An optimal general type-2 fuzzy controller for Urban Trafﬁc Network

An optimal general type-2 fuzzy controller for Urban Traffic Network
Mohammad Hassan Khooban a,n
, Navid Vafamand b
, Alireza Liaghat b
, Tomislav Dragicevic a
a
Department of Energy Technology, Aalborg University, Aalborg DK-9220, Denmark
b
School of Electrical and Computer Engineering, Shiraz University, Shiraz, Iran
a r t i c l e i n f o
Article history:
Received 12 January 2016
Received in revised form
8 September 2016
Accepted 10 October 2016
Available online 3 November 2016
This paper was recommended for publica-
tion by Oscar Camacho
Keywords:
Traffic Signal Control
Optimal General Type-2 Fuzzy Controller
(OGT2FC)
Modified Backtracking Search Algorithm
(MBSA)
a b s t r a c t
Urban traffic network model is illustrated by state-charts and object-diagram. However, they have
limitations to show the behavioral perspective of the Traffic Information flow. Consequently, a state space
model is used to calculate the half-value waiting time of vehicles. In this study, a combination of the
general type-2 fuzzy logic sets and the Modified Backtracking Search Algorithm (MBSA) techniques are
used in order to control the traffic signal scheduling and phase succession so as to guarantee a smooth
flow of traffic with the least wait times and average queue length. The parameters of input and output
membership functions are optimized simultaneously by the novel heuristic algorithm MBSA. A com-
parison is made between the achieved results with those of optimal and conventional type-1 fuzzy logic
controllers.
& 2016 ISA. Published by Elsevier Ltd. All rights reserved.
1. Introduction
One of the main difficulties in big modern cities is the over-
populations of automobiles. Traffic signal control has a crucial role
in the transport safety and the smoothness of traffic flow. Parti-
cularly, in order to prevent automobiles overcrowding in urban
streets, not only off-line timing, but also real-time control of traffic
signals has been proposed recently [1–4].
A crossroad is the main node of the municipal transport net-
work. Collection of the traffic data and control of the traffic flow
surrounding it is a challenging research subject. There are two
major techniques in signal control: off-line signal control and real-
time signal control. Because of the stochastic nature of traffic flow,
the off-line traffic control technique can be only used for less
crowded crossroads. The real-time technique optimizes the signal
control based on the information collected by sensors. There are
several control techniques in literatures. For example in [3,4], a
novel technique based on video reorganization is presented.
In some research activities, optimization algorithms are used.
The major optimization algorithms include fuzzy logic system
(FLS), neural network-fuzzy (NNF), multi-objective genetic algo-
rithms (MOGA), and Markov Process [5–9].
Lately, Diakaki et al. [2] have suggested a model for traffic
monitoring, and have used optimal linear quadratic regulator for
controlling the model. While this study looks encouraging for
smart control of traffic, the model and the control procedure have
several limitations. In order to overcome these defects, a more
comprehensive model and a robust control method have been
suggested in [10,11]. The modeling procedures in [10,11] and also
in [2] are based on the so-called store-and-forward modeling
method which requires some statistical data related to traffic, for
instance saturation flows and turning motion rates to create the
model. So, this modeling procedure has flaws that it is fairly so-
phisticated to create a model from statistical data and that it might
be too perfect to consider for an actual traffic problem.
The fuzzy logic controller (FLC) is credited with being an ap-
propriate method for designing robust controllers that are capable
of delivering a satisfactory efficiency against uncertainty; therefore
the FLC has become a common solution to reactive traffic signal
control in recent years [12]. The type-1 FLCs have the popular
problem that they cannot be completely used for the linguistic and
numerical uncertainties related to variable environmental condi-
tions as they apply accurate type-1 fuzzy sets. Type-1 fuzzy sets
employ the uncertainties related to the FLC inputs and outputs by
applying accurate and crisp membership functions that the gainer
believes that uncertainty is inhibited [13]. When the type-1
membership functions have been selected, all the uncertainty is
eliminated, since type-1 membership functions are completely
accurate [13,14]. The linguistic and numerical uncertainties related
to variable environmental conditions create problems in specifying
the accurate consequents membership functions over the design
process.
Contents lists available at ScienceDirect
journal homepage: www.elsevier.com/locate/isatrans
ISA Transactions
http://dx.doi.org/10.1016/j.isatra.2016.10.011
0019-0578/& 2016 ISA. Published by Elsevier Ltd. All rights reserved.
n
Corresponding author.
E-mail addresses: mhkhoban@gmail.com, khooban@sutech.ac.ir,
mhk@et.aau.dk (M.H. Khooban).
ISA Transactions 66 (2017) 335–343

Lately, many researchers [15–18] pay attention to general type-
2 fuzzy sets and systems because of their ability to deal with un-
certainties and disturbances. Zadeh [17] in 1975 presented Type-2
fuzzy sets as an extension of type-1 fuzzy sets and have been used
in engineering areas successfully. For instance, [19] demonstrates
the efficient performance of IT2FLSs in comparison to type-1 fuzzy
logic systems (T1FLS) when faced with various uncertainties such
as dynamic uncertainties, rule uncertainties, external disturbances
and noises. Available information for making the rules in a fuzzy
logic system can be uncertain. Unlike interval type-2 fuzzy sets
(IT2FS) and type-1 fuzzy sets (T1FS), general type-2 fuzzy sets can
deal with the rule uncertainties. In literatures, only IT2FLSs have
been mainly applied until now because general type-2 fuzzy sets
and systems are computationally complex. Liu [19] proposed a
useful fast process for computing centroid and type reduction of
GT2FLS by using a recent plane representation theorem. In [20–23]
, an in-depth description of the zSlices-based representation,
which enables the representation of and computation with general
type-2 fuzzy sets and their associated third dimensions at a level
of precision and associated computational overhead, which can be
chosen as required by the respective application has given. Bilgin
et al. [24] addressed the need to enhance transparency in Ambient
Intelligent Environments by developing more natural ways of in-
teraction, which allow the users to communicate easily with the
hidden networked devices rather than embedding obtrusive ta-
blets and computing equipment throughout their surroundings. A
novel zSlices based general Type-2 Fuzzy PI (zT2-FPI) controller
where the SMFs are adjusted in an on-line manner through a
single tuning parameter is presented in [25].
Motivated by the aforementioned researches, the purpose of
this paper is to present an Optimal General Type-2 Fuzzy Con-
troller (OGT2FC) for controlling the traffic signal scheduling and
phase succession to guarantee a smooth flow of traffic with the
least wait times and average queue length. The parameters of in-
put and output membership functions are optimized simulta-
neously by a novel heuristic algorithm called Modified Back-
tracking Search Algorithm (MBSA). Simulation results indicate the
superiority of the proposed controller over the non-optimal type-1
fuzzy controller and optimal type-1 fuzzy controller.
2. General type-2 fuzzy sets and systems
A GT2FS in a universal set X can be defined as
∫μ̃ =
( )
∈ ( )
̃
A
x x
x X
/
1
A
∫μ ( ) =
( ( ))
∈
∈[ ]
( )
̃ x
z u u
u J
J
/
, 0, 1
2
A
x
x
x
where in this formula μ ( )̃ xA
is called a secondary membership
function (MF) and ( )z ux is called secondary grade; Jx
is the domain
of the secondary MF which is called primary membership and u is
a fuzzy set in [0, 1]. Fig. 1 illustrates a GT2FS where the upper and
lower MFs are triangular and its secondary MF is also triangular.
When ( )=z u 1x IT2FS is obtained that demonstrate a uniform un-
certainty in the primary membership function and is simply de-
scribed by its lower μ̲ ( )˜ xA
and upper μ¯ ( )˜ xA
MFs. Because of cal-
culation simplicity, especially in the type reduction, many re-
searchers use interval type-2 fuzzy sets instead of general type-2
fuzzy sets [16,18,20].
Lately, Liu [19] presented a new method for GT2FSs which is
theoretically and computationally effective. Because this method
resembles the α-cut for type-1 fuzzy sets, it is named a α-plane for
type-2 fuzzy sets. ̃
αA is the denotation of An α-plane representa-
tion for a GT2FS ̃A. It is the union of all primary MFs whose sec-
ondary grades are greater than or equal to the special value α:
∫μ̃ =
( )
∈ ( )
α
̃
αA
x x
x X
/
3
A
∫μ
α
( ) =
( ( ) ≥ )
∈
∈ [ ]
( )
̃
α
x
z u u
u J
J
/
, 0, 1
4
A
x
x
x
Then a GT2FS ̃A based on α-plane representation theorem can
be demonstrated in the following form:
α̃ = ⋃ ̃
( )α
α
∈[ ]
A A/
51,0
It is a beneficial representation because α ̃
αA/ can be seen as an
IT2FS with the secondary grade of level α. As a result, several
IT2FSs may be made from the decomposition of a general type-2
fuzzy set with a corresponding level of α for each, where
α = { … ( − ) }K K K0,1/ , , 1 / , 1 . In simpler terms, a general type-2
fuzzy logic system can be seen as a huge collection of IT2FLSs with
one IT2FLS for each value of α. However, Liu [24] showed that
using only 5 to 10 α-plane can get the required accuracy for cen-
troid calculation. Fig. 2 illustrates the new design for a general
type-2 fuzzy system based on α-plane representation.
In general, a GT2FLS is made of a fuzzifier; fuzzy rule-based;
fuzzy inference engine; type reducer and defuzzifier. Fuzzifier
maps real values into fuzzy sets. Singleton fuzzifier whose output
is a single point of a unity membership grade is used in this paper
Fig. 1. A general type-2 fuzzy set with triangular upper and lower MFs where the secondary MF is triangular.
M.H. Khooban et al. / ISA Transactions 66 (2017) 335–343336

because it is simple. Fuzzy rule base includes fuzzy IF-THEN rules.
In the following The jth
rule in the GT2FLS is shown:
~
…
~ ~
= … ( )R x F x F y G j M: If is and is then is 1, 2, , 6
j j
n n
j j
1 1
where xi for = …i n1, , and y are the input and output of the
GT2FLS, ̃Fi
j
and ̃G
j
are general type-2 antecedent and the con-
sequent sets. A mapping from input GT2FSs to output GT2FSs is
given by the inference engine that merges rules. Because α-plane
representation for fuzzy set is used, the firing set for each related
IT2FS is shown as following:
( ) ( ) ( )= ̲ ¯
( )α α α
⎡
⎣⎢
⎤
⎦⎥F X f X f X,
7
j j j
( ) μ μ μ̲ = ̲ * ̲ *…* ̲
( )α ˜ ˜ ˜
α α α
f X
8
j
F F F
j j
n
j
1 2
( ) μ μ μ¯ = ¯ * ¯ *…* ¯
( )α ˜ ˜ ˜
α α α
f X
9
j
F F F
j j
n
j
1 2
Here ( )̲ α
f Xj
and ( )¯
α
f X
j
are the lower and upper MFs of the jth
rule with level of α, and * indicates product t-norm. a type reducer
changes the output of the inference engine which is a type-2 fuzzy
set into a type-1 fuzzy set before defuzzification. Five kinds of
reducers which are based on calculating the centroid of an IT2FS
are demonstrated in [25]. The output of the type reduction in
IT2FLS is defined with its left-end point yl
and right-end point yr
due to uniformly secondary grade of IT2FLS.
KM iterative algorithms, introduced two algorithms for calcu-
lating these two end points in [25–27], as presented by Mendel
and Karnik. In comparison to the other type reduction methods,
center of sets (COS) is used a lot because of its computation sim-
plicity [15]. If a singleton fuzzifier is used, product inference en-
gine and COS type reducer, left and right end points for each part
of GT2FLS based on α- representation theorem can be shown as
follows:
θ θ
θ ξ=
∑ ¯ + ∑ ̲
∑ ¯ + ∑ ̲
=
( )
α α
α α
= = +
= = +
α
α α
α α
y
f f
f f 10
l
j
L j
l
j
j L
M j
l
j
j
L j
j L
M j
l
T
l
1 1
1 1
where in this formula θ
αl
j
is the left-end point of jth
consequent set
with level of α θ θ θ ξ= … = = ∑ ¯ +α α α α
α
α
α
α
α α
̲ ¯
=
⎡
⎣
⎤
⎦
⎡
⎣
⎢
⎢
⎤
⎦
⎥
⎥
D f, , , , , ,l l l
M
T
l
j f j
Dl
f
j
Dl
l j
L j1
1
ξ ξ ξ∑ ̲ = …α= + α α α
⎡
⎣
⎤
⎦f and , ,j L
M j
l l l
M
T
1
1
. In addition,
θ θ
θ ξ=
∑ ¯ + ∑ ̲
∑ ¯ + ∑ ̲
=
( )
α α
α α
= = +
= = +
α
α α
α α
y
f f
f f 11
r
j
R j
r
j
j R
M j
r
j
j
R j
j R
M j
r
T
r
1 1
1 1
where θ αr
j
is the right end point of jth
consequent set with level of
α θ θ θ ξ ξ= … = = ∑ ¯ + ∑ ̲α α α α
α
α
α
α
α α α α
̲ ¯
= = +
⎡
⎣
⎤
⎦
⎡
⎣
⎢
⎢
⎤
⎦
⎥
⎥
D f f, , , , , , andr r r
M
T
r
j f j
Dr
f
j
Dr
r j
R j
j R
M j
r
1
1 1
ξ ξ= …α α
⎡
⎣
⎤
⎦, ,r r
M
T
1
Meanwhile, performing KM iterative algorithm can
specify R and L for each individual IT2FLS of level α. From the
combination of all of these obtained intervals into a type-1 fuzzy
set like Fig. 3, a crisp output can be obtained using centroid de-
fuzzification as:
( ) ( )
( )
α
α
α
α=
∑ +
∑
=
∑ +
+
= { … ( − ) }
α α α
α
α α α
12
y
y y y y
K
K K K
2 1
, 0, 1/ , , 1 / , 1
l r l r
where +K 1 shows the number of the α-planes or in other words
it determines the number of individual IT2FLSs.
3. Optimization technique
Based on the updating process of genetic algorithm (GA) and
differential evolution (DE), a new backtracking search algorithm
(BSA) is developed as an evolutionary optimization (EO) technique
in [28]. Similar to other EOs, the application of basic principles of
Fig. 2. Architecture for a general type-2 fuzzy Logic system.
Fig. 3. Output of each individual IT2FLSs.
M.H. Khooban et al. / ISA Transactions 66 (2017) 335–343 337

BSA to mathematical optimization begins with the random gen-
eration of an initial population [34]. An objective function is used
to evaluate each individual of the initial set. It profits from a
random mutation techniques which employs only the one direc-
tion individual instead of multiple ones comparing to the DE al-
gorithm. Moreover, a non-uniform crossover procedure is im-
plemented along with the BSA which is more complicated than the
GA's crossover. In the crossover step, the variables of two selected
individuals are combined to each other and in the mutation step
the new solutions are randomly generated in order to diversify the
solution search space and allow the BSA to eventually escape from
locally optimal solutions. Furthermore, two new selection strate-
gies are defined and used by the BSA. According to this informa-
tion, the BSA can be divided into the following three steps. It
should be noted that the proposed modification strategies are
explained during the process of the BSA for the sake of
explicitness.
3.1. Initialization
Initially the N number of individual is randomly generated and
positioned at random locations on the search space. Here, the N
represents the D parameters involved in the fuzzy controller. The
individual position is characterized as an initial vector of length N
and D as the problem dimension. The initial position vector can be
represented as:
=
… …
⋮ ⋮ ⋮ ⋮ ⋮
… …
⋮ ⋮ ⋮ ⋮ ⋮
… … ( )
⎡
⎣
⎢
⎢
⎢
⎢
⎢
⎢
⎤
⎦
⎥
⎥
⎥
⎥
⎥
⎥
X
X X X
X X X
X X X 13
j D
i i j i D
N N j N D
1,1 1, 1,
,1 , ,
,1 , ,
Each array Xi j, is generated randomly using the uniform dis-
tribution U as it can be expressed:
= + (∙)( − ) ( )X X U X X 14i j min j i j max j min j, , , , ,
Where, Xmin j, and Xmax j, are the lower and upper bounds of the jth
member of the problem dimension. The (∙)Ui j, is the random
function generator in the range of [ ]0,1 from the uniform prob-
ability distribution function for the jth member of the ith agent.
First modification
Chaotic opposition-based learning is suggested in this study to
enhance the quality of the final solution and speed up the con-
vergence rate of the algorithm. If the random initialization is not
far away from the optimal solution, then the convergence is ex-
pected to be faster. In this learning scheme, the chaotic systems
and the opposition-based learning technique are combined to
generate the initial population. The population is divided into two
groups. The first group is initialized based on the chaotic theory
and the second one is generated according to the opposition-based
learning. Here, a sinusoidal based chaotic operator is used to
generate a chaotic number as follows:
( )π= = …
= (∙) ( )
+C C k k
C U
sin , 1, ,
15
i j k i j k max
i j i j
, , 1 , ,
, ,1 ,
where k is the iteration counter of sinusoidal iterator and
kmax is the maximum iteration number for chaotic mechanism.
Accordingly, the first group ( =N N/21 ) is generated as follows:
( )= + (∙) − = … ( )X X C X X i N, 1, , 16i j
C
min j i j k max j min j, , , , , , 1max
Where Xi j
C
, is the jth variable of the ith individual which is initially
generated by the chaotic system. The rest of the population is
initialized as follows [29]:
= + − ( )X X X X 17i j
o
min j max j i j, , , ,
Where Xi j
o
, is the jth variable of the ith individual which is
initially generated by the opposition-based learning scheme.
3.2. Selection-I
The BSA uses the Xold
vector to update the historical population
at the beginning of each iteration. This vector is initialized like the
X and updated through the following ‘If-Then’ rule [28]:
(∙) (∙) ⋈ ( )XU UIf Then :X 18
old
1 2
where ⋈: is an operation which is used for updating process of
Xold
. It means that the BSA has a memory and remembers the
unchanging population from the previous generation until it is
changed. After determining the current Xold
, the following process
is implemented to modify the order of the individuals in the Xold
in the current iteration as follows:
( )( )= ( )X X randperm N 19
old old
where ( )randperm N is a random shuffling function and randomly
selects the index from N.
3.3. Mutation, crossover and selection II
Following the embodiment of the selection-I process, a muta-
tion is a next implementation step in the BSA. Based on this, the
trial solutions are generated as follows [28]:
( )= + ( − ) ( )X X M X XF. . 20
trial old
where F is a parameter that controls the amplitude of the search di-
rection vector −X Xold
and considered to be × rndn3 in which
~ ( )rndn N 0,1 (N is the standard normal probability distribution func-
tion). Moreover, M is the binary integer-valued vector of size ×N D
which is used for the crossover purpose. It is generated as follows:
(∙) (∙)
=
=
∈
=
=
=
=
∈
=
=
( )
U U
i :N
j :D
j Q
M
M
i :N
j :D
j Q
M
M
If
For 1
For 1
If
1
Else
0
End
End
End
End
For 1
For 1
If
1
Else
0
End
End
End
21
i j
i j
i j
i j
1 2
1
,
,
2
,
,
where
( )( )
( )
= (∙)
= ( )
Q u ceil U D
u randperm D
1 : .
22
1
( )= ( )Q randi D 232
with ( )ceil S as a function which rounds S toward the upper integer.
Also, the randi produces random integers from a uniform discrete
distribution.

It is notable that the trial solutions with better fitness values
are replaced with the same position of population X for the next
iteration of the BSA.
Second modification
Following the embodiment of chaotic opposition-based learn-
ing process, a new self-adaptive learning strategy is presented and
applied to the original BSA in order to diversify the problem search
space. Based on the learning experience, a self-adaptive learning
guides the decision making for the modification rules. Three
modification rules are adopted to extend the exploration and ex-
ploitation area in the method, which are expressed as follows:
– Modification rule 1:
( )
( )υ υ
υ
=
+ −
+ − ( )
⎪
⎪
⎧
⎨
⎩
X
X Best Worst
X Worst X
if 0
otherwise 24
i mod
i new
i new i new
, 1
,
, ,
where Xi mod, 1 is the generated trial solution for the ith individual
through the modification rule 1. The Best and Worst are the
current best and worst solutions, respectively. The υ is for-
mulated as follows:
υ
ϕ
=
( )
ϕ−
⎛
⎝
⎜
⎞
⎠
⎟
⎛
⎝
⎜
⎜
⎞
⎠
⎟
⎟
e
d
1
2
cos
5
25
d2
2
2
where ϕ is a random number in the range of [ − ]d d2.5 , 2.5 . The
amount of d is set to change with the iteration number and
expressed as follows:
( )
= ( )
− −
−⎛
⎝
⎜
⎜⎜
⎛
⎝
⎜
⎛
⎝
⎜
⎞
⎠
⎟
⎞
⎠
⎟
⎞
⎠
⎟
⎟⎟
d e 26
iter iter
iter
ln 1 1max
max
0.5
where itermax is the maximum number of iterations for the
MBSA.
( )= + (∙) − ( )X X X XU 27i mod i new i m new i new, 2 , , ,
through the modification rule 2. The m is randomly chosen in-
dividual index from the population.
( )= + (∙) − ( )X X Best MeanU 28i mod i new i, 3 ,
through the modification rule 3. The Mean is the mean value of
the current population which is provided as follows:
=[ … … ] ( )Mean me me me 29j D1
∑=
( )=
X
me
N 30
j
i
N
i j new
1
, ,
While modification rule 1 improves both the global and local
search capacity, modification rules 2 and 3 enhance the local and
global search capacities, respectively. The selection of modification
rules depends on the following normalized probability model:
=
+ + ( )
Prob
Prob
Prob Prob Prob 31
r
r
1 2 3
= +
( )
Prob Prob
ac
0. 85 0. 15
10 32r r
r
= + = …
( )
ac ac
w
N
ii N1, ,
33
r r
ii
r
r
( )
( ) ( )
=
− +
+…+
= …
( )
w
N i
N
i N
log 1
log 1 log
1, ,
34
i
where Nr is the number of population which selects the mod-
ification rule r.
Finally, the roulette wheel mechanism on the basis of nor-
malized probability (31) is employed in order to choose the rth
modification rule for each individual.
4. An intelligent approach for the single crossroad
4.1. State-space equations
The two-phased signalized crossroad is shown in Fig. 4 [30]. In
this figure, the Leg 1 and Leg 3 are phase 1, and the Leg 2 and Leg
4 are phase 2.
The average queue length is a significant parameter that char-
acterizes the traffic state of a crossroad. The queue is defined as
follows
( + )= ( )+ ( )− ( ) ( ) ( )Q n Q n q n d n S n1 35i i i i i
where =i M1,2, . . . , is the index of the traffic flows;
= –n N0,1, . . . , 1 is the index of the discretized time intervals;
( )Q ni , based on number of automobiles, is the queue length of the
ith flow at the beginning of the nth time interval; ( )q ni
is the
number of automobiles that join the ith queue in the nth time
interval; ( )d ni is the number automobiles that leave the ith queue
in the nth time interval; and ( )S ni , which is equal to 0 (for stop-
ping) or 1 (for going), is the signal state of the ith flow in the nth
time interval. qi
and di are typically considered random signals.
In fixed time monitoring and fuzzy smart control, the control
variables are studied as follows [30]. For phase 1 crossroad,
( ) = ( )S S S S, , , 0,1,0,11 2 3 4 , which means traffic signal is green in lanes
2 and 4 and red in lanes 1 and 3. So, the automobiles can move in
Fig. 4. Two phase signalized crossroad. (For interpretation of the references to
color in this figure,the reader is referred to the web version of this article.)

lanes 2 and 4 and they should stay in lanes 1 and 3. On the other
side, for phase 2 crossroad, ( ) = ( )S S S S, , , 1,0,1,01 2 3 4 , which means
traffic signal is green in lanes 1 and 3 and red in lanes 2 and 4. So,
the automobiles can move in lanes 2 and 4 and they should stay in
lanes 2 and 4.
Integrating the average queue length due to the time, results
the average waiting time of automobiles in the queue. Let T is
short sufficiently, the automobiles arrivals can be assumed as
being identical in each time interval. Therefore, integrating (24)
yields
( )+ = ( )+ ( )+ ( )− ( ) ( )
( )
W n W n TQ n Tq n Td n S n1
1
2
1
2 36i i i i i i
where ( )W ni is the average waiting time of the ith queue since the
start of the time interval to the beginning of the nth time interval.
Eqs. (35) and (36) are the state-space equations characterizing
the dynamic development of the traffic state at a separate cross-
road. The length of queue and the average waiting time are two
common efficiency indices for signal controls. The average waiting
time is utilized here as the efficiency index. So, the cost function is
[31–33]:
( ) ∑= ( )
( )=
⎪ ⎪
⎪ ⎪
⎧
⎨
⎩
⎫
⎬
⎭
min W N W N
37i
M
i
1
To expedite the formulation, the state-space equations and the
cost function can be overridden in matrix form as:
( + ) = ( )+ ( ) ( )+ ( ) ( )X n AX n B n S n C n1 38
( )= ( ) ( )y n CX n 39
where
( )= ( ) ( ) ( ) ( ) ( ) ( )⎡⎣ ⎤⎦X n Q n Q n Q n W n W n W n.. . .. .M M
T
1 2 1 2 are the state
variables and ( )= ( ) ( ) ( )⎡⎣ ⎤⎦S n S n S n S n.. . M
T
1 2 are the control variables.
The different coefficient matrices and vectors are [32]
=
( ) = −
( ) …
( ) …
⋮ ⋮ ⋱ …
⋮ …
( ) … ( )
( ) …
⋮ … ⋱
… ( )
=
( ) = ( )⋯ ( ) ( )⋯ ( )
( )
⎡
⎣
⎢
⎢
⎤
⎦
⎥
⎥
⎛
⎝
⎜
⎜
⎜
⎜
⎜
⎜
⎜
⎜
⎜
⎜
⎜
⎜
⎜
⎜
⎜
⎜
⎞
⎠
⎟
⎟
⎟
⎟
⎟
⎟
⎟
⎟
⎟
⎟
⎟
⎟
⎟
⎟
⎟
⎟
⎡
⎣
⎢
⎢
⎤
⎦
⎥
⎥
⎡
⎣⎢
⎤
⎦⎥
A
I
TI I
B n
d n
d n
Td n d n
Td n
Td n
C
I
I
C n q n q n Tq n Tq n
0
0 0
0 0
0 0
1
2
0
0
1
2
0
0
0 0
1
2
0
0
1
2
1
2 40
M
M M
m
M
M
M
M M
1
2
1
2
1 1
4.2. Optimal Type-2 fuzzy controller for a single crossroad
Designing an optimal Type-2 fuzzy controller for a single sig-
nalized crossroad comprises three steps. In the following, we will
present these steps in detail.
The first step in designing of a fuzzy controller is determination
of input and output variables. Since the goal is to minimize the
cost function (37), ( )W ni are chosen as the input variables. Fur-
thermore, the states ( )Q ni are directly related to the number
joining and leaving automobiles (i.e. ( )q ni
and ( )d ni ). Therefore, we
select ( )Q ni as the other input variables. Consequently, the input
variables of the fuzzy controller are state variables of the state-
space system given in (38) and (39). Also, the control variables
( )S ni are the output of the fuzzy controller.
The second step in designing of a fuzzy controller is fuzzification.
The fuzzification is the process of making a crisp quantity fuzzy. The
input variables ( )W ni and ( )Q ni for =i 1, . . , 4 are divided into three
fuzzy membership functions: “Low(L),” “Medium(M),” and “High(H)”
and the output variables ( )S ni are divided into two set: “Going(G)” and
“Stopping(S)”. Going is the continuation of green phase and Stopping
is the red phase. Best locations of the left and right feet or base
points of the triangle and also best location of the triangle peak are
set by MBSA. In other words, the MBSA optimization is performed to
compute the lower and upper MFs of the type-2 MF. To do this, the
MBSA is carried out based on the fitness function (37), the initial
population given in section III, and random values for ( )q ni
[35].
The next step in designing of a fuzzy controller is determina-
tion of fuzzy rules. A more precise analyze of the single intersec-
tion traffic problem in Fig. 4, reveals that the traffic lights of the
legs 1 and 3 (traffic light pair S1- S3) and the legs 2 and 4 (traffic
light pair S2-S4) should change simultaneously and conversely. In
other word, when the pair (S1-S3) changes to green, the other pair
(S2-S4) turns to red. Furthermore, the two pairs (S1-S3) and (S2-S4)
are dependent to each straight road overall queue length and
waiting time (i.e. ( )+ ( )Q n Q n1 3 , ( )+ ( )Q n Q n2 4 , ( )+ ( )W n W n1 3 , and
( )+ ( )W n W n2 4 ). Consequently, one fuzzy controller is needed for
these two pairs. The the jth fuzzy rule the controller is as follows:
{ }
{ } { }
{ } { }
+ +
+ +
R If Q Q is F Q Q is F
is and is
Then S and S are G and S and S are not G
: , ,
W W F W W F
j j j
j j
1 3 1 2 4 2
1 3 3
j
2 4 4
j
1 3 2 4
The parameters + +W Wi i 2 and + +Q Qi i 2 for =i 1,2 are specified
by fuzzy membership functions as demonstrated in Figs. 5–7.
Consequently, the number of fuzzy rules for the controller is
computed as × × × =3 3 3 3 81. Some examples of fuzzy rules of
the controller are provided in Table 1. Another part of the fuzzy
controller is inference engine. An inference engine is a computer
program that attempts to obtain answers from a Rule-base. It is
the “brain” that expert systems use to reason about the data in the
Rule-base for the ultimate purpose of formulating new results.
The last part of designing a fuzzy controller is defuzzification.
The defuzzification is the conversion of a fuzzy quantity to a
precise quantity, just as fuzzification is the conversion of a precise
quantity to a fuzzy quantity. But before applying the defuzzifica-
tion, in type-2 fuzzy sets and systems we have another part,
namely type-reducer. A type-reducer is required to convert them
Fig. 5. Fuzzy membership functions for +W W1 3, +W W2 4.

into a T1 FS before defuzzification can be performed. The precise
numerical amount of the automobiles that join the ith queue in
the nth time interval for the specified inputs is computed using Eq.
(37). General scheme of the proposed controller is shown in Fig. 8.
5. Discussion of results
The proposed controller can appropriately control traffic
stream under both regular and abnormal traffic circumstances. In
simulation, the sampling time is considered 5 s (T ¼ 5), while a
full cycle time of traffic light is considered 100 s. Simulation has
been done with the following assumptions:
a) The crossroad is composed of four ways, and each way consists
of three lanes.
b) The entrance of automobiles into crossroad is independent on
each lane.
c) Pedestrian crossing at the crossroad is embedded.
d) Sensors are located at specified distance from the crossroad,
the highest number of automobiles in the queue that can be
identified by sensors, is 30 automobiles.
e) Maximum and minimum time to cross the crossroad is re-
spectively 40 and 5 s.
The number of automobiles that leave the ith queue in the nth
time interval is obtained by the following equation
( )( )= ( ) + ( ) ( ) ( )d n Q n q n d nmin , 41i i i si
So that saturation stream rate is
β( ) = ( )+ ( ) ( )d n d n q n 42si cons i.
for =i 1,2,3,4. The dcons. parameter is greater equal fifty, ( ≥d 50cons. ).
The β parameter is in the interval [0 1], so that it’s changes are
listed in Table 2. In fact, the β parameter indicates the level of
uncertainty, where β = 0 corresponds to the highest possible un-
certainty, while β = 1 corresponds to the situation in which the
exact travel time through the network is known with complete
certainty. The traffic data was saved every 5 s and was utilized in
the simulations. The qi and di variables are random signals with
normal distribution.
5.1. Results of fixed-time monitoring
The traffic signals of Leg 1 and Leg 3 in Fig. 4 were supposed
green in 140 s and red in after 60 s. On the other side, the traffic
signals of Leg 2 and Leg 4 in Fig. 4 were supposed red in 140 s and
green in after 60 s. The purposes of optimization are length re-
duction of queues and reducing the average waiting time of au-
tomobiles in the queues. The total numbers of automobiles in
queues on crossroad were shown in every 5 s in Fig. 9. The si-
mulation time is 1000 s.
Fig. 6. Fuzzy membership functions for +Q Q1 3, +Q Q2 4.
Fig. 7. Fuzzy membership functions for S1, S2, S3, and S4.
Table 1
Example of fuzzy rules of the proposed controller.
Rule j F j
1 F j
2 F j
3 F j
4 G j
1 Low Low Low Low Going
2 Low Low Low Medium Stopping
3 Low Low Low High Stopping
4 Low Low Medium Low Going
5 Low Low Medium Medium Going
19 Low Medium Medium Low Stopping
51 Medium High High Low Stopping
79 High High High Low Going
80 High High High Medium Going
81 High High High High Going
… … … … … …
Fig. 8. General scheme of the proposed controller.
Table 2
Level of uncertainty presented by β parameter.
The Position of Traffic β
Non Saturation β ≥ 0.7
Saturation 0.4 β≤ ≤ 0.6
Super Saturation 0.1 β≤ ≤ 0.3
Instable β=0

5.1.1. Results of fuzzy intelligent control
The output of controller is the set of control variables (Si). These
control variables for the Leg 1 and Leg 3 of crossroad were shown
in Fig. 10. Also, the total numbers of automobiles in queues on
crossroad, when the fuzzy controller has been used, were shown
in every 5s in Fig. 11. The simulation time is 1000 s.
In order to show the superiority of proposed method, the
Table 3 is presented, where it can be observed that in most tests
the proposed model has a better performance compared with
other methods. The results in this section show that the
non-optimal and optimal type-1 fuzzy controllers and the interval
Type-2 fuzzy controller can still work but there exists room for
improvement. On the other hand, the Optimal General Type-2
Fuzzy Controller resulted on length reduction of queues and re-
duction of the average waiting time of automobiles in the queues.
The proposed strategy can achieve control goals with faster re-
sponse and smoother manner. The proposed method can increase
capacity of the road network in acceptable time and decrease
congestion. Moreover, the computing time of proposed control
strategy is less than 5sec.
6. Conclusion
In this paper an Optimal General Type-2 Fuzzy Controller
(OGT2FC) was proposed for a single crossroad. The parameters of
input and output membership functions were optimized si-
multaneously by a novel heuristic algorithm namely Modified
Backtracking Search Algorithm (MBSA). The overall purpose of the
optimization was length reduction of queues and reduction of the
average waiting time of automobiles in the queues. Simulation
results show the superiority of the proposed controller to the non-
optimal and optimal type-1 fuzzy and the interval type-2 con-
trollers (Table 3).
Acknowledgements
The authors would like to thank the reviewers for their very
useful comments, which help us to improve the quality of the
paper. The reviewers' efforts are gratefully appreciated. The au-
thors wish to thank Prof. Jerry M. Mendel and Dr. Maryam Ho-
mayounzadeh for their assistance in the present work.
References
[1] B. De Schutter, Optimal traffic light control for a single intersection. In: Pro-
ceedings American control conference, p. 2195–2199; 1999.
[2] Diakaki C, Papageorgiou M, Aboudolas K. A multivariable regulator approach
to traffic-responsive network-wide signal control. Control Eng Pract
2002;10:183–95.
[3] K. Iwaoka, T. Otokita and S. Niikura, Optimization strategy on urban signal
control. In: Proceedings the 8th world congress on ITS ’01; 2001.
[4] Oda T, Leo J. Optimization of coordinated traffic signal timings in urban road
network. Trans ISCIE 1998;11(7):364–74 [in Japanese].
[5] L.A. Klein, Traffic parameter measurement technology evaluation. In: Proceedings
of the IEEE-IEE vehicle navigation and information systems conference, 1993, pp.
529–533.
[6] F. Woelk, S. Gehrig, and R. Koch, A monocular image based intersection as-
sistant”. Intelligent vehicles symposium 2004 , IEEE, 2004, pp. 286–291.
[7] Chen XF, Shi ZK. A dynamic optimization method for traffic signal timings
based on genetic algorithm. [Chinese Association for System Simulation.Beij-
ing]. J Syst Simul 2004;16(6):1155–61.
[8] X.F. Chen, Z.K. Shi, Real-coded genetic algorithm for signal timing optimization
of a single intersection. In: Proceedings of 2002 international conference on
machine learning and cybernetics, vol 3, 2002, pp.1245–1248.
[9] S. Takahashi, H. Nakamura, H. Kazama, and T. Fujikura, Genetic algorithm
approach for adaptive offset optimization for the fluctuation of traffic flow. In:
IEEE Proceedings of the 5th International Conference on Intelligent
Fig. 9. The total number of automobiles in queues on crossroad without controller.
Fig. 10. The control variables for Leg 1 and Leg 3.
Fig. 11. The total number of automobiles in queues on crossroad with (OT2FC).
Table 3
The Comparison between average waiting in any lane of (OGT2FC).
Average Waiting
Time (s)
Queue 1 Queue 2 Queue 3 Queue 4 Summation of
Queues
Fixed Time Control 11,000 s 5000 s 11,000 s 5000 s 32,000 s
The Controller
Presented in [30]
1350 s 1335 s 1300 s 1314 s 5300 s
Optimal Type-1
Fuzzy Control
1215 s 1295 s 1250 s 1273 s 5033 s
Interval Type-2
Fuzzy Control
1101 s 1173 s 1127 s 1149 s 4550 s
Optimal Type-2
Fuzzy Control
1005 s 985 s 910 s 930 s 3830 s
Improvement
Percentage
90.86 80.3 91.72 81.4 88.125

Transportation Systems, 2002, pp. 768–772.
[10] Y. Wakasa, K. Iwaoka and K. Tanaka, Modelling and robust control of traffic
signal systems. In: Proceedings European Control Conference 2003 (CD-ROM),
2003.
[11] Y. Wakasa, K. Hanaoka, K. Iwaoka and K. Tanaka, Controllability of network-
wide traffic signal systems and state-feedback controller tuning. In: Proceed-
ings 11th World Congress on ITS Nagoya, Aichi 2004, Japan (CD-ROM) 2004.
[12] Chinyere OU, Francisca OO, Amano OE. Design and simulation of an intelligent
traffic control system. IJAET 2011;1(Issue 5):47–57.
[13] Mendel J. Uncertain Rule-Based Fuzzy Logic Systems: Introduction and New
Directions. Upper Saddle River, NJ: Prentice-Hall; 2001.
[14] John R. Type 2 fuzzy sets: an appraisal of theory and applications. Int J Un-
certain Fuzziness Knowl Based Syst 1998;6(6):563–76.
[15] Chiclana F, Zhou S-M. Type-reduction of general type-2 fuzzy sets: the type-1
OWA approach. Int J Intell Syst 2013;28(5):505–22.
[16] Ghaemi Mostafa, Hosseini-Sani Seyyed Kamal, Khooban Mohammad Hassan.
Direct adaptive general type-2 fuzzy control for a class of uncertain non-linear
systems. IET Sci, Meas Technol 2014;8(6):518–27.
[17] Zadeh LA. The concept of a linguistic variable and its application to approx-
imate reasoning–I. Inf Sci 1975;8(3):199–249.
[18] Khooban Mohammad, Hassan, Niknam Taher. A new and robust control
strategy for a class of nonlinear power systems: adaptive general type-II fuzzy.
Proc Inst Mech Eng, Part I: J Syst Control Eng 2015 [0959651815571621].
[19] Liu F. An efficient centroid type-reduction strategy for general type-2 fuzzy logic
system. IEEE Comput Intell Soc Walter J Karplus Summer Res Grant Rep 2006.
[20] Wagner C, Hagras H. Toward general type-2 fuzzy logic systems based on
zSlices. Fuzzy Syst, IEEE Trans 2010;18(4):637–60.
[21] Bilgin, A., Hagras, H., van Helvert, J. and Alghazzawi, D., A linear general type-2
fuzzy logic based computing with words approach for realising an ambient
intelligent platform for cooking recipes recommendation.
[22] Kumbasar T, Hagras H. A self-tuning zSlices-based general type-2 fuzzy PI
controller. Fuzzy Syst IEEE Trans 2015;23(4):991–1013.
[23] Mendel JM. Uncertain rule-based fuzzy logic systems: introduction and new
directions. Upper Saddle River, NJ: Prentice-Hall; 2001.
[24] Liu F. An efficient centroid type-reduction strategy for general type-2 fuzzy
logic system. Inf Sci 2008;178(9):2224–36.
[25] Karnik NN, Mendel JM. Centroid of a type-2 fuzzy set. Inf Sci 2001;132:195–
220.
[26] Mendel JM, Wu D. Perceptual computing: aiding people in making subjective
judgments. Hoboken, NJ: Wiley-IEEE Press; 2010.
[27] Wu D, Mendel JM. Enhanced karnik-mendel algorithms. IEEE Trans Fuzzy Syst
2009;17(4):923–34.
[28] Civicioglu P. Backtracking search optimization algorithm for numerical opti-
mization problems. Appl Math Comput 2013;219(15):8121–44.
[29] Shaw B, Mukherjee V, Ghoshal SP. A novel opposition-based gravitational
search algorithm for combined economic and emission dispatch problems of
power systems. Int J Electr Power Energy Syst 2012;35(1):21–33.
[30] Azimirad E, Pariz N, Naghibi Sistani MB. A novel fuzzy model and control of
single intersection at urban traffic. [March]. Netw, IEEE Syst J 2010;4(1):107–
11 [March].
[31] Khooban Mohammad, Hassan Davood Nazari Maryam, Abadi Alireza, Alfi,
Siahi Mehdi. Swarm optimization tuned Mamdani fuzzy controller for dia-
betes delayed model. Turk J Electr Eng Comput Sci 2013;21(Suppl 1):2110–26.
[32] Khooban Mohammad Hassan. Design an intelligent proportional-derivative
PD feedback linearization control for nonholonomic-wheeled mobile robot. J
Intell Fuzzy Syst: Appl Eng Technol 2014;26(4):1833–43.
[33] Alfi Alireza, Kalat Ali Akbarzadeh, Khooban Mohammad Hassan. Adaptive
fuzzy sliding mode control for synchronization of uncertain non-identical
chaotic systems using bacterial foraging optimization. J Intell Fuzzy Syst
2014;26(5):2567–76.
[34] Modiri-Delshad M, Rahim NA. Solving non-convex economic dispatch pro-
blem via backtracking search algorithm. Energy 2014;77:372–81.
[35] Talab Hamid, Saadat Hadiseh, Mohammadkhani, Haddadnia Javad. Controlling
multi variable traffic light timing in an isolated intersection using a novel
fuzzy algorithm. J Intell Fuzzy Syst 2013;25(1):103–16.

An optimal general type-2 fuzzy controller for Urban Trafﬁc Network

More Related Content

What's hot (20)

Similar to An optimal general type-2 fuzzy controller for Urban Trafﬁc Network (20)

More from ISA Interchange (20)

Recently uploaded (20)

An optimal general type-2 fuzzy controller for Urban Trafﬁc Network