714 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 11, NO.

3, SEPTEMBER 2010

Distributed Geometric Fuzzy Multiagent Urban

Traffic Signal Control
Balaji Parasumanna Gokulan, Student Member, IEEE, and Dipti Srinivasan, Senior Member, IEEE

Abstract—Rapid urbanization and the growing demand for

faster transportation has led to heavy congestion in road traffic
networks, necessitating the need for traffic-responsive intelligent
signal control systems. The developed signal control system must
be capable of determining the green time that minimizes the
network-wide travel time delay based on limited information of the
environment. This paper adopts a distributed multiagent-based
approach to develop a traffic-responsive signal control system, i.e.,
the geometric fuzzy multiagent system (GFMAS), which is based
on a geometric type-2 fuzzy inference system. GFMAS is capable
of handling the various levels of uncertainty found in the inputs
and rule base of the traffic signal controller. Simulation models
of the agents designed in PARAMICS were tested on virtual road
network replicating a section of the central business district in
Singapore. A comprehensive analysis and comparison was per-
formed against the existing traffic-control algorithms green link Fig. 1. Breakdown of a three-phase cycle at an intersection.
determining (GLIDE) and hierarchical multiagent system (HMS).
The proposed GFMAS signal control outperformed both the ensure a free flow of vehicles with minimum wait time along
benchmarks when tested for typical traffic-flow scenarios. Further a specific direction. The breakdown of a three-phase cycle at
tests show the superior performance of the proposed GFMAS
in handling unplanned and planned incidents and obstructions. an intersection is shown in Fig. 1 to elucidate the terms split,
The promising results demonstrate the efficiency of the proposed phase, cycle length, and right of way.
multiagent architecture and scope for future development. Traffic signal timing optimization or split adjustment by
Index Terms—Geometric type-2 fuzzy set, multiagent system, changing the green time of a phase in coordinated signals is a
PARAMICS, urban traffic signal control. potential solution for alleviating congestion and maximizing the
throughput at the controlled intersection. It also helps in main-
taining the degree of saturation of all the links connected to an
I. I NTRODUCTION
intersection without compromising the safety of vehicles inside
RAFFIC congestion is a major recurring problem faced the road network. Developing a mathematical model of an inter-
T in many countries of the world due to the increased level
of urbanization and the availability of cheaper vehicles. Opti-
section for calculating the green time required for the specific
traffic demand is difficult due to the nonstationary characteris-
mization of traffic signals can alleviate the congestion levels tics of vehicular flow at intersections. Furthermore, the presence
to a certain extent by evenly distributing the delay among all of signals in neighboring intersections causes platoon formation
the vehicles at an intersection, thereby reducing the travel time (pseudorandom behavior), limiting the usage of stochastic con-
of vehicles inside the road network and providing a temporal trol models. In particular, for a large-scale traffic-management
separation of vehicles with right of way in a link. Traffic signals system, it is difficult to predict the change in traffic state due to
control the movement of traffic by adjusting the split of each a specific traffic signal control parameter variation because of
phase assigned in a total cycle time and by modifying the offset. the aforementioned behavior of traffic flow. This necessitates
Split refers to the total time allocated to each phase in a cycle, the use of a distributed control architecture where each inter-
right of way refers to the lanes with green signal during a section is individually controlled by intelligent agents, which
specific phase, and offset is the time lag between the start of autonomously decide the desired signal control policy based
green time for successive intersections, which is required to on local and communicated congestion information from other
agents at intersections connected to the outgoing links. The
Manuscript received December 5, 2008; revised June 18, 2009, general structure of the agent-controlled intersection is shown
September 24, 2009, and March 25, 2010; accepted April 27, 2010. Date of pub- in Fig. 2. The main objective of such a distributed multiagent
lication June 10, 2010; date of current version September 3, 2010. This work architecture is to achieve a coordinated signal control to en-
was supported by the National University of Singapore under the Research
Grant WBS: R-263-000-425-112. The Associate Editor for this paper was sure lower levels of network-wide congestion by predicting
Z. Li. the future demand. For a signal control that is continuously
The authors are with the Department of Electrical and Computer Engineer- controlled by such agents, the problem reduces to infinite-
ing, National University of Singapore, Singapore 117576 (e-mail: g0501086@
nus.edu.sg; [email protected]). horizon approximation. The infinite-horizon approximation
Digital Object Identifier 10.1109/TITS.2010.2050688 would make the distributed architecture infeasible because of
1524-9050/$26.00 © 2010 IEEE
GOKULAN AND SRINIVASAN: DISTRIBUTED GEOMETRIC FUZZY MULTIAGENT URBAN TRAFFIC SIGNAL CONTROL 715

of vehicles at intersections. To overcome the limitations of

vehicle-actuated controllers, adaptive traffic controllers that
determine the green time based on the prediction of traffic using
the present and past information, employing hybrid techniques
such as a fuzzy system, neural networks, or an evolutionary
process, have been experimented with an isolated intersection.
Traffic signal control using fuzzy logic [7]–[9], the Markov de-
cision process [1], [2], and queue theory [10] have also been at-
tempted on an isolated intersection. However, the assumption of
isolated intersection simplifies the network optimization prob-
lem and obliterates the large overhead of coordination required
between neighboring intersections necessary to reduce the
overall delay experienced by vehicles inside the road network.
Fig. 2. Agent-based traffic signal control loop.
Traffic-responsive methods like SCOOT [11], the Sydney
the necessity to maintain a record of the entire history of the cooperative adaptive traffic system (SCATS) [12], and green
traffic states. However, this problem can be solved by assuming link determining (GLIDE) [13] have been applied to road net-
the traffic characteristics to be Markovian [1], [2], where the works with a large number of interconnected intersections and
predictions are based on the data available at that time instant. are used in many countries as the primary traffic-management
In this paper, a multiagent distributed architecture signal system. They have a centralized architecture where decisions
control system based on geometric type-2 fuzzy sets utilizing concerning the green time, offset, and cycle time are centrally
Markovian properties is proposed. This paper is organized made in traffic control centers, based on the data collected from
into nine sections. Section II discusses some of the related the inductive loop detectors placed near the stop line of each
research work done in traffic signal control design, followed by intersection. Their main drawback is the amount of information
Section III, which describes the main problems associated with to be communicated to the central server and the data-mining
traffic signal control. Section IV gives a brief background of operations required in extracting relevant information to calcu-
type-2 fuzzy sets. Section V describes in detail the proposed late the degree of saturation for each lane [14]. The increase
geometric fuzzy multiagent system (GFMAS) signal-control in the size of the network further exacerbates the problem. In
scheme. Section VI describes the performance measures used [14], a hierarchical multiagent system (HMS) was introduced
for evaluating the controller integrity and robustness, followed to reduce the communication overhead by providing more
by Section VII, which provides the information on benchmarks decision-making capability to the intersection agents whose
and the simulation setup. Section VIII presents the simulation functions were monitored by regional agents and have been
results, and Section IX summarizes this paper. explained in detail in Section VII-C. The increase of efficiency
in performance when using a distributed architecture to large-
scale traffic networks has been demonstrated in [15]–[17]. A
II. R ELATED W ORKS
few other traffic signal controls designed and applied for large
Early traffic signal-control schemes were typically designed networks have been presented in [9], [16], and [17] but were
for isolated intersections, as these form the basic components based on a simple type-1 fuzzy inference system.
of a road traffic network and can easily be modeled. One of the
first mathematical models for calculating the green time with an
III. T RAFFIC -C ONTROL P ROBLEM
objective to reduce the average delay experienced by vehicles
inside a road network was proposed in [3]. The green time The problem of reducing the overall mean delay and in-
was calculated offline based on historical traffic flow pattern; creasing the mean speed of vehicles inside the road network
however, the design was not capable of handling any sudden can be divided into subproblems of finding the green time
variations in the traffic pattern. Furthermore, offline estimation required at each intersection that produces a lower delay and
methods were also prone to losses when switching between a higher speed. These are comparative values, and therefore,
signal plans, particularly when rapid traffic changes occurred. no specific optimal values are available. The green time for
These limitations were overcome by real-time vehicle-actuated each intersection needs to be calculated based on the current
methods like MOVA [4], LHOVRA [5], and SOS [6], where demand and the future predicted value. The calculated green
the extension or termination of the green signal of a phase time produces a lower delay if the prediction is based on
was decided based on a refinement gap-extension process, data sampled at a high frequency, as a greater amount of data
mathematical optimization, or relative delays of traffic flows. samples is available to build the appropriate traffic model.
Vehicle-actuated methods perform better if the loop detectors In a large-scale traffic network, this poses a serious problem
that collect the occupancy and vehicle detection are placed with the amount of overhead required in communicating the
in appropriate positions near the intersections. Increasing the information to a decision-making system. This necessitates the
number of detectors increases the accuracy of vehicle detection. use of distributed localized decision-making units capable of
As most of the vehicle-actuated methods use a specified time calculating the green time required for a specified intersection.
block extension of green time on detecting the presence of A distributed multiagent system provides such a solution that
a vehicle, even a sparse traffic flow can increase the delay suits the need of the traffic signal control. Agent-based control
716 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 11, NO. 3, SEPTEMBER 2010

has widely been used in many applications [18]–[20]. Some IV. F UZZY S ETS
of the agent-based traffic control systems are described in
Brief introductory background information regarding the
[21]–[24]. However, in most of these papers, the drivers were
type-2 fuzzy sets, geometric fuzzy sets, and the defuzzification
modeled as agents rather than the infrastructure.
procedure is provided in this section.
The major limitation is the limited sensing capability of the
sensors attached, restricting the view of the distributed system.
Furthermore, any damage to the local sensors affects the action A. Type-2 Fuzzy Sets
of the localized controller and, by extension, the entire network
performance. However, the use of a multiagent system ensures Type-1 fuzzy sets are the extension of the concept of crisp
robustness by providing redundancy and graceful degradation sets, where each element of a set or the logical assertion is
in the case of single-point failure. allowed to be in an intermediate region between truth and
Traffic signal control has uncertainty associated with it at falsehood. A type-1 fuzzy set maps the domain of the possible
various levels of control. The traffic data like flow, occupancy, values to a corresponding measure of possibility, which is also
and queue collected from the loop detectors placed near the called membership grade, i.e.,
intersections have nonstationary noise associated with them.
The level of noise varies in accordance with various factors A = {(x, µA (x)) |∀x ∈ X} (1)
like the environmental condition, the vehicle length, the driver
behavior (clearance zone problem) [25], etc. Furthermore, the where 0 ≤ µA (x) ≤ 1, and µA (x) is the membership grade
traffic count data are inaccurate due to the vehicles terminating associated with values of x. The type-1 fuzzy set defines the
between the upstream and downstream detectors and can be partial truth of input in various clustered regions in the primary
assumed as the addition of noise to the sensor data. On the domain.
communication side, the uncertainties present include the loss The type-2 fuzzy system proposed by Zadeh [29] and devel-
of data due to mechanical failure of connections, speed of oped by Karnik et al. [30] is capable of handling uncertainties
transmission, and additive noise. The traffic flow by itself is associated with the input by assigning a range of partial truth
stochastic in nature at the point of entry into the network, which, values rather than a crisp value. The type-2 fuzzy system is
however, becomes pseudorandom due to platoon formation an extension of type-1 fuzzy sets (1) and can be represented
when controlled by signals. This poses a challenge to the design by Ã, which is characterized by a type-2 membership function
of a signal control for a network with connected intersections, µÃ (x, u), where x ∈ X, and u ∈ Jx ⊆ [0, 1], i.e.,
where the traffic state of the intersection controlled is influenced
by the traffic state in adjacent junctions. à = {((x, u), µÃ (x, u)) |∀x ∈ X, ∀u ∈ Jx ⊆ [0, 1]} (2)
Deriving a mathematical model that emulates all of the
uncertainties associated with this problem is very difficult and is where 0 ≤ µÃ (x, u) ≤ 1, X is the domain of the fuzzy set Ã,
likely to be inaccurate [14], [26]. Computational-intelligence- and Jx is the domain of the secondary membership function
based methods like neural networks [26] can be used for at x. The block diagram of a type-2 fuzzy system is shown in
signal control design. However, a large training data set that Fig. 3. The major limitation of the generalized type-2 fuzzy
encompasses all the uncertain data is difficult to obtain for sets is the computational complexity [30] associated with the
the proper training of neural networks. Since traffic flow is defuzzification process to obtain a crisp output from the type-2
usually controlled based on rules, a rule-based inference system fuzzy consequent sets. This limitation has been overcome by
capable of handling all the uncertainties associated with the an interval type-2 fuzzy set [31] that has shown improved
inputs and outputs would be more appropriate for the design of performance without the burden of increased computational
a suitable signal-control system. A fuzzy system provides such cost associated with general type-2 fuzzy sets.
a solution without the necessity to derive a proper mathematical In interval type-2 fuzzy sets, the secondary membership
model or define the exact relationship between input and output grade is fixed to be unity, and the uncertainties associated
parameters. However, the type-1 fuzzy set assigns a crisp value with inputs are represented by the union of all points between
to the input at the point of fuzzification discounting the input the lower and upper primary membership grades, which is
uncertainty. termed footprint of uncertainty. Type reduction is the process
A type-2 fuzzy set, where each input is modeled as a type-1 of converting the type-2 fuzzy consequents into type-1 fuzzy
fuzzy set, preserves the fuzziness and provides better approxi- consequents and is usually performed utilizing the embedded
mation than type-1 fuzzy sets. Type-2 fuzzy inference systems sets. The number of embedded sets available within the up-
have been shown to perform better than type-1 fuzzy systems in per and lower boundaries is innumerable, and computing the
situations of nonstationary noisy input, random data-generation centroid of each of these embedded sets is computationally
mechanism, and rule-based control in [27] and [28]. This makes intractable. In [27] and [32], different type-reduction methods
the type-2 fuzzy inference system more suitable for traffic based on the iterative method to compute switching points and
signal control. The limitations are the computational require- uncertainty bounds have been proposed. In [33], a geomet-
ments that can be overcome by using a geometric fuzzy system. ric type-reduction method with reduced computational com-
Therefore, a distributed agent architecture with a geometric plexity and reduced sorting than in [27] and [32] have been
type-2 fuzzy decision system—GFMAS—is proposed in this proposed and is better suitable for the Mamdani-type fuzzy
paper. systems.
GOKULAN AND SRINIVASAN: DISTRIBUTED GEOMETRIC FUZZY MULTIAGENT URBAN TRAFFIC SIGNAL CONTROL 717

Fig. 3. Structure of a type-2 fuzzy system with a single input.

V. M ULTIAGENT S YSTEM A RCHITECTURE

The overall structure of the proposed multiagent architecture
is shown in Fig. 5. It shows nine distributed agents communi-
cating with their immediate neighbors to perform the desired
action. The proposed agent system does not have a specific
overall agent-management structure as all the agents are ho-
mogenous and have equal decision-making capabilities. The
structure of each individual agent is shown in Fig. 6. Each agent
is composed of five modules concurrently operating. The agent
receives data from the sensors attached to the intersection asso-
ciated with it and directly calculates the desired green time for a
phase based on the averaged flow rate, queue count collected by
Fig. 4. Ordered-coordinate geometric consequent set showing two of the the data-collection module, and neighboring intersection status
closed polygons.
collected by the communication module.

B. Geometric Fuzzy Sets and Defuzzification

A. Input Data
A geometric type-2 fuzzy set can be defined as a collec-
tion of polygons in 3-D space with the edges forming the The flow rate of vehicles in a link is calculated based on the
triangle [33]. In geometric defuzzification, the type reducer difference between the upstream arrival flow rate and the down-
and the defuzzifier are combined into a single block, thereby stream departure flow rate for each lane in the link. The data
reducing the computational requirements associated with the are sampled at the frequency of the time step in the simulation.
type reducer and directly providing the crisp output from the Time step refers to the number of discrete events simulated
type-2 fuzzy consequent sets. The centroid of the geometric per second. By varying the time step value, sampling can be
type-2 fuzzy consequent can be calculated as the center of performed in a coarse or fine manner, as desired. Increasing the
the geometric shape of the final consequent set obtained by time step would provide more data that can be used to construct
using the Bentley–Ottman plane sweep algorithm [34] after a realistic traffic-arrival distribution and increase the possibility
the calculation of the firing levels of all the rules. The plane of noise compensation by averaging. Mathematically, the flow
sweep can be performed by ordering the discretized values of value calculated can be expressed as
the coordinate points in the consequent type-2 fuzzy set, and
cg
/∇T 
n l 
the centroid (3) can be calculated by finding the closed poly-  up
fij down
− fij

gon formed by the edges of the ordered coordinates with the k=1 i=1 j=1
origin, i.e., fph = (4)
n.l.(cg /∇T )
n−1
where cg is the current green time for the phase, n is the number

(xi + xi+1 )(xi yi+1 − xi+1 yi )
centroid = i=0
. (3) of links having right of way during the phase, l is the number
n−1 up
 of lanes in the link, ∇T is the sampling period, and fij and
3 (xi yi+1 − xi+1 yi ) down
i=0 fij are the upstream and downstream vehicle flow rates,
respectively. The queue data are calculated as the number of
Fig. 4 shows a consequent type-2 fuzzy set where the lower vehicles queuing in lanes that had the right of way during the
and upper coordinates are arranged in a specified order. The phase that just ended. The queue value gives only the remaining
two different shaded regions indicate the triangle formed by vehicle count and does not take into consideration the vehicle
using two adjacent points on the consequent with the origin to count added after the end of a phase. As it is essential to
construct the final closed polygon. calculate the green time based on the maximum congested lane
718 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 11, NO. 3, SEPTEMBER 2010

Fig. 5. Overall agent architecture—nine coordinating distributed agents.

B. Input Fuzzifier
The collected averaged data are passed on to the geometric
type-2 inference system for calculating the green time required
for the phase during the next cycle period. A geometric infer-
ence system has the same functional blocks as those of a type-2
fuzzy system, except that the type reducer and the defuzzifier
are merged into a single block called a geometric defuzzifier.
The initial step in the inference system is the fuzzification of
the inputs, in which a measure of possibility is given to each
input. Since the fuzzy set used is an interval type-2 fuzzy set,
the inputs are assigned upper and lower membership grades,
and the secondary membership grade associated with each
primary membership grade is assigned a value of unity. The
union of all the membership grades between the upper and
Fig. 6. GFMAS signal control—individual agent architecture.
lower bounds would give the footprint of uncertainty of the
input in the classified region. In this paper, the membership
in a link, the maximum value of queue formed is used as the functions were designed in trapezoidal shape and divided into
input data. The following equation shows the mathematical three regions. The lower and upper bounds in each region
expression for obtaining the maximum queue from the links: was decided based on the maximum and minimum flow rates
experienced at an intersection during the specified time period
 up
Q = max{ q11 − q11 down
 prev  up
+ q11 down
, q12 − q12
 of weekdays. Therefore, each input is associated with a range
of membership grades rather than a crisp point, thus retaining
prev  up down
 prev
+ q12 , . . . . . . q21 − q21 + q21 , . . . . . .} (5) the fuzziness. In a similar manner, the queue count is designed
as a type-2 fuzzy set, and each input is assigned a boundary
where Q is the maximum queue value at an intersection, and membership grade. The final input is the data received from
up down prev the communication module, which is a type-1 fuzzy set. The
qij , qij , and qij represent the upstream traffic count,
downstream traffic count, and the previous stored queue for the fuzzified inputs are shown in Fig. 7.
specific lane in a link, respectively.
It is assumed that the arrival rate of the vehicles can be
C. Coordination via Communication
approximated to a uniform distribution, as the average value
is calculated from data sampled at a high frequency and can The communication module is responsible for receiving the
adequately compensate the queue increase after the end of each information regarding the traffic status of the intersections con-
phase. nected to the outgoing links. Communication is asynchronous,
GOKULAN AND SRINIVASAN: DISTRIBUTED GEOMETRIC FUZZY MULTIAGENT URBAN TRAFFIC SIGNAL CONTROL 719

The module calculates the percentage of green time with

respect to the maximum permissible green time. Green time can
serve as a direct indication of the traffic state at an intersection
if the value is dynamically calculated based on the traffic data
and the traffic flowing in the outgoing links. However, the
relationship becomes highly nonlinear due to the influence of
traffic in the incoming links and platoon formation. The level of
uncertainty associated with green timing in a phase can easily
be handled by using type-1 fuzzy sets as the only source of
uncertainty is the communication noise.
The green timing calculated solely based on the local traffic
data would cause increased congestion levels in the outgoing
links. This can cause already-congested neighboring intersec-
tions to experience even higher inflow of vehicles and extension
of queue in the link beyond the incoming intersection, resulting
in deadlocks. Therefore, coordination based on communicated
congestion data becomes essential during the periods of heavy
and medium traffic flow. A reduced value of green time is allo-
cated to a specific phase if the intersection connected to the out-
going link is already congested. This prevents queue spillback
and deadlock formation. This coordination is achieved based
on the rule base information and highly congested neighbor
communicated data.
Since the communicated congestion data are similar to the
consequent green time, except for their representation in per-
centage, as both have a similar classification. The membership
functions are designed by dividing into three equal regions
with the overlap calculated based on Webster’s equation. As
the coordination is with respect to the maximally congested
neighbor, the maximum value of the communicated congestion
data from the neighboring intersections is used as input. The
communication module, in short, performs the functions of data
Fig. 7. Membership grades of antecedents. (a) Traffic flow. (b) Queue and
consequents. (c) Green time. reception, transmission of congestion status to neighbors, and
data-mining operation.

D. Inference System
The inference engine is the core of a multiagent system.
The inputs and outputs to the type-2 fuzzy inference engine
are shown in Fig. 9. The inference engine calculates the lower
and upper threshold of firing levels for each rule in the rule
base. A total of 27 rules were created based on the three inputs,
as shown in Fig. 10. When the communicated congestion data
indicate low traffic congestion in the neighboring intersections,
a greedy signal policy is used to calculate the green time with
Fig. 8. Road traffic network with six intersections and congestion data com-
munication between the agents. minimal cooperation. For adjacent intersections with medium
and high traffic-congestion levels, the signal policy is suitably
modified to produce a smaller duration of green time than that
and the data received from the neighbors are stored in the buffer with a greedy signal policy. This ensures smaller inflow of
memory. Any standard agent communication protocol like the vehicles into the already-congested intersection, enabling the
Foundation for Intelligent Physical Agents protocol can be intersection to clear its traffic faster.
adopted to perform this function. In this paper, simple broadcast The structure of the rule base is as shown below.
is used. Fig. 8 illustrates the communication information in a
six-intersection road network. The arrows indicate the flow of “If Flow input is in Low region and Queue input is in
congestion information between neighboring agents. It can be High region and Communicated data is in Low region
seen that the link connecting intersection with agent5 to agent6 then Green time is in High region”
is a one-way link. Therefore, traffic-congestion information at The “AND” operation is performed by using the t-norm (min-
the sixth intersection is not passed to agent5. imum function). Since the geometric inference system used is a
720 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 11, NO. 3, SEPTEMBER 2010

Fig. 9. Geometric fuzzy inference system.

Fig. 10. Rule base for a fuzzy inference engine.

type-2 fuzzy system, the lower and upper firing threshold values
for the consequent set for each rule are calculated using the
following equations, respectively:
Fig. 11. (a) Clipped consequent set. (b) Consequent set with ordered coordi-
nates after using a plane sweeping algorithm.
    
f l (x) = sup µq (x1) ∗ µF l (x1)
x 1
x1∈q x2∈f l x3∈cd
    region and in descending order for the lower membership func-
∗ µf l (x2) ∗ µF l (x2) ∗ µcd (x3) ∗ µF l (x3) /x (6) tions. The coordinates are obtained based on the plane sweeping
 
2
 
3 algorithm that gives the nonoverlapping edges of the trapezoid
 and the point of intersection of different regions of green time.
f¯l (x) = sup µ̄q (x1) ∗ µ̄F1l (x1)
x Fig. 11 shows the process of arranging the coordinate points
x1∈q x2∈f l x3∈cd and also removing overlap regions.
    Once sorted, triangles can be constructed by using two adja-
∗ µ̄f l (x2) ∗ µ̄F2l (x2) ∗ µ̄cd (x3) ∗ µ̄F3l (x3) /x (7)
cent points on the consequent set and the origin. The resultant
average value of the centroid of all the triangles constructed on
where µF l (x1), µF l (x2), and µF l (x3) are the lower member- the consequent set meets the geometric center of gravity of the
1 2 3
ship functions, and µ̄F1l (x1), µ̄F2l (x2), and µ̄F3l (x3) are the output type-2 fuzzy set which is the closed polygon formed by
upper membership functions of antecedent inputs for the lth connecting all edges to origin.
rule. [(µq (x1), µf l (x2), µcd (x3)), (µ̄q (x1), µ̄f l (x2), µ̄cd (x3))] For better accuracy of the calculated value of the centroid,
are the lower and upper membership functions of the input it is essential to discretize the consequent set into a number of
queue, flow, and communicated congestion data, respectively. points so that a greater number of overlapping polygons can be
The supremum value is attained when the terms inside the used to average the centroid value, but the tradeoff would be the
bracket attain the least upper bound value. The output type- computational cost associated with the process.
2 fuzzy set is obtained by the process of clipping the output The signal control decisions are based on traffic in the incom-
consequent set regions in accordance to the firing levels. ing and outgoing links but not on the traffic in lanes without
To obtain the crisp value of output green time from the type- right of way. Since the green time for each phase is calculated
2 consequent set obtained above, a geometric defuzzification online, it is difficult to consider the competing phase timings.
process was used. The process utilizes the geometrical shape However, the maximum green time limitation imposed on
of the output set to find the centroid and requires arranging all each phase minimizes the possibility of indiscriminate increase
the coordinate points in ascending order for the upper boundary in green time for a phase. The signal control also begins to
GOKULAN AND SRINIVASAN: DISTRIBUTED GEOMETRIC FUZZY MULTIAGENT URBAN TRAFFIC SIGNAL CONTROL 721

allocate more time to a phase in a recursive manner once a the memory, and an average value of the time delay TAD is
queue buildup is detected. It can be said that the traffic at lanes calculated as
without right of way is taken care of with a delay of one cycle n

period. TD
i=1
TAD = (9)
Tv
E. Signal Plan Implementation
where n is the number of intersections in the road network, TD
The green time decided by the inference engine for the is the time delay experienced by vehicles at each intersection,
specified phase in a cycle is applied to the traffic signal through and Tv is the total number of vehicles that entered and left the
the action implementation module. The module serves as an network during the measurement period. The delay parameter
interface between the hardware and the inference engine. The has widely been adopted for characterizing traffic signal-control
action implementation module also serves to verify whether the schemes. In [37], it has been proved that the average queue
timing calculated is within the permissible range (minimum of size at any intersection is directly proportional to the average
5 s and maximum of 60 s) for the phase. In case of delay in delay experienced by a vehicle inside the network, which makes
the availability of signal timings from the inference engine, it the average time delay parameter a suitable entity to classify
continues to use the previously used signal split values. This is the congestion level at an intersection. Another important work
the main advantage of using a distributed modular architecture that showed the linear relationship of the detector occupancy
that prevents failures due to communication delays. This mod- with the average delay was [4]. Moreover, the Highway Ca-
ule also communicates with the backup system regarding the pacity Manual (HCM2000) [35] uses the average control delay
signal timing in progress. incurred by vehicles at the intersection to classify the level of
service offered by each signal control. All of these aforemen-
tioned works justify the suitability of using the average delay
F. Backup System
parameter to evaluate the performance of traffic signal control.
In case of a failure to continuously receive the inputs for ten The current vehicle mean speed is the other parameter used
consecutive cycles or if the action policy remains unchanged for the evaluation of the performance of the proposed architec-
for 20 consecutive cycles, the backup system takes over the ture. It is essential to use the speed parameter along with the
operation of the controller and sets the signal in pretimed mode. delay parameter to avoid errors caused due to inaccuracy in the
The value of the green timing for the phase is calculated using calculated travel delay value and the number of vehicles inside
Webster’s signal calculation method in [28] and [29] using the the network [26], [38], [39]. The current vehicle mean speed is
last collected nonzero input. The optimal value of cycle time inversely proportional to the delay. These two parameters reflect
required at an intersection can be calculated as the overall traffic condition in the road network and have been
adopted in this paper.
C = (1.5 ∗ L + 5)/(1 − Y ) (8)
VII. S IMULATION S ETUP AND P LATFORM
where C is the cycle time, L is the lost time, and Y is the
The multiagent system presented in the previous section
summation of the ratio of flow to saturation flow in each arm
was tested on a simulated real-world scenario and an extreme-
of the intersection. For the detailed procedure on how to obtain
stress scenario to effectively evaluate its performance in a
the lost time and the calculation of split timing, see [36].
large interconnected network. This section describes in detail
The backup system returns control to the agent once changes
the simulation environment chosen and the platform used to
in the decision value are detected. This ensures smooth flow of
conduct the tests.
traffic to the maximum possible extent, except during network
saturation or full-capacity flow period, without affecting the
network. A. Traffic Network Design
The traffic network used to evaluate the performance of the
proposed GFMAS signal control was a section of the central
VI. S YSTEM E VALUATION AND P ERFORMANCE M EASURES
business district in Singapore (see Fig. 12). The network was
The performance of the proposed GFMAS signal control is modeled using version 6.0 of the PARAMICS traffic simulator.
evaluated based on two measures: 1) the average mean delay of The signal phasing sequence, signal plans, and traffic count data
vehicles and 2) the mean speed of vehicles currently inside the necessary to create the origin–destination (OD) matrix for the
network. The delay at each signalized intersection is computed PARAMICS software were obtained from the Land Transport
as the difference between the actual travel time of vehicles Authority of Singapore. The section of the network considered
across the intersection and the travel time in case of no signal for simulation is prone to frequent traffic jams because of the
control. The actual delay is calculated as the sum of the accel- high traffic flow in this region. The network constructed is
eration, deceleration, and stopped time delay for each vehicle at similar to the network used in [14], [26], [38], and [39], and
an intersection. In microscopic traffic-simulation platforms like details are shown in Table I.
PARAMICS, the delay value calculated at each intersection of The network has been calibrated for the real-time data for
the network at every time step of the simulation is stored in the simulations conducted in [26], [38], and [39]; therefore, no
722 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 11, NO. 3, SEPTEMBER 2010

is obtained based on the collated information of vehicle count

at intersections.
The GLIDE [13] signal controller is the modified version
of the SCATS currently used in Singapore for the control and
management of road traffic. The signal plans used were derived
from the actual signal settings used in Singapore road networks.
GLIDE is a centrally managed signal controller whose deci-
sions are based on the data collected from the loop detectors at
each intersection, with the main objective of maintaining the
degree of saturation. GLIDE has been simulated to the best
possible extent for the same network and similar simulation
scenarios as in [14] and [39], which are used to benchmark the
proposed signal controller.
Another benchmark used is the HMS designed based on
hybrid neural network in [39] and applied to the same network
as GLIDE. The HMS signal control is a semidistributed con-
trol system with different levels of hierarchical architecture to
determine the required green time. The semidistributed nature
of the controller provides partial autonomous decision-making
capability and reduces the communication requirements and
Fig. 12. Section of the road network used for simulation. data mining operations. However, the limitations are real-time
TABLE I computational needs restricted by the use of genetic algorithms
D ETAILS OF THE N ETWORK S IMULATED based training and communication requirements (although the
amount of data to be communicated is less than that of a
centralized architecture, it is still considerably high).

VIII. R ESULTS AND D ISCUSSIONS

Two types of simulation scenarios were used to compare
the performance of the proposed multiagent-based traffic signal
recalibration is required. All the benchmarks and the proposed
control and the benchmarks. A 6-h simulation scenario with
algorithm have been tested for similar traffic-flow patterns, and
typical morning and afternoon peaks was used to study the con-
all comparisons are made to the simulated results and not to the
trol efficiency of the multiagent signal control when subjected
actual delay experienced.
to dynamically varying traffic conditions within a short duration
of time. The other scenario was a long-term simulation with
B. Agent Implementation multiple peaks to emulate the infinite-horizon problem where
The agents were implemented as dll plugins in PARAMICS. the traffic arrival can be in platoons (pseudorandom) or com-
Each agent uses local data received from the sensors directly pletely stochastic with multiple repetitive traffic patterns. This
connected to the intersection, and the communicated congestion is a highly fictitious scenario used to test the settling time and
data from the neighboring intersections connected to the outgo- response speed of the signal control by creating repetitive high
ing links to determine the green time required during the next and low traffic peaks within a short period of time. For each
cycle for the particular phase using a geometric type-2 fuzzy simulation scenario, ten simulation runs with different random
decision system. seeds were conducted. The random seeds ensure the release of
the vehicle into the network to be stochastic even though the
OD matrix provides the exact quantity of vehicles that would
C. Benchmarks
be released within a specified period. The random seed also
It is very difficult to obtain a reliable benchmark for es- ensures the starting point of simulations to be different in each
timating the improvements achieved by using the proposed run. Since the vehicle release strategy is unique in all the simu-
traffic signal control. It would be unfair to compare the hybrid lation runs, the comparison of point data of travel time delay or
intelligent system with fixed signal controllers or controllers speed is of least significance. Therefore, the travel time delay
developed for isolated intersections as they are developed for is averaged over all the runs, and a comparison of the results
simplified scenarios. Moreover, the results obtained in some of the delay value after a specific heavy-traffic period (decided
of the papers cannot directly be compared as they were tested based on the input OD matrix) is shown in Tables II and V.
for different traffic networks, traffic loading, and signal plans The smaller difference in the variance of data in simulation runs
[9], [16]. In view of these issues, the network used for testing makes the averaging an acceptable option.
the proposed signal controller is the same as the one used for To verify the efficiency of the GFMAS signal control for
the benchmarks in [26] and [39]. The OD matrix providing the planned and unplanned incidents, additional tests like ran-
details of the network loading during a specified period of time dom incidents on major and minor links and closure of
GOKULAN AND SRINIVASAN: DISTRIBUTED GEOMETRIC FUZZY MULTIAGENT URBAN TRAFFIC SIGNAL CONTROL 723

TABLE II TABLE IV
M EAN D ELAY AND S PEED C OMPARISON FOR THE C ONFIDENCE I NTERVAL OF THE 6-h T WO -P EAK S IMULATION
S HORT-T ERM T WO -P EAK S CENARIO

TABLE III
C OMPARISON OF V EHICLE C OUNT IN THE N ETWORK
FOR THE 6-h T WO -P EAK S IMULATION

heavy-traffic lanes were conducted. These simulation scenarios

are analogous to the impulse and step input and provide an Fig. 13. Plot of a 6-h two-peak scenario for simulation runs with random
insight into the responsiveness of the GFMAS signal control. seeds.
The simulation of all failure scenarios for complete reliability
TABLE V
testing is beyond the scope of this paper. However, planned and M EAN D ELAY C OMPARISON FOR THE L ONG -T ERM
unplanned incidents give us an insight into the performance un- E IGHT-P EAK S IMULATION
der unforeseen conditions. The simulation results are discussed
in the following sections.

A. Two-Peak Scenario
The two-peak simulation is a typical traffic pattern, where
the morning and afternoon heavy-traffic situation is simulated.
Table II shows the comparison of results in terms of delay
and speed for the proposed GFMAS signal control and the
benchmarks at the end of the peak traffic period. It can be seen current mean speed of the vehicles. To prove the repeatability
that there is a significant improvement in the delay experienced of results, an analysis of worst and best simulation run results
by vehicles during the peak traffic time when using GFMAS is performed. The standard deviation of the time delay for all
signal control, in comparison with HMS and GLIDE. It can the ten simulation runs and the 90%, 95%, and 99% confidence
also be seen that, for both HMS and GLIDE, the performance intervals are calculated, as shown in Table IV. As can be
degrades during the second peak period and can be attributed seen from Table IV, even for the 99% confidence interval, the
to the settling time. Table III shows the percentage reduction delay value fluctuation is restricted to a maximum of 9.33 s.
in the number of vehicles retained inside the network after Fig. 13 shows the 6-h two-peak simulation results of two runs
a specific block of time period. Since the entire traffic-flow conducted with random starting seed values along with the peak
generation is stochastic, the comparison between the different traffic region.
control strategies is valid only when the number of vehicles
released into the road network at the end of specific time periods
B. Multiple-Peak Scenario
remains close to each other. Table III confirms the closeness
and, hence, the validity of the comparison. The comparison is The multiple-peak scenario is the extreme-traffic pattern
done with the best performing benchmark HMS. simulation stress test to verify the integrity, robustness, and re-
The results clearly indicate the effectiveness of GFMAS sig- sponsiveness of the signal control when subjected to repetitive
nal control in clearing traffic at intersections and improving the high traffic conditions within a short interval. It can be seen
724 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 11, NO. 3, SEPTEMBER 2010

TABLE VI
V EHICLE S TATISTICS FOR THE 24-h E IGHT-P EAK S CENARIO

Fig. 14. Plot of 24-h eight-peak scenario for simulation runs with random seeds.

from Table V that the GLIDE signal control performance starts undergoes no any degradation and performs better than both
to degrade after the fifth traffic peak period. The main reason GLIDE [13] and HMS [19], [39]. Since it has been shown that
for this degradation is the inability to clear the vehicles present traffic begins to saturate after fifth peak period in GLIDE, a
inside the network within a short duration of time before the comparison of the number of vehicles inside the network is not
start of the next high traffic period. GLIDE signal control adopts shown in Table VI. Fig. 14 shows the results obtained in two of
a prespecified value for the cycle length during the off-peak the simulation runs for eight peak traffic-input conditions. The
periods and only changes it to another prespecified value during scenarios tested so far serve to show the applicability of the
the peak period. This makes the green time allocated to each GFMAS signal control to traffic signals for normal operating
phase insufficient to clear the vehicles when the traffic input conditions.
increases, resulting in queue spillback. This claim is supported
by results shown in Table V, which indicates a steady increase
C. Planned-Event Simulation Scenario
in the delay as more intersections begins to get congested,
leading to deadlocks. HMS [14], [39] signal control also starts Planned-event simulation scenarios are of practical impor-
showing degradation in performance after the sixth peak period, tance and recreate the conditions of road blockage due to
which can be attributed to the increase in the number of vehicles preorganized or planned events such as clearance of roadside
waiting for green time. This can be attributed to the conflict trees, scheduled maintenance of traffic-management systems
in decisions between agents of different hierarchies, causing like variable message signs, electronic road pricing unit main-
smaller green time to be allocated to congested intersections. tenance, or special events like “Formula One car racing.”
Therefore, the number of vehicles leaving the simulated section Simulations were conducted by closing some lanes in the links
of the network increases as the vehicle input is kept constant and reducing the total load capacity of the link. This scenario
based on the OD matrix and can be seen in Table VI. How- is similar to the step input in a control system. The results
ever, results show that the proposed GFMAS signal control are shown in Fig. 15 and indicate that the closure of minor
GOKULAN AND SRINIVASAN: DISTRIBUTED GEOMETRIC FUZZY MULTIAGENT URBAN TRAFFIC SIGNAL CONTROL 725

Fig. 15. Planned incident—GFMAS signal control response to lane closure.

Fig. 16. Unplanned incident—GFMAS signal control random incident response.

roads could improve the travel time. The main reason for this PARAMICS. The responsiveness of GFMAS to such intermit-
is the routing mechanism used in PARAMICS. The routing is tent disturbances, which are analogous to impulse inputs, was
usually performed by calculating the feedback costs associated studied using this scenario. The detection and clearance of the
with each link network-wide for every prespecified interval incidents is assumed to be handled by different subunits of the
(5 min in this case). This causes the vehicles to opt for major traffic-management system, which currently do not come under
links, thereby achieving better optimization and lesser travel the GFMAS architecture.
delay time. However, when multiple lanes were closed, a slight Fig. 16 shows the various scenarios simulated for this study.
degradation is seen during the first peak traffic period, which The incidents were created during the first peak period to
subsequently settles down in the later peak periods to a slightly examine the response and time taken to settle down. The
higher delay value than in the normal operation. incidents were simulated in the link connecting Victoria Street
with Rochor Road and in the link connecting Rochor Road to
Bencoolean Street, as indicated in the network map shown in
D. Unplanned Events
Fig. 12. It can be observed that two simultaneous incidents at
Unplanned-event simulation refers to the incidents and acci- different links cause the GFMAS signal control performance to
dents that happen in a random manner and cannot be controlled. degrade. It is noted that, although the incidents were simulated
Incidents were randomly created using the incident file in very close to the peak traffic period and the alleviation of traffic
726 IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS, VOL. 11, NO. 3, SEPTEMBER 2010

congestion is affected to a certain extent, the proposed GFMAS [5] Vagverket, Signal Design With LHOVRA, Project Handbook.
signal control is able to effectively handle the increased traffic [6] P. Kronborg, F. Davidsson, and J. Edholm, “SOS-self optimizing signal
control, development, and field trials of the SOS algorithm for self opti-
congestion and bring down the average delay experienced by mizing signal control at isolated intersections,” 1997, TFK Rep.
the vehicles. [7] C. P. Pappis and E. M. Mamdani, “A fuzzy logic controller for
traffic junction,” IEEE Trans. Syst., Man, Cybern., vol. SMC-7, no. 10,
pp. 707–717, Oct. 1977.
IX. C ONCLUSION [8] Y. Sazi Murat and E. Gedizlioglu, “A fuzzy logic multi-phased signal
control model for isolated junctions,” Transp. Res. Part C, vol. 13, no. 1,
A distributed multiagent signal control (GFMAS) that uses pp. 19–36, Feb. 2006.
[9] J. Niittymaki and E. Turunen, “Traffic signal control on similarity logic
a geometric type-2 fuzzy inference engine has been proposed reasoning,” Fuzzy Sets Syst., vol. 133, no. 1, pp. 109–131, Jan. 2003.
and successfully tested on a simulated highly complex traffic [10] P. B. Mirchandani and N. Zou, “Queuing models for analysis of traffic
network by subjecting it to challenging simulation scenarios. adaptive signal control,” IEEE Trans. Intell. Transp. Syst., vol. 8, no. 1,
pp. 50–59, Mar. 2007.
The proposed architecture utilizes the input and rule uncertainty [11] P. B. Hunt, D. I. Robertson, R. D. Bretherton, and R. L. Winton, “SCOOT:
handling capabilities of a type-2 fuzzy architecture to carry A traffic responsive method of coordinating signals,” TRRL, Crowthorne,
the uncertainty to the consequent part and compute the output. U.K., Rep. LR1014, 1981.
[12] P. R. Lowrie, “The Sydney coordinated adaptive traffic system—
Some of the significant features of the proposed signal control Principles, methodology, algorithms,” in Proc. Int. Conf. Road Traffic
architecture are summarized as follows: Signal., London, U.K., Mar. 30–Apr. 1, 1982, pp. 67–70.
[13] C. K. Keong, “The GLIDE system—Singapore’s urban traffic control
1) fully distributed multiagent architecture with no central system,” Transp. Rev., vol. 13, no. 4, pp. 295–305, Oct. 1993.
supervising or synchronizing agent; [14] M. C. Choy, “Cooperative, hybrid multi-agent systems for distributed,
2) demonstrated the use of type-2 fuzzy decision system for real-time traffic signal control,” Ph.D. dissertation, Dept. Elect. Comput.
Eng., Nat. Univ. Singapore, Singapore, 2005.
traffic system with uncertain and noisy data’; [15] M. Balmer, K. Nagel, and B. Raney, “Large-scale multi-agent simulations
3) coordination between agents through communication of for transportation applications,” J. Intell. Transp. Syst., Technol. Plan.
decision values; Oper., vol. 8, no. 4, pp. 205–221, Oct. 2004.
[16] I. Kosonen, “Multi-agent fuzzy signal control based on real-time simula-
4) design of the internal belief model as a part of the decision tion,” Transp. Res. Part C, vol. 11, no. 5, pp. 389–403, Oct. 2003.
system; [17] P. Mirchansani and L. Head, “A real-time traffic signal control system:
5) asynchronous synchronization of agents using local Architecture, algorithms, and analysis,” Transp. Res. Part C, vol. 9, no. 6,
pp. 415–432, Dec. 2001.
memory. [18] F.-Y. Wang, ABCS: Agent-Based Control Systems. Tucson, AZ: Univ.
In the proposed architecture (GFMAS), the congestion state Arizona, 1998.
communicated by each agent to its neighbors is used to deter- [19] F.-Y. Wang, “aDCS: Agent-Based Distributed Control Systems,” Univ.
Arizona, Tucson, AZ, 1999.
mine the cooperation level internally within the framework of [20] Q. Wu and F.-Y. Wang, “A mobile-agent based distributed intelligent
the type-2 decision system. This increases the cooperation level control system architecture for home automation,” in Proc. IEEE Int.
and effectively avoids the communication bottleneck associated Conf. Syst., Man, Cybern., Tucson, AZ, 2001, pp. 1599–1605.
[21] F.-Y. Wang, “Toward a revolution in transportation operations: AI
with the flow of control from agents at a higher hierarchy to for complex systems,” IEEE Intell. Syst., vol. 23, no. 6, pp. 8–13,
lower hierarchy agents. The proposed architecture also reduces Nov./Dec. 2008.
the requirement of clock synchronization required in HMS by [22] K. Dresner and P. Stone, “A multiagent approach to autonomous inter-
section management,” J. Artif. Intell. Res., vol. 31, no. 1, pp. 591–656,
utilizing localized memory and broadcast communication. Jan. 2008.
A comparison with benchmark signal controls GLIDE and [23] F.-Y. Wang, “Agent-based control for networked traffic management
HMS showed that GFMAS signal control outperformed them systems,” IEEE Intell. Syst., vol. 20, no. 5, pp. 92–96, Sep. 2005.
[24] R.-S. Chen, D.-K. Chen, and S.-Y. Lin, “ACTAM: Cooperative multi-
under all the simulation scenarios and was capable of alleviat- agent system architecture for urban traffic signal control,” IEICE Trans.
ing the congestion experienced at the intersections. The promis- Inf. Syst., vol. E88-D, no. 1, pp. 119–126, Jan. 2005.
ing results suggest the possibility of improvement in traffic [25] J. Bonneson, D. Middleton, K. Zimmermann, H. Charara, and M. Abbas,
“Intelligent detection-control system for rural signalized intersections,”
conditions when applied to a real traffic network. However, this Texas Transp. Inst., College Station, TX, Rep. FHWA/TX-03/4022-2,
would require calibration of the various parameters involved Aug. 2002.
in the GFMAS signal control. Further improvements could [26] D. Srinivasan, M. C. Choy, and R. L. Cheu, “Neural networks for real-
time traffic signal control,” IEEE Trans. Intell. Transp. Syst., vol. 7, no. 3,
be achieved by employing evolutionary techniques and online pp. 261–272, Sep. 2006.
reinforcement learning for fine-tuning of the membership func- [27] H. Wu and J. M. Mendel, “Uncertainty bounds and their use in the design
tions and the rule base, thereby increasing the possibility of of interval type-2 fuzzy logic systems,” IEEE Trans. Fuzzy Syst., vol. 10,
no. 5, pp. 622–639, Oct. 2002.
applying the signal control technique to real traffic networks. [28] Q. Liang, N. N. Karnik, and J. M. Mendel, “Connection admission control
in ATM networks using survey based type-2 fuzzy logic systems,” IEEE
Trans. Syst., Man, Cybern. C, Appl. Rev., vol. 30, no. 3, pp. 329–339,
R EFERENCES Aug. 2000.
[1] X.-H. Yu and A. R. Stubberud, “Markovian decision control for traffic [29] L. A. Zadeh, “The concept of linguistic variables and its application to
signal systems,” in Proc. 36th IEEE Conf. Decision Control, San Diego, approximate reasoning,” Inf. Sci., vol. 8, pp. 199–249, 1975.
CA, Dec. 1997, vol. 5, pp. 4782–4787. [30] N. N. Karnik, J. M. Mendel, and Q. Liang, “Type-2 fuzzy logic systems,”
[2] W. W. Recker, B. V. Ramanathan, X.-H. Yu, and M. G. McNally, IEEE Trans. Fuzzy Syst., vol. 7, no. 6, pp. 643–658, Dec. 1999.
“Markovian real-time adaptive control of signal systems,” J. Math. Com- [31] Q. Liang and J. M. Mendel, “Interval type-2 fuzzy logic systems:
put. Model., vol. 22, no. 4–7, pp. 355–375, Aug.–Oct. 1995. Theory and design,” IEEE Trans. Fuzzy Syst., vol. 8, no. 5, pp. 535–550,
[3] F. Webster, “Traffic signal settings,” Road Res. Lab., London, U.K., Road Oct. 2000.
Res. Tech. Paper r39, 1958. [32] N. N. Karnik and J. M. Mendel, “Centroid of a type-2 fuzzy set,” Inf. Sci.,
[4] J. R. Peirce and P. J. Webb, “MOVA control of isolated traffic signals- vol. 132, no. 1–4, pp. 195–220, Feb. 2001.
recent experience,” in Proc. 3rd Int. Conf. Road Traffic Control, London, [33] S. Coupland and R. John, “Geometric type-1 and type-2 fuzzy logic
U.K., May 1990, pp. 110–113. systems,” IEEE Trans. Fuzzy Syst., vol. 15, no. 1, pp. 3–15, Feb. 2007.
GOKULAN AND SRINIVASAN: DISTRIBUTED GEOMETRIC FUZZY MULTIAGENT URBAN TRAFFIC SIGNAL CONTROL 727

[34] S. Coupland and R. John, “New geometric inference techniques for Dipti Srinivasan (M’90–SM’02) received the
type-2 fuzzy sets,” Int. J. Approx. Reason., vol. 49, no. 1, pp. 198–211, M.Eng. and Ph.D. degrees in electrical engineering
Sep. 2008. from the National University of Singapore (NUS),
[35] Highway Capacity Manual 2000—HCM2000, Transp. Res. Board, Nat. Singapore, in 1991 and 1994, respectively.
Res. Council, Washington, DC, 2000. She was with the Computer Science Division,
[36] J. E. Clarke, “Assessing the sensibility of signal timing split optimization University of California, Berkeley, as a Postdoctoral
in addressing congestion,” in Proc. Inst. Transp. Stud. Annu. Meeting Researcher from 1994 to 1995. In June 1995, she
Exhib., 2007, pp. 24–29. joined the faculty of the Department of Electrical and
[37] J. D. Little, “A proof for the queuing formula: L = λW,” Oper. Res., Computer Engineering, NUS, where she is currently
vol. 9, no. 3, pp. 383–387, May 1961. an Associate Professor. From 1998 to 1999, she was
[38] P. G. Balaji, D. Srinivasan, and C. K. Tham, “Coordination in distributed a Visiting Faculty Member with the Department of
multi-agent systems using type-2 fuzzy decision systems,” in Proc. FUZZ- Electrical and Computer Engineering, Indian Institute of Science, Bangalore,
IEEE, Jun. 2008, pp. 2291–2298. India. She has published/presented more than 170 technical papers in interna-
[39] M. C. Choy, D. Srinivasan, and R. L. Cheu, “Neural networks for con- tional refereed journals and conference proceedings. Her main areas of interest
tinuous online learning and control,” IEEE Trans. Neural Netw., vol. 17, are neural networks, evolutionary computation, intelligent multiagent systems,
no. 6, pp. 1511–1531, Nov. 2006. and the application of computational intelligence techniques to engineering op-
timization, planning, and control problems in intelligent transportation systems
and power systems.
Balaji Parasumanna Gokulan (S’09) is currently Dr. Srinivasan is a member of the Institute of Engineers Singapore.
working toward the Ph.D. degree with the De-
partment of Electrical and Computer Engineering,
National University of Singapore, Singapore.
His research interests include multiagent systems,
application of hybrid computational intelligence
techniques to engineering applications, type-2 fuzzy
inference, and intelligent transportation systems.