSponse Tool), attempts to replicate three stages of
human response to operational problems: verification/classification, action and impact monitoring.
The basic parameters to be taken into account in
making decisions are time of d a y / d a y of week,
traffic phenomena (number of vehicles, arrival rate)
and traffic restraints both in the reference area and in
its surroundings (Deeter and Ritchie, 1993).
As far as urban traffic management and control
are concerned, short term forecasting, congestion
monitoring and incident detection have been recognized as highly interdependent fields. In fact, incidents are usually identified by comparing current
network conditions with expected conditions. These
are derived from traffic predictions formulated on
the basis of data collected by properly located traffic
detectors and the operating signal control system
(McDonald, 1994). Some prototype systems have
been developed in the framework of DRIVE programme (e.g. in projects INVAID I, II and
LLAMD-LEADER). Extensive field tests are planned
in order to assess their performance (Sellam and
Boulmakoul, 1994).

4.

OR

and

AI

approaches

to road

traffic

control

Traffic control has been defined as " a non-capital


intensive technique to promote safe, efficient and
convenient movement of persons and goods, making
a better use of existing roads" (Gartner and Gershwin, 1983). Urban traffic control (UTC) systems
differ in a number of aspects. The major distinctive
features are related to:

1) spatial organization of control measures


(centralized or distributed systems);
2) aggregation level of control measures (isolated
intersection, arterial or network);
3) operation mode (fixed-time, traffic flow actuated, demand-responsive).
First generation control systems choose signal
timing plans among those stored in a library, that are
computed off-line by means of historical traffic data.
A plan can be selected according to the time-of-day,
by direct operator choice or by matching plans to
current traffic conditions. Second generation control
is an on-line strategy that computes and implements
real-time signal plans based on surveillance data and
predicted traffic volumes. Third generation control
was conceived to implement and evaluate a fully
responsive, on-line, traffic control system. Timing
plans are revised in a shorter period than second
generation (every 3-5 minutes) and cycle length
(that is, the minimum time in which a complete
succession of signals occurs) is required a priori to
vary during the control period. Last generation control systems take account of the strict interdependence between the system architecture design (number, kind and location of detectors and computers)
and the implemented control strategy. Even if actuation methods cannot guarantee the best possible performance per se, properly calibrated traffic-actuated
signals can provide considerable advantages over
fixed-time control (Bruno and Improta, 1994).
4.1. OR approaches

Mathematical programming methods for UTC can


be grouped in four classes, corresponding to problems of increasing complexity:
- single junction control;
- arterial coordination and synchronization;
- network coordination and synchronization;
- combined traffic assignment and signal setting.
These models are presented, analysed and discussed
in the following subsections.
4.1.1. Single junction control
The main goals of single junction control can be
listed as follows:
- capacity factor maximization;
- total rate of delay minimization;

M. Bielli, P. Reverberi / European Journal o f Operational Research 92 (1996)550-572

- cycle time minimization.


These can be subject to the following kinds of
(generally linear) constraints:
undersaturation constraints;
- minimum green-time constraints;
maximum red-time constraints;
constraints on minimum and maximum values for
the cycle time;
- compliance constraints between stage cycle and
green times (for stage-based approaches);
- incompatibility crossing constraints (for phasebased approaches).
Stage-based approaches seek optimal green times
after requiring the sequence of stages and the duration of transition periods to be specified in advance
(a stage is that part of a plan where signals do not
change). No direct method is available to constrain
the duration of the effective red and green times for
the groups (which are sets of streams receiving the
same signals from the controller) (Improta, 1991; for
further insights, see Allsop, 1991b).
Such approaches have evolved to allow the computation of optimal timings by using knowledge
about crossing compatibilities among the different
streams. This implies to analyse possible sequences
of compatibility cliques, whose edges correspond to
a set of groups that are mutually compatible. Any
clique is a possible stage, while any ordered set of
cliques that satisfies some imposed constraints is a
possible stage sequence. The optimization problem
consists of calculating the optimal clique sequence,
the green time for each clique and the cycle time
(Stoffers, 1968; Zuzarte Tully, 1977; Zuzarte Tully
and Murchland, 1978). All maximal sequences of
stages which give a single green interval for each
group in each cycle are generated, although no automatic procedure has been developed yet in order to
remove superfluous stages. H e n c e , a considerable
amount o f manual intervention is required to reformulate the constraints when zero-green-time cliques
and corresponding transition periods are deleted
(Heydecker and Dudgeon, 1987).
Phase-based approaches determine optimal cycle
time, green timing and scheduling by directly addressing group signal timings through knowledge of
the incompatibilities among the streams, rather than
through the stage mediation. This allows for constraints such as minimum green times and group-to-

559

group clearance times to be expressed explicitly.


Furthermore, the stage sequence and structure of the
transition periods need not be specified before the
optimization takes place (Heydecker and Dudgeon,
1987; Moller, 1987). Mathematical programming
models have been proposed that optimize control
variables simultaneously. Such variables are represented on a time axis that can be scaled in time units
(Improta and Cantarella, 1984) or in proportion to
the cycle time (Cantarella and Improta, 1988). Binary variables are required for compatibility constraints. Phase-based approaches, however, require a
preliminary decision about the composition of the
streams and their assignment to the lanes; generally,
the flow ratios of the streams have to be assumed
constant during the cycle. SICCO (Single Intersection Capacity factor and Cycle time Optimization)
(Cantarella and Improta, 1988) and SIGSIGN (SIGnal deSIGN) (Silcock, 1990) are among the computer codes that use a phase-based approach.
Signal timing optimization presents both theoretical and computational difficulties. Actually, it is
necessary to select the best sequence of signal
changes and the associated timings. These decisions
are based upon estimates of current queue lengths as
well as on vehicle detector data about the traffic flow
that is arriving/leaving the junction within the next
few seconds. Actually, the state of traffic at a junction is specified by both the number of vehicles
queuing in each of the streams, that is influenced by
the signal control plan, and the number of vehicles
arriving in the next future.
On the other hand, the state of the controller is
specified by the set of green signals, the changes - if
any - which are currently underway, the times at
which they will be completed and the expire times of
minimum or maximum allowed durations. It follows
that the state space to be investigated is huge, since it
is given by the product of all possibilities for each of
these state variables. Several techniques have been
devised to abate the computational complexity of the
signal timing optimization problem. Generally, these
work on a state space reduction achieved through
simplifying assumptions such as those of uniform
vehicular arrivals beyond the detection period, the
restriction of the signal control states to a few stages
and time discretization.
Many researchers approached the single junction-

560

M. Bielli, P. Reverberi / European Journal of Operational Research 92 (1996) 550-572

optimizing traffic responsive control as a Dynamic


Programming (DP) problem (Robertson and Bretherton, 1974; Gartner, 1982; Henry, Farges and Tuffal,
1983). This optimization process divides the horizon
length into N stages, where by horizon is meant any
period for which flows arriving at the stop lines are
known or can be predicted. At each stage the state of
the intersection is defined by the state of the signal
(green or red) and the queue length on each of the
approaches.
The decision variable indicates whether the signal
is to be switched at the current stage or not. The final
objective is to determine the optimal sequence of
switching decisions at all stages of the horizon,
whose typical length is 5-15 minutes. A recursive
optimization function that aims at minimizing the
total delay based on the queuing discharge processes
occurring at the intersection is provided. The main
shortcomings of dynamic programming methods are
that they require future arrival information for the
entire horizon and an extensive computational effort.
Moreover, since DP optimization is carried out backwards, it precludes the opportunity for modifying
control decisions in response to real-time updated
traffic data (Gartner, 1991).
Hence, a procedure that uses only available flow
data has been carried out in which a 'projection
horizon' of k intervals is the period for which traffic
flow information is needed. The procedure calculates
an optimal policy for the entire horizon, but implements it only for the first r < k intervals. Next, the
projection horizon is shifted forward of r intervals
and the process is repeated, thus realizing the socalled 'rolling horizon' optimization (Wagner, 1977).

4.1.2. Arterial system control


Traffic lights tend to group vehicles in platoons
with fairly uniform headways. It is desirable to set
up control actions that maintain a continuous movement of vehicle platoons through successive intersections. Little (1966) proposed a general formulation of
the arterial coordination as a mixed-integer linear
program, taking as objective function a weighted
sum of the bandwidths of all arterials, where each
bandwidth was assumed to be linearly dependent on
the offsets (a bandwidth measures the length of the
platoon that passes a junction and does not have to
stop at the next one due to the traffic ligh0.

The presence of integer variables as well as the


introduction of several redundant variables and constraints generated a lot of troubles in the practical
use of the method. Nevertheless, Litfle's methodology has been implemented in an off-line computer
procedure, named MAXBAND (Little, Kelson and
Gartner, 1981). The procedure uses Webster's theory
(Webster, 1958), stating that total delay at an intersection is minimized by assigning the available cycle
time to the competing traffic streams in proportion to
their saturation ratios. If b (b') denotes the outbound
(inbound) bandwith and k is the target ratio of
inbound-to-outbound bandwith, this can be expressed
as follows:

Max

(b+kb')

s.t.

(1-k)b'>_(1-k)kb,
b=b',

k ~ l,

k = 1.

Note that the traffic engineer may wish to favor one


direction of traffic over the other.
Given splits, queue clearances, bandwidth target
ratio as well as lower and upper bounds for cycle
time, link speeds and speed changes among the
different links, MAXBAND finds cycle time, offsets,
bandwidths, link progression speeds and left-turn
phase patterns in order to maximize the objective
function subject to cycle time, bandwidth-ratio, interference, loop-integer, speed and speed-change constraints. It must be specified that loop-integer constraints are a set of consistency constraints on signal
offsets, requiring that the algebraic sum of relative
offsets associated with the links of any closed path
(loop) has an integer value. Gartner (1972) demonstrated that, for a connected network of v vertices
and l links, only l - v + 1 loop constraints are
linearly independent.
A recent improvement of MAXBAND is the procedure MULTIBAND, in which traffic volume variations along an arterial (due to turn-in and tum-out
traffic) are considered. Thus, it is possible to assign a
different bandwidth to each directional road section
of the arterial, that is weighted with respect to its
contribution to the overall objective function (Gartner
et al., 1991).
Signal timing optimization is generally carried out
by means of a sequential decision process involving
the following steps (Gartner, 1976):

M. Bielli, P. Reverberi / European Journal of Operational Research 92 (1996) 550-572

(a) a master cycle is found depending on the


requirements of the most loaded junctions;
(b) based on the identified master cycle, green
splits are computed for the various junctions;
(c) a set of optimal offsets among signals is
determined.
This procedure cannot ensure a globally optimal
solution, that could be found only through a simultaneous optimization process. However, experience has
shown that sequential procedures provide good solutions for arterial and network signal coordination
problems in practice.
4.1.3. The network synchronization problem
The bandwidth-based methods present some practical advantages. In fact, they use relatively little
input and provide space-time diagrams that represent
a simple graphic tool to evaluate a n d / o r modify the
results obtained. However the bandwidth-based regulation is often inadequate when applied to urban
networks, given that it does not take into account
road users' delay. On the other hand, mathematical
programming models for network synchronization of
traffic control generally assume as ultimate goal the
total rate of delay minimization (usually a non-linear
function that is approximated by a piecewise linear
function). With respect to single junction control,
these methodologies introduce the following additional sets of constraints:
- congruence constraints between green and red
times for each pair of conflicting streams;
- congruence constraints between cycle time, green
and red times for each stream;
loop-integer constraints.
MITROP (Mixed-Integer TRaffic Optimization
Program) represents the only simultaneous network
optimization model of all control variables (cycle
time, green splits and offsets) proposed so far
(Gartner, Little and Gabbay, 1976). It aims at minimizing the total users' delay by using convex piecewise linear delay functions. Nevertheless, the convexity assumption may be violated in real cases;
moreover, the program requires large computer
memory and high running time.
Link performance functions used in most mathematical programming schemes are based on the assumption that the delay on a link depends only on
the offset between signals at upstream and down-

561

stream junctions. This hypothesis is not true in general. It is then necessary to define new formulations,
expressing the delay as a more general function of
network regulation parameters. At present, only
heuristic procedures allow to remove the limitative
assumption conceming the independence of the delay on a link on signal control features of other links.
Among these, the TRANSYT method (Robertson,
1969; Vincent, Mitchell and Robertson, 1980) is
surely the most widely used in real-world applications for fixed-time plan calculations. The method
works on the basis of the following assumptions:
(a) All major network junctions are controlled by
signals or priority rules.
(b) All signals have a common cycle time or a
multiple of it.
(c) Traffic pattern is known and constant.
TRANSYT follows a hill-climbing strategy for
minimizing a performance index (PI) given by a
weighted sum of delay and stops on the network,
expressed as follows:
N

PI = ~ ( W w i d i + K / I O 0 -- kiPi),
i=1

where:
N = Number of links of the network.
W = Overall cost per average passenger car unit
(pcu) of delay.
K = Overall cost per 100 pcu stops.
d i = delay on link i.
wi = Delay weighting on link i.
k~ = Stop weighting on link i.
Pg = Number of stops on link i.
A demand-responsive approach to traffic network
control is represented by SCOOT (Split, Cycle and
Offset Optimization Technique), developed at TRRL
in the United Kingdom, whose success is testified by
over forty installations both in Europe and overseas
(Hunt et al., 1981; McDonald and Hounsell, 1991).
SCOOT control system has flow and occupancy
detectors placed on specific links, that allow arrival
patterns at the downstream signals to be predicted;
these in turn represent the primary input of the signal
setting optimization model (Bell, 1992). It adopts a
traffic model similar to the one used in TRANSYT,
attempting to determine the signal timings that mini-

562

M. Bielli, P. Reverberi / European Journal of Operational Research 92 (1996) 550-572

mize a performance index based on a linear combination of delays and stops. Signal timings are adjusted smoothly based on the latest traffic situation,
thus avoiding abrupt changes in control plans.
Each junction is treated by the green split optimizer independently from other junctions. A few
seconds before each stage changes, it decides whether
the change should be advanced or delayed by a small
amount of time based on the minimization of the
maximum degree of saturation on the approaches to
the junction. The offset optimizer decides whether or
not to alter all scheduled change times for the stages
at a junction, for every cycle. Any such change has
the effect of altering the offsets between that junction and the adjacent ones. This decision is made by
comparing the values of the sum of the PIs on all
adjacent streets for the assessed offsets. The cycle
time optimizer computes the common cycle time for
the region under control every 2.5 or 5 minutes (in
some cases, SCOOT can assign a junction a cycle
time equal to half of the common cycle time, if this
provides better results). The adopted criterion is that
the most heavily loaded junction in the region should
operate at a maximum degree of saturation of about
90%.
Various demand-responsive systems have been
realized or are currently being developed in several
countries; these include, among others, SCATS
(Lowrie, 1982) in Australia, OPAC (Gartner, 1983)
in USA, PRODYN (Henry, Farges and Tuffal, 1983)
in France, UTOPIA (Mauro and Di Taranto, 1989) in
Italy. PRODYN has represented for a long time the
state-of-the-art in real-time traffic control with conventional techniques. Although it has been tested in a
limited number of areas, it resulted in about 16%
improved performance with respect to the fixed-time
plans of TRANSYT (Wild, 1994a). It is a decentralized system based on a decomposition of the road
network into a number of small traffic subareas,
where it is possible to apply a Dynamic Programming approach to signal optimization that aims at
minimizing total delay. For this purpose, a Bayesian
estimation of queue lengths within a time horizon of
70 seconds is used and a signal plan generation is
made in real-time (every 5 seconds). Flexible signal
phase times and acyclic signal settings are additional
features of the system.

4.1.4. Combined traffic assignment and signal optimization


From the pioneering work of Allsop (1974), it has
been widely recognized that vehicle delays at junctions highly depend on signal setting. As a consequence, changing the adopted control strategy does
have a redistributional effect on traffic flows, given
that traveling costs for road users are altered consequently. Conversely, if changes in routes selected by
road users take place, the optimality of operating
signal settings is compromised (van Vuren, 1991).
Based on the foregoing observations, traffic signal
setting models have been defined where user behaviour is explicitly taken into account; these can be
referred to as equilibrium network traffic signal setting models and viewed as network design models
where regulation parameters assume the role of design variables.
When the flow pattern is assigned, link flows as
well as path choice are assumed independent of the
adopted regulation. Under this hypothesis, traffic
signal setting models for single junction as well as
network control can be used to compute regulation
parameters (Cantarella, Improta and Sforza, 1991).
Such models can be formally expressed as
Min
pEP

Z(p, q),

where q is the assigned flow pattern, p is the vector


of control parameters, P is the set of feasible vectors
p and Z ( p , q) is the system performance index. If
this index relates to the total user delay, the previous
expression becomes
Min
p~P

Z( p, q ) = D ( p , q ) = r , di( P, q)ql,

where di(p, q) is the unitary delay on link (or


junction approach) i.
On the other hand, when traffic signal setting is
fixed, the link flow pattem can be found through a
traffic assignment model q = qb(p). According to
Wardrop's first principle (Wardrop, 1952), for which
each driver chooses the path that minimizes his own
generalized travel cost, such model results in the
following convex program:
Min
qEQ

I(p,q),

M. Bielli, P. Reverberi/ EuropeanJournalof OperationalResearch92 (1996)550-572


where Q is the set of feasible flow pattern and
I ( p , q) is the integral travel cost:

I( p, q) = ~ foqici( p, t) dr,
where ci( p, q) is the cost-flow function for link i.
Hence, the travel cost is assumed to be a function of
flow on the different links and vector p of network
control system parameters.
When both signal setting parameters and link
flows are not fixed, it is necessary to perform a
combined evaluation that requires modelling the interaction between flow pattern and signal setting.
This interaction can be seen as a Stackelberg twoplayer game between a leader (maybe a traffic
agency), that chooses a control strategy in order to
optimize a system performance index, and a follower
(in this case road users), that aims at minimizing its
individual travel costs (Fisk, 1984). Formally:
Min Z ( p , q ) ,
p~P,
q~Q.
P,q

Such combined assignment and signal optimization problem can be formulated by means of a
bilevel programming model in which the upper level
relates to the leader's decisions, while the lower
level concerns the follower's decisions in light of the
choices made by the leader. Note that the manager's
criteria for evaluating his decision are based on
public interest concerns, while the individual tripmaker is assumed to select routes for his own exclusive benefit. Signalized network equilibrium is. therefore defined by control variables that minimize the
collective travel cost (system optimum), according to
the Wardrop's second principle for a normative system, together with a traffic flow pattern that guarantees the minimization of user travel cost for each
assigned regulation (user optimum), thus satisfying
Wardrop's first principle for a descriptive system.
The bilevel optimization problem requires to find the
minimum total cost solution among all those which
respect Wardrop's first principle. Therefore, it can be
formalized through
Min{Z(p, q*):[l(p,

q * ) = Min l ( p , q ) ] }

subject to conservation, additivity and non-negativity


constraints on traffic flows, as well as further constraints on regulation parameters.

563

Bilevel programming problems are generally hard


to solve because evaluation of the upper-level objective function requires solving the lower-level optimization problem. Since the lower-level problem is
in effect a nonlinear constraint, the whole problem
turns out to be nonconvex, so that a global optimum
might be difficult to find (Yang et al., 1992). Exact
and approximate algorithms developed so far in order to solve bilevel programming models can be
applied to small or, at best, medium-size problems
with a limited number of variables and constraints,
but they cannot be used to cope with large-scale
network problems, for computational burden limitations (Hansen, Jaumard and Savard, 1992; Bard and
Moore, 1990; Dempe, 1987).
With reference to traffic network control, global
optimization models usually assume green times as
the only signal setting variablesi This represents a
strongly limitative assumption in urban areas, where
signal coordination between adjacent junctions has
the potential for greatly improving system performance and thus plays a fundamental r o l e in the link
travel cost-function definition. The values of all other
variablescan only b e determined through heuristic
procedures that iteratively perform signal setting and
traffic assignment until a fixed convergence criterion
is met. These methods are at present the most viable
approach to bilevel programming problems, both for
their computational features and their closeness to
the actual decision-making process in a S~ackelberg
game. In fact, they dramatically reduce the computational effort needed to achieve a solution; although
a n optimal Solution:: is not guaranteed t o b e found
(Cant~ella, Improta and Sforza, 1991), so t h a t Dicks o n (1981) showed that the total travel cost can
increase its value during the execution of such procedures.
In: order, to compute a mutually consistent flow
pattern and control policy for a set :of coordinated
junctions, Allsop a n d Charlesworth (1977) introduced a heuristic method consisting of the following
steps:
Step 1. Perform an initial assignment.
Step 2. Determine the signal setting which minimizes delay at each junction.
Step 3. Estimate cost-flow relationships, by using
TRANSYT simulation module.

564

M. Bielli, P. Reverberi / European Journal of Operational Research 92 (1996) 550-572

Step 4. Perform an equilibrium assignment, according to TRAFFIC method (Nguyen, 1976).


Step 5. Perform a new signal setting, according to
TRANSYT.
Step 6. Repeat Steps 3 - 5 till a negligible change
in traffic assignment pattern does occur.
Based on the same scheme, Gartner et al. (1980)
proposed two iterative procedures, the first using the
Webster's method and an equilibrium assignment,
the second based on MITROP method and an equilibrium assignment. Further developments of the
same kind of procedure may be found in Cantarella,
Improta and Sforza (1991).
4.2. AI approaches
Mathematical programming models for road network control are generally developed based on the
assumption of traffic flow undersaturation, while
temporary oversaturations do occur frequently in urban networks, especially during peak hours. Moreover, OR methods are often used for off-line computations, while in recent years there has been an
increasing need for methodological tools suitable for
real-time traffic control. Yagar (1991) identifies four
major reasons explaining the failure of conventional
control systems:
high frequency of signal timing changes;
- low accuracy and reliability of traffic predictions;
- inadequacy of traffic performance indices;
inability to deal with occasional events.
The main factors affecting the performance of
conventional traffic control systems and approaches
can be briefly summarized as follows:
1) Congestion. None of the conventional systems
is able to cope with queuing phenomena that expand
locally when these start interacting in different network areas. In such cases, some higher-level, strategic control is necessary, where local problems are
treated simultaneously, various options for the different areas are evaluated, priorities are set up and
remedial control actions are implemented accordingly;
2) Sudden traffic pattern changes. Conventional
methods are not able to respond effectively to unforeseen, significant non-recurrent traffic flow redistributions on the road network (possibly due to accidents, roadworks, etc.), given that they are designed
either to cope with predefined 'average' traffic con-

ditions (fixed-time, traffic actuated systems) or react


to small changes in traffic flows (demand responsive
systems);
3) Traffic data. Conventional signal control methods are able to deal with traffic flow counts measured by roadside detectors, but they cannot take
account of structural changes in travel demand and
users' route choice, as reflected by other kinds of
information sources, like, for instance, Origin-Destination (O-D) trip matrices;
4) System integration. A major drawback of current UTC systems is that, in most cases, signal
control is applied as an isolated control strategy
which has little or no interaction with other traffic
management measures, like variable message signs,
demand management systems, route guidance and
navigation, motorway control, ramp metering, public
transport management, pollution monitoring, etc. In
order to achieve the highest benefits, such systems
should be integrated in a common road transport
environment that makes some new functions possible, by implementing the principles of cooperative
control and cooperative equilibrium (Mauro, Morello
and Wrathall, 1993). Most of the current research
and development in the area of traffic and transportation systems is pointing towards such an integrated
view.
It follows from the foregoing statements that current UTC systems, no matter how sophisticated, are
unable to operate totally unsupervised. Actually, traffic engineers still play a fundamental role in designing and managing the systems in order to cope with
real traffic conditions through day-by-day operations.
By means of various tools and devices (such as TV
cameras, sensor-reading databases or radio links),
UTC operators monitor traffic behaviour, identify
critical events, evaluate the performance of currently
implemented control strategies and adopt remedial
actions when necessary (e.g., by forcing a particular
signal plan). It is this 'man in the loop' paradigm,
that is a prevailing feature of today UTC installations, which motivates the need for the development
of AI and knowledge-based (KB) methods for on-line
traffic control.
Relevant contributions are expected from AI and
KB techniques with respect to different levels of
application to urban traffic control, such as transportation network analysis and planning, integration
of UTC with other traffic management systems and

M. B&lli, P. Reverberi/ European Journal of Operational Research 92 (1996)550-572

UTC system management. Focusing our attention on


the latter point, it is worth to distinguish the following phases:
(a) Traffic data acquisition. A comprehensive
view of current traffic flow volume, composition and
distribution on the network is required to improve
on-line management actions. Some advanced systems, like SCOOT, already include a database (called
ASTRID) connected to signal control. However, this
is meant as a tool for data collection in light of
planning purposes, rather than as a support to on-line
traffic analysis and real-time control decisions. The
major requirements to achieve a complete and reliable information set include:
inference on missing data, to provide plausible
traffic flow values where sensors are not installed
a n d / o r misfunctioning;
estimation of turning movement coefficients at
intersections, to infer the distribution of measured
flows with respect to different directions;
- estimation of the O-D trip matrix relative to the
reference area.
(b) Traffic monitoring. Improved methods are
required for traffic data interpretation. Traffic actuated control strategies classify detected traffic patterns and select proper signal plans based either on a
state space description (vector method) or, more
empirically, on a set of threshold values. However,
this simple approach needs to be extended in several
respects to better account for the complexity and
variability of traffic phenomena. Further improvements are needed in order to analyse congestion over
the dimensions of space and time, thusi understanding
its causes. Critical intersections and other congestion
sources in the network should be identified according to empirical criteria to set u p priorities when
solving traffic problems. Finally, predicting traffic
situations on a short-term basis (5-10 minutes)would
enable to define improved control strategies, capable
of anticipating and regulating network flow distribution. Since it is very difficult to develop quantitative
models for short-term prediction, especially in large
networks, qualitative approaches that aim a t predicting only relevant changes can be seen as an effective
alternative (Cuena, 1989; Martin et al,, 1990).
(c) Supervision. Control quality and effectiveness
depend largely on data collected in phases (a) and
(b). However, 'expert' control requires additional
-

565

capabilities, such as evaluating continuously the performance of currently implemented control strategies, individuating where and when remedial actions
are needed, assessing the relative merit and likely
effects of alternative control actions. AI techniques
enable to deal simultaneously with multiple problems occurring in the traffic network (e.g. congestion
and critical situations at different sites), that may
require to manage conflicting objectives (e.g. favouring traffic in an area vs. restraining access to reduce
environmental pollution). Moreover, they allow at
the same time to make decisions on the basis of a
temporal analysis of traffic behaviour (and not merely
as a reaction to current travel demand) and to cope
with non-recurring congestion due to unforeseen
events.
Hierarchically intelligent traffic control systems
have been proposed for both urban areas and fleeways, based on the same theoretical approach. A
significant reduction in computational complexity is
achieved by logically decomposing the control problem into three easier-to-solve decision-making levels,
arranged in ascending order of required intelligence
(Saridis, 1991). The on-line control level is the lowest level in the hierarchy and corresponds to single
intersections for urban roads or entry ramps for
freeways. At this level, the operating system optimizes signal cycles while taking into account relevant physical constraints. Some of the regulating
parameters are set by the higher coordination level of
the hierarchy, that supervises a group of intersection
or access ramp controllers in order to allow efficient
traffic progressions.
The coordinator receives the traffic assignment
task from a centrally located organization level, that
represents the highest level of the hierarchy. It aims
at improving the overall system performance by
establishing a communication link between human
operators (traffic engineers) and the control system.
This kind o f architectural design allows to adapt
control strategies to the different traffic patterns with
little computational effort, due to its high learning
capabilities.
4.2.1. Knowledge-based architecture for traffic signal control
In the framework of the DRIVE programme, several projects have been launched in order to develop

566

M. Bielli, P. R e v e r b e r i / European J o u r n a l o f Operational Research 92 (1996) 5 5 0 - 5 7 2

and test M-based prototypes to be integrated in


current UTC system operational processes (some of
them are presented in Bielli, Ambrosino and Boero,
1994). The objectives of the projects are to assess
potential benefits of KB methods in several kinds of
traffic management tasks. Generally, the proposed
model builds an 'intelligent layer' on top of conventional UTC systems. Such layer would have to take
real-time traffic flow data from the existing infrastructures via a traffic control computer (TCC) and
provide suggestions about control actions, to be
transmitted back to the roadside receivers through
the TCC.
Most of the prototypes consist in a blackboard
architecture that is made of several knowledge-based
modules (or knowledge sources, KS) associated to
different key traffic control tasks, including data
completion, traffic evaluation, classification and diagnosis, traffic prediction and signal control. A central object database provides a common representation of the traffic network and a data exchange
facility for the various modules. Finally, an action
scheduler controls individual knowledge source operations in response to some changes introduced in
the blackboard-stored data.
The essential features of blackboard problemsolving models are the integration of different
knowledge representation methods and the execution
of procedures in a distributed computing environment, that allows real-time traffic control and management systems to achieve greater efficiency (Bielli, 1992), especially in the field of signal plan
selection (Ambrosino et al., 1991) and incident detection (Ritchie, 1990). In particular, prototype
IUTCS (Intelligent Urban Traffic Control System)
works by reasoning on a symbolic and qualitative
model of the real world that is maintained in the
global working memory and updated according to
the output of some KS and traffic data coming from
roadside sensors (Ambrosino, Bielli and Boero,
1994).
Traffic situation on the network is evaluated
through a quite straightforward rule-based shallow
reasoning system. At a local level, traffic flows on
each link are classified according to the degree of
saturation (or occupancy rate, or) v/C, where v is
the actual flow and C is the capacity of the link,
indicating how near the link is at its maximum

Table 1
Quantity space for traffic flow analysis in the IUTCS test case of
Hamburg city
L i n k load levels:
Flow

v < 20%C
20%C < v < 50%
50%C < v < 80%
80%C < u

Occupancy
or
or
or
or
0.2
0.3
0.4

<
<
<
<
<
<
<

0.2
0.2
0.2
0.2
o r < 0.3
o r < 0.4
or

Load level
very low
low
undersaturation
saturation
saturation
preeongestion
congestion

Intersection load levels:


Capacity factor

Load level

CF > 5
5 > CF > 2
2 > CF > 1.25
1.25 > CF > 1
1 > CF

very low
low
undersaturation
precongestion
congestion

capacity. Analogously, traffic conditions at any intersection are evaluated through a global indicator called
capacity factor (CF), that, for a given signal setting,
measures the maximum common multiplier of traffic
flow values relative to a junction for which undersaturation still holds.
Qualitative indicators for both parameters are defined by setting up a classification of load levels
based on proper correspondences to a quantity space
(Table 1 shows the instantiation of such classification in the IUTCS test case of Hamburg city center).
Traffic condition trends are also defined based on a
three-valued qualitative scale (increasing, decreasing
or steady flows), so that a complete qualitative description of traffic flow on a link at a given time
would have the following form: (low, + ), meaning
'low traffic with increasing volume'.
The Traffic Control (TC) KS of IUTCS models
decision-making processes underlying the selection
of proper control actions in face of dynamically
changing traffic conditions. Actually, it evaluates the
performance of the traffic network under the current
control setting, identifies remedial actions for critical
intersections and assesses the impact of proposed
local changes on adjacent links and intersections.
Control actions are available in the form of a predefined library of Signal Plans (SP) for intersections. A
global performance index based on parameter CF -

M. Bielli, P. Reverberi / European Journal of Operational Research 92 (1996) 550-572

as provided by the Data analysis KS - is continuously evaluated to assess the adequacy of the adopted
control strategy with respect to the current traffic
situation on the network. This evaluation enables
TCKS to assess current and available signal plans at
intersections for identifying a candidate SP list. In
order to handle the combinatorial explosion arising
in the solution search process, some heuristic mechanisms are needed. These consist in compatibility
constraints that embody both general, domain-independent SP selection criteria and specific, site-dependent rules. Examples of site-independent criteria are
shown in Table 2, with reference to a given set of
network links and intersections.
IUTCS environment has been developed by implementing several k n o w l e d g e representation
schemes and inference mechanisms on top of LISP,
including object-oriented, rule-based and constraintbased reasoning. Object-oriented programming has
also been extensively used to implement higher-level
representation functionalities and is supported by
CLOS, the Common LISP Object System (Steele,
1990). IUTCS has been tested in laboratory using
both detectors data collected in the test site of Hammerbrook area (in Hamburg city center) and: MISSION microsimulation environment. Although further experimentation is needed, promising results
have already been obtained in terms of the signal
control system performance (Ambrosino, Bielli and
Boero, 1994).
IUTCS prototype is intended to provide a first
step towards the development of a general model for
the integration of knowledge processing functionalities in the next generation of UTC systems. The
blackboard architecture is open with respect to fur-

567

ther developments. For instance, in a centralized


control approach, different modules can be run in
parallel on different processors, while in a distributed
control architecture the same theoretical approach
could be hierarchically replicated at the intersection,
area and network levels.
M-based traffic-actuated control systems have
also been developed in the framework of the DRIVE
programme, for instance in CLAIRE (Bell, Scemama
and Ibbetson, 1991) and SAPPORO (Wild, 1994)
projects. Such systems aim at establishing a feedback
loop between traffic flows on the network and control actions. Since no exact mathematical relationship
between traffic behaviour and control does exist,
traffic flow simulations are carried out by using
currently observed data and a model of time-dependent traffic evolution. These result in an estimated
flow distribution within a time horizon of 1 up to 10
minutes, on which evaluation of current signal plans
and - if necessary - signal adjustments are performed and compared with the original plans. This
cycle can be repeated several times until a final
decision is made based on a performance index value
and corresponding signal plans for the next time
horizon are switched to the network junctions.

5. Conclusions

Traffic engineering problems have been usually


faced through optimization approaches that require
the development of complex mathematical models.
When dealing with real-world problems, defined on
medium and large-size networks, these models often
impose an excessive computational burden on the

Table 2
Site-independent SP selection rules
Rule 1:

if CFi > 2 in all key junctions


then restrict candidates SPs to those with shorter cycle times

Rule 2:

if CFi < 2 in all key junctions


then restrict candidates SPs to those with longer cycle times

Rule 3:

if CFi < 2 on junction i since t > n- T (T = sampling time; n = 3)


and SPk I(cycle-time(SPk)) = cycle time (current SP)
and direction of coordination (SP) - current direction of coordination
and CFi > 2
then SPk is a candidate
else constrain selection to SPs with longer cycle times.

568

M. Bielli, P. Reverberi / European Journal of Operational Research 92 (1996) 550-572

available hardware. In general, numerical computation techniques are not able to find a solution in
reasonable time except for the single intersection
level. As a consequence, it has been necessary to
devise a set of practical heuristic problem-solving
algorithms that, for their very nature, cannot guarantee to achieve the optimum (as it has been shown for
the case of combined traffic assignment and signal
setting).
A significant reduction in computational complexity can also be obtained by decomposing logically a
general traffic engineering problem into easier-tosolve decision-making levels, arranged in ascending
order of required intelligence. A centrally located
organization level, representing the top of the hierarchically intelligent traffic control system, aims at
improving the overall system performance by establishing a communication link between human operators and the control system. At a local level (namely,
network links and intersections), the control system
would have to deal with elementary problem instances that can be addressed by common optimization approaches.
Moreover, in recurring congested situations, that
cause oversaturation flows to affect some links of the
network, pure quantitative optimization models show
some inconveniences. Hence, traffic engineers still
play a significant role as supervisors of the control
system, since they can decide the implementation of
proper strategies or remedial actions in response of
the observed traffic network conditions, with the aim
of improving the values taken by a set of established
performance criteria.
Qualitative reasoning patterns of AI-based techniques represent a viable approach in dealing with
uncertain, missing and erroneous data. In such cases,
that occur frequently in practice, determining the
optimal value of all the relevant parameters is either
not possible or not necessary in light of the correct
decisions to be made. In fact, qualitative methods
allow to incorporate knowledge which is present
only as human expertise or heuristics and is expressed in non-numerical or symbolic form, thus
overcoming some of the major limitations of conventional mathematical and algorithmic procedures.
However, since AI techniques are not appropriate
for dealing with every subtask comprised in a complex real-world problem, research efforts in the near

future should be directed towards the identification


of a correct combined use of such methods with
conventional OR approaches. This would allow to
exploit the large amount of expected benefits deriving from the possible sinergies between the two
disciplines in different practical applications, as it
has been shown for the case of road traffic control.
The current research trend is to combine computer
simulation and AI tools in hybrid systems that include advanced methodologies and knowledge-based
systems, for the purpose of designing and building a
traffic simulation intelligent front end. This is a kind
of expert system interfacing the user and performing
both symbolic and algorithmic computations. Combined systems provide a range of facilities that include different modelling paradigms (among others,
object-oriented, logic, data-oriented and rule-based
programming) and interactive access to the model
building and experiment running phases.
Enhancements originating from the integration of
such different approaches could be realized effectively in the framework of Decision Support Systems, taking into account the large number of conflic*ang goals and objectives arising from the heterogeneous actors that are affected by the traffic management and control strategies evaluated in the decision-making process.

References
Allsop, R.E. (1974), "Some possibilities for using Iraffic control
to influence trip destination and route choice", in: Proceedings of the Sixth International Symposium on 'Transportation
and Traffic Theory', Elsevier, Amsterdam, 345-374.
Allsop, R.E. (1991a), "Road traffic: An introduction", in: M.
Papageorgiou (ed.), Concise Encyclopedia of Traffic and
Transportation Systems, Pergamon, Oxford, 380-386.
Allsop, R.E. 0991b), "Signal control at individual junctions:
Stage-based approach", in: M. Papageorgiou (ed.), Concise
Encyclopedia of Traffic and Transportation Systems, Pergamon, Oxford, 478-483.
Allsop, R.E., and Charlesworth, J.A. (1977), "Traffic in a signal
conlrolled road network: An example of different signal timings inducing different routing", Traffic Engineering & Control 18, 262-264.
Ambrosino, G., Bielli, M., and Boero, M. (1994), "Artificial
Intelligence approach to road traffic control", in: M. Bielli, G.
Ambrosino and M. Boero (eds.), Artificial Intelligence Appli-

M. Bielli, P. Reoerberi / European Journal o f Operational Research 92 (1996) 550-572


cations to Traffic Engineering, VSP International Science
Publishers.
Ambrosino, G., Bielli, M., Boero, M., and Mastretta M. (1991),
"Expert Systems approach for road traffic control", in: M.
Papageorgiou (ed.), Concise Encyclopedia of Traffic and
Transportation Systems, Pergamon, Oxford, 124-130.
Baras, J.S., Levine, W.S., Dorsey, A.J., and Lin, T.L. (1979),
"Advanced filtering and prediction software for urban traffic
control systems", Transportation Studies Center, University of
Maryland, Baltimore, MD.
Barcelo', J. (1991a), "Simulation of urban traffic: software environments", in: M. Papageorgiou (ed.), Concise Encyclopedia
of Traffic and Transportation Systems, Pergamon, Oxford,
483-491.
Barcelo', J. (1991b), "Traffic management systems", in: M.
Papageorgiou (ed.), Concise Encyclopedia o f Traffic and
Transportation Systems, Pergamon, Oxford, 541-550.
Barcelo', J., and Martin, R. (1994), "Assessment of vehicle
guidance systems and strategies by simulation", in: Proceedings of TRISTAN H International Conference, Capri, Italy,
239-253.
Barcelo', J., Ferrer, J.L., and Montero, L. (1989), " A simulator
generator for microscopic urban traffic simulation',, Technical
Report, Universitat Politecnica de Catalunya.
Bard, J.F., and Moore, J.T. (1990), " A branch and bound algorithm for the bilevel programming problem", SIAM Journal
on Scientific and Statistical Computing 11,281-292.
Bell, M.G.H. (1992), "Future directions in traffic signal control",
Transportation Research A 26/4, 303-313.
Bell, M.G.H., Scemama, G., and Ibbetson, L. (199I), "CLAIRE:
An expert system for congestion management,', in: Proceedings o f the DRIVE Conference, Brussels, 596-614.
Bell, M.G.H., Busch, F., Shield, C.M., and Anderson, J.M. (1992),
"Assignment in the integration of urban traffic control and
dynamic route guidance", in: Proceedings of the Second
International Seminar on 'Urban Traffic Networks', Capri,
Italy.
Bellei, G., and Bielli, M. (1992), "Sensitivity analysis of a
dynamic equilibrium model for route and arrival time choice",
in: Proceedings of the Second lnternational Seminar on 'Urban
Traffic Networks', Capri, Italy.
Ben-Akiva, M., de Palma, A , and Kaysi (1994), "The impact of
predictive information on guidance efficiency:An analytical
approach", in: Proceedings o f TRlSTAN ll International Conference, Capri, Italy, 579-590.
Ben-Akiva, M., Koutsopoulos, H.N., and Mukundan, A. (1992),
" A dynamic traffic model system", in: Proceedings o f the
Second International Seminar on 'Urban Traffic Networks',
Capri, Italy.
Bianco, L., and Bielli, M. (1993), "System aspects and optimization models in air traffic control planning", in: L. Bianco and
A.R. Odoni (eds.), Large Scale Computation and Information
Processing in Air Traffic Control, Springer, Berlin.
Bianco, L., Bielli, M., and Speranza, M.G. (eds.) (1993), Special
Issue on 'Urban Traffic Management', European Journal of
Operational Research 71/2.
Bielli, M. (1992), " A DSS approach to urban traffic

569

management", European Journal of Operational Research


61/3, 106-113.
Bielli, M., Ambrosino, G., and Boero, M. (eds.) (1994), Artificial
Intelligence Applications to Traffic Engineering, VSP International Science Publishers.
Bielli, M., Ambrosino, G., Boero, M., and Mastretta M. (1991),
"Artificial Intelligence techniques for urban traffic control",
Transportation Research A 25/5, 319-325.
Bobrow, D.G. (1985), Qualitative Reasoning About Physical Systems, MIT Press, Cambridge, MA.
Bonsall, P.W., and Kirby, H.K. (1986), "The role of expert
systems in transport", in: P. Bonsall and M. Bell (eds.),
Information Teclmology Applications in Transport, VNU Science Press, Utrecht, 353-382.
Bruno, G., and Improta, G. (1992), "Traffic control under poilurant emission constraints", in: Proceedings of the Second
International Seminar on 'Urban Traffic Network', Capri,
Italy.
Bruno, G., and Improta, G. (1994), "Urban traffic control: current
methodologies", in: M. Bielli, G. Ambrosino and M. Boero
(eds.), Artificial Intelligence Applications to Traffic Engineering, VSP International Science Publishers, 69-93.
Busch, F. (1991), "Incident detection", in: M. Papageorgiou,
(ed.), Concise Encyclopedia of Traffic and Transportation
Systems, Pergamon, Oxford, 219-225.
Cantarella, G.E., and Improta, G. (1988), "Capacity factor and
cycle time optimization: A graph theory approach", Transportation Research B 22, 1-23.
Cantarella, G.E.. and Sforza, A. (1991), "Traffic assignment", in:
M. Papageorgiou (ed.), Concise Encyclopedia of Traffic and
Transportation Systems, Pergamon, Oxford, 513-520.
Cantarella, G.E., lmprota, G., and Sforza, A. (1991), "'Road
network signal setting: Equilibrium conditions", in: M. Papageorgiou (ed.), Concise Encyclopedia of Traffic and Transportation Systems, Pergamon, Oxford, 366-371.
Cascetta, E. (1984), "Estimation of trip matrices from traffic
counts and survey data: A generalised least-squares estimator",
Transportation Research B 18, 289-299.
Cascetta. E., and Cantarella, G.E. (1991), " A day-to-day and
within-day dynamic stochastic assignment model", Transportation Research A 25/5, 277-291.
Cascetta, E., and Nguyen, S. (1988), "A unified framework for
estimating or updating Origin-Destination matrices from traffic Counts", Transportation Research B 22, 437-455.
Chang, E.C., and Huarg, K. (1993), "Freeway Incident Management Expert System Design", Transportation Research Record
No. 1399 on 'Artificial Intelligence', Transportation Research
Board, Washington, DC.
Cremer, M. (1991), "Origin-Destination matrix: Dynamic estimation,, in: M. Papageorgiou (ed.), Concise Encyclopedia of
Traffic and Transportation Systems, Pergamon, Oxford, 310315.
Cuena, J. (1989). "AURA: A second generation expert system for
traffic control in urban motorways", in: Proceedings o f the
Ninth International Workshop on Expert Systems and their
Applications, EC2, Avignon, France.
Cuena, J., and Molina, M. (1994), "Knowledge-based systems for

570

M. Bielli, P. Reverberi / European Journal of Operational Research 92 (1996) 550-572

motorway traffic control", in: M. Bielli, G. Ambrosino and


M. Boero (eds.), Artificial Intelligence Applications to Traffic
Engineering, VSP International Science Publishers, 157-175.
Cuena, J., Molina, M., and Marlin, G. (1992), "An architecture
for knowledge-based traffic management for the Expo-92
Sevilla urban ring", in: Proceedings of the Second OECD
Workshop on "Knowledge-based expert systems in transportation', Montreal, Canada.
de Kleer, J., and Brown, S. (1984), " A qualitative physics based
on confluences", Artificial Intelligence 24, 7-83.
de Palma, A. (1991), "Road networks: Dynamic equilibrium
models", in: M. Papageorgiou (ed.), Concise Encyclopedia of
Traffic and Transportation Systems, Pergamon, Oxford, 372380.
Deeter, D.L., and Ritchie, S.G. (1993), " A prototype real-time
expert system for surface street traffic management and control", in: Proceedings of the Third ASCE Conference, Seattle,
USA.
Dempe, S. (1987), " A simple algorithm for the linear bilevel
programming problem", Optimization 18, 373-385.
Dickson, TJ. (1981), " A note on traffic assignment and signal
timings in a signal controlled road network", Transportation
Research B 15, 267-271.
Dougherty, M.S., Kirby, H.R., and Boyle, R.D. (1994), "Using
neural networks to recognise, predict and model traffic", in:
M. Bielli, G. Ambrosino and M. Boero (eds.), Artificial
Intelligence Applications to Traffic Engineering, VSP International Science Publishers, 233-250.
Eldor, M. (1977), "Demand predictors for computerized freeway
control systems", in: Proceedings of the Seventh International
Symposium on 'Transportation and Traffic Theory', Institute
of Systems Science Research, Kyoto, Japan.
Faghri, A., and Hua, J. (1992), "Evaluation of applications of
neural networks in transportation engineering", Transportation
Research Record No. 1358, Transportation Research Board,
Washington, DC.
Fisk, C.S. (1984), "Game theory and transportation systems
modelling", Transportation Research B 18, 301-318.
Florian, M. (1994), "Recent advances in methods and models in
urban transportation planning", in: Proceedings of TRISTAN
H International Conference, Capri, Italy.
Forbus, K.D. (1984), "Qualitative process theory", Artificial
Intelligence 24, 85-108
Gartner, N.H. (1972), "Constraining relations among offsets in
synchronized networks", Transportation Science 6, 88-93.
Gartner, N.H. (1976), "Area traffic control and network equilibrium", in: M. Florian (ed.), Traffic Equilibrium Methods,
Springer, Berlin, 274-297.
Gartner, N.H. (1982), "Demand-responsive decentralized urban
control. Part 1: Single intersection policies", Report DOTRSPA-DPB-50-81-24, Office of University Research, US Department of Transportation.
Gartner, N.H. (1983), "OPAC: A demand-responsive strategy for
traffic signal control", Transportation Research Record No.
906, Transportation Research Board, Washington, DC.
Gartner, N.H. (1991), "Road traffic control, Demand responsive",

in: M. Papageorgiou (ed.), Concise Encyclopedia of Traffic


and Transportation Systems, Pergamon, Oxford, 386-391.
Gartner, N.H., and Gershwin, S.B. (1983), "Analitycal models for
transportation system management", in: Proceedings of the
AIRO Conference, Guida Editori Napoli, 793-804.
Garlner, N.H., Little, J.D.C., and Gabbay, H. (1976), "Simultaneous optimization of offsets, splits and cycle time", Transportation Research Record No. 596, Transportation Research
Board, Washington, DC, 6-15.
Gartuer, N.H., Assmann, S.F., Lasaga, F., and Hou D.L. (1991),
" A multiband approach to arterial traffic signal optimization",
Transportation Research B 25, 55-74.
Gartaer, N.H., Gershwin, S.B., Little, J.D.C., and Ross, P. (1980),
"Pilot study of computer-based urban traffic management",
Transportation Research B 14, 203-217.
Goslin, G.D. (1987), "Identification of artificial intelligence applications in air traffic control", Transportation Research A
21, 27-38.
Hajek, J.J., and Hurdal, B. (1993), "Comparison of rule-based
and neural network solutions for a structured selection problem", Transportation Research Record No. 1399 on 'Artificial
Intelligence'; Transportation Research Board, Washington, DC.
Hansen, P., Jaumard, B., and Savard, G. (1992), "New branchand-bound rules for linear bilevel programming", SlAM Journal on Scientific and Statistical Computing 5, 1194-1217.
Hecht-Nielsen, R. (1990), Neurocomputing, Addison-Wesley,
Reading, MA.
Henry, J.J., Farges, J.L., and Tuffal, J. (1983), "The PRODYN
real-time traffic algorithm", in: Proceedings of the Fourth
IFAC / IFORS Conference on "Control in Transportation Systems', Baden Baden.
Heydecker, B.G., and Dudgeon, I.W. (1987), "Calculation of
signal settings to minimize delay at a junction", in: N.H.
Gartner and N.H.M. Wilson (eds.), Transportation and Traffic
Theory, Elsevier, Amsterdam, 159-178.
Hua, J., and Faghri, A. (1993), "Dynamic traffic pattern classification using artificial neural networks", Transportation Research Record No. 1399 on 'Artificial Intelligence', Transportation Research Board, Washington, DC.
Hunt, P.G., Robertson, D.I., Bretherton, R.D., and Winton, R.I.
(1981), "SCOOT: A traffic-responsive method of coordinating signals", Report 1014 TRRL, Transport and Road Research Laboratory, Crowthorne, UK.
Kuipers, B.J. (1986), "Qualitative simulation", Artificial Intelligence 29, 289-338.
Improta, G. (1991), "Signal control at individual junctions:
phase-based approach", in: M. Papageorgiou (ed.), Concise
Encyclopedia of Traffic and Transportation Systems, Pergamon, Oxford, 473-478.
Improta, G., and Cantarella, G.E. (1984), "Control system design
for an individual signalized junction", Transportation Research B 18, 147-167.
Keller, H., and Ploss, G. (1987), "Identification of O-D network
flows from counts for urban traffic control", in: Proceedings
of the Tenth International Symposium on 'Transportation and
Traffic Theory', Elsevier, Amsterdam, 267-284.

M. Bielli, P. Reverberi / European Journal of Operational Research 92 (1996) 550-572


Lesort, J.B. (1991), "Prediction of traffic flow", in: M. Papageop
giou (ed.), Concise Encyclopedia of Traffic and Transportation Systems, Pergamon, Oxford, 329-331.
Little, J.D.C. (1966), "The synchronization of traffic signals by
mixed-integer linear programming", Operations Research 14,
568-594.
Little, J.D.C., Kelson, M.D., and Gartner, N.H. (1981),
"'MAXBAND: A program for setting signals on arteries and
triangular networks", Transportation Research Record No.
795, Transportation Research Board, Washington, DC, 40-46.
Lowrie, P.R. (1982), "The Sydney coordinated adaptive traffic
system. Principles, methodology, algorithms", in: Proceedings of the lEE International Conference on 'Road signalling',
London.
Maher, M.J. (1983), "Inferences on trip matrices from observations on link volumes: A Bayesian statistical approach",
Transportation Research B 17, 435-447.
Mahmassani, H., and Chang, G.L. (1986), "Experiments with
departure time choice dynamics of urban commuters", Transportation Research 20, 297-320.
Mahmassani, H., Hu, Y., and Jayakrislman, R. (1992), "Dynamic
traffic assignment and simulation for advanced network informatics", in: Proceedings of the Second International Seminar
on 'Urban Traffic Networks', Capri, Italy.
Martin G, Toledo F., and Moreno S. (1994), "Qualitative simulation of traffic flows for urban traffic control", in: M. Bielli, G.
Ambrosino and M. Boero (eds.), 'Artificial Intelligence Applications to Traffic Engineering', VSP International Science
Publishers.
Martin G, Moreno S., Toledo F., and Rosich F. (1990), "Qualitative simulation for temporal reasoning in urban traffic control",
in: Proceedings of the Tenth International Conference on
Expert Systems and their Applications, EC2, Avignon, France.
Mauro, V., and Di Taranto, C. (1989), "UTOPIA", in: Proceedings of the Sixth IFAC-IFIP-IFORS Conference on 'Control,
Computers, Communications in Transport', Paris, France.
Manro, V., Morello, E., and Wrathall, C. (1993), "Target IRTE
architectures - Theory and practice of the 5T architectural
design", in: Proceedings of the Technical Days on "Advanced
Transport Telematics', Brussels.
McDermott, D., and Chamiak, E. (1985), Introduction to Artificial Intelligence, Addison-Wesley, Reading, MA.
McDonald, M. (1994), "Integrated urban traffic management",
Area 3 Report, AT'r Project EEC, Proceedings of the Technical Days, Brussels.
McDonald, M., and Honnsell, N.B. (1991), "Road traffic control:
TRANSYT and SCOOT", in: M. Papageorgion (ed.), Concise
Encyclopedia of Traffic and Transportation Systems, Pergamon, Oxford, 400-408.
McShane, W.R., Lieberman, E.B., and Goldblatt, R. (1976), "Developing a predictor for highly responsive system-based control", Transportation Research Record No. 596, Transportation Research Board, Washington, DC.
Moiler, K. (1987), "Calculation of optimum fixed-time signal
programs", N.H. Garmer and N.H.M. Wilson (eds.), in:
Transportation and Traffic Theory, Elsevier, Amsterdam,
179-198.

571

Morello, E., and Sala, G. (1993), "Automatic incident detection


in HERMES", in: Proceedings of the Technical Days on
'Advanced Transport Telematics', Brussels.
Ngnyen, S. (1976), " A unified approach to equilibrium methods
for traffic assignment", in: M. Florian (ed.), Traffic Equilibrium Methods, Springer, Berlin, 148-182.
Nicholson, H., and Swarm, C.D. (1974), "The prediction of traffic
flow volumes based on spectral analysis", Transportation
Research 8, 533-538.
Papageorgiou, M. (1991a), "Traffic control modes", in: M. Papageorgiou (ed.), Concise Encyclopedia of Traffic and Transportation Systems, Pergamon, Press, Oxford, 530-531.
Papageorgiou, M. (1991b), "On-ramp control: Coordinated traffic-responsive strategies", in: M. Papageorgiou (ed.), Concise
Encyclopedia of Traffic and Transportation Systems, Pergamon, Oxford, 289-294.
Papageorgiou, M., and Schmidt, G. (1991), "Freeway traffic
modelling", in: M. Papageorgiou (ed.), Concise Encyclopedia
of Traffic and Transportation Systems, Pergamon, Oxford,
162-167.
Payne, H.J., Helfenbein, E.D., and Knobel, H.C. (1976), "Development and testing of incident detection algorithms", Technical Report FHWA-RD-76-20, Federal Highway Administration, Washington DC.
Ritehie. S.G. (1990), " A knowledge-based decision support architecture for advanced traffic management". Transportation Research A 24/1, 27-34.
Robertson, D.I. (1969), "TRANSYT method for area traffic
control", Traffic Engineering & Control 11,276-281.
Robertsou, D.I., and Bretherton, R.D. (1974), "Optimum control
of an intersection for any known sequence of vehicular arrivals", in: Proceedings of the Second IFAC-IF1P-IFORS
Symposium on 'Traffic Control and Transportation Systems',
Monte Carlo, Monaco.
Saridis, G.N. (1991), "Intelligent traffic control systems", in:
Concise Encyclopedia of Traffic and Transportation Systems,
Pergamon, Oxford, 225-229.
Sellam, S., and Boulmakoul, A. (1994), "Intelligent intersection:
Artificial intelligence and computer vision techniques for automatic incident detection", in: M. BieUi, G. Ambrosino and M.
Boero (eds.), "Artificial Intelligence Applications to Traffic
Engineering', VSP International Science Publishers, 189-200.
Silcock, J.P. (1990), "SIGSIGN: A phase-based optimisation
program for individual signal-controlled junctions", Traffic
Engineering & Control 31,291-298.
Spiess, H. (1987), "'A maximum likelihood model for estimating
Origin-Destination matrices", Transportation Research B 21,
395-412.
Steele G. (1990), Common LISP, Digital Press, Belford, CA.
Stoffers, K.E. (1968), "Scheduling of traffic lights: A new approach", Transportation Research 2, 199-234.
Sussman, J.M, (1992), "Intelligent Vehicle Highway Systems A challenge awaits the transportation and OR/MS community", OR~MS Today 19/6, 18-24.
Transportation Research Circular (1993), "Primer on Intelligent
Vehicle Highway Systems", Transportation Research Board
No. 412, National Research Council, Washington DC.

572

M. Bielli, P. Reverberi / European Journal of Operational Research 92 (1996) 550-572

van Vuren, T: (1991), "Signal control and traffic assignment", in:


M. Papageorgiou (ed.), Concise Encyclopedia of Traffic and
Transportation Systems, Pergamon, Oxford, 468-473.
van Zuylen, H., and Willumsen, L.G. (1980), "The most likely
trip matrix estimated from traffic counts", Transportation
Research B 14, 281-293.
Vincent, R,A., Mitchell, A.I., and Robertson, D.I. (1980),
"TRANSYT: Version 8", Transport and Road Research Laboratory Report 888, Crowthorne, UK.
Wagner, H.M. (1977), Principles of Operations Research, Prentice-Hall, Englewood Cliffs, NJ.
Wardrop, J.G. (1952), "Some theoretical aspects of road traffic
research", Proceedings of the Institute of Civil Engineers.
Part H l, 325-378.
Webster, F.V. (1958), "Traffic signal setting", Road Research
Technical Paper 39, HMSO, London.
Wie, B.W. (1991), "Traffic assignment, Dynamic", in: M. Papageorgiou (ed.), Concise Encyclopedia of Traffic and Transportation Systems, Pergamon, Oxford, 521-524.
Wild, B. (1994), "Using artificial intelligence in traffic engineer-

ing - Perspectives and potential applications", in: M. Bielli,


G. Ambrosino and M. Boero (eds.), Artificial Intelligence
Applications to Traffic Engineering, VSP International Science Publishers.
Yagar, S. (1991), "Traffic control system trends", in: M. Papageorgiou (ed.), Concise Encyclopedia of Traffic and Transportation Systems, Pergamon, Oxford, 536-541.
Yang, H., Sasaki, T., Iida, Y., and Asakura, Y. (1992), "Estimation of Origin/Destination matrices from link traffic counts on
congested networks", Transportation Research B 26, 417434.
Zuzarte Tully, I.M. (1977), "Synthesis of sequences for traffic
signal controllers using techniques of the theory of graphs",
Ph.D. Thesis, University of Oxford Engineering Laboratory
Report OUEL.
Zuzarte Tully, I.M., and Murchland, J.D. (1978), "Calculation
and use of the critical cycle time for a single traffic controller",
in: Proceedings of the PTRC Summer Annual Meeting, PTRCP152, 96-112.

You might also like