Expert Systems with Applications 39 (2012) 4390–4401

Contents lists available at SciVerse ScienceDirect

Expert Systems with Applications

journal homepage: [Link]/locate/eswa

Comparative performance analysis of various binary coded PSO algorithms

in multivariable PID controller design
Muhammad Ilyas Menhas ⇑, Ling Wang, Minrui Fei, Hui Pan
School of Mechatronics Engineering and Automation, Shanghai University, Shanghai, China

a r t i c l e i n f o a b s t r a c t

Keywords: In this paper, comparative performance analysis of various binary coded PSO algorithms on optimal PI
PID control and PID controller design for multiple inputs multiple outputs (MIMO) process is stated. Four algorithms
Swarm intelligence such as modified particle swarm optimization (MPSO), discrete binary PSO (DBPSO), modified discrete
Binary PSO binary PSO (MBPSO) and probability based binary PSO (PBPSO) are independently realized using MATLAB.
Particle swarm optimization
The MIMO process of binary distillation column plant, described by Wood and Berry, with and without a
PID tuning
decoupler having two inputs and two outputs is considered. Simulations are carried out to minimize two
objective functions, that is, time integral of absolute error (ITAE) and integral of absolute error (IAE) with
single stopping criterion for each algorithm called maximum number of fitness evaluations. The simula-
tion experiments are repeated 20 times with each algorithm in each case. The performance measures for
comparison of various algorithms such as mean fitness, variance of fitness, and best fitness are computed.
The transient performance indicators and computation time are also recorded. The inferences are made
based on analysis of statistical data obtained from 20 trials of each algorithm and after having compar-
ison with some recently reported results about same MIMO controller design employing real coded
genetic algorithm (RGA) with SBX and multi-crossover approaches, covariance matrix adaptation evolu-
tion strategy (CMAES), differential evolution (DE), modified continuous PSO (MPSO) and biggest log mod-
ulus tuning (BLT). On the basis of simulation results PBPSO is identified as a comparatively better method
in terms of its simplicity, consistency, search and computational efficiency.
Ó 2011 Elsevier Ltd. All rights reserved.

1. Introduction responses. In fact, optimal control performance can only be

achieved after identifying the finest set of three gains, that is, pro-
Despite numerous advancements in process control methodolo- portional gain (Kp), integral gain (Ki) and derivative gain (Kd). Many
gies, Proportional–Integral–Derivative (PID) control is still the most approaches have been reported in literature for tuning parameters
efficient and widely used feedback control strategy. This is due to its of PID controller. The conventional PID tuning techniques include
simplicity and satisfactory control performance. PID controller was Z–N, Cohen Coon, and relay feedback methods (Cohen & Coon,
introduced in 1910 and its use and popularity had grown particu- 1953; Ziegler & Nichols, 1942). The modern techniques are based
larly after the Ziegler–Nichols empirical tuning rules in 1942 on artificial intelligence techniques such as neural network, fuzzy
(Åström & Hägglund, 2001; Ziegler & Nichols, 1942). The develop- logic and evolutionary computation; these are the most recent
ment in artificial intelligence and digital technology have resulted techniques (Astrom & Hagglund, 1995).
in many intelligent control schemes such as fuzzy logic control Recently, many attempts have been made by several research-
(Goshal, 2004; Lee, 1990), neural network control (Fukuda & ers to tune the PID controller parameters using various EAs, such
Shibata, 1992) and adaptive control (Astrom & Wittenmark, 1995; as genetic algorithm (GA), covariance matrix adaptation evolution
Zuo, 1995). But no other technique could replace PID algorithm strategy (CMAES), particle swarm optimization (PSO), differential
and more than 90% of industrial controllers are still based on PID evolution (DE), tribes algorithm (TA), ant colony optimization
control (Ang, Chang, & Li, 2005). The wide use of PID control has sus- (ACO), and discrete binary particle swarm optimization (DBPSO)
tained research on finding the key methodology for PID tuning to ob- (Bingul, 2004; Chang, 2007, 2009; Chen, Cheng, & Lee, 1995;
tain best possible performance out of the PID control (Marsh, 1998). Coelho & Bernert, 2009; Duan, Wang, & Yu, 2006; Gaing, 2004;
The optimally combined three terms functioning of PID control- Jan, Tseng, & Liu, 2008; Kim, Maruta, & Sugie, 2008; Menhas,
ler can provide treatment for both the transient and steady state Wang, Fei, & Ma, 2011; Mukherjee & Goshal, 2007; Wang, Zhang,
& Wang, 2006; Willjuice & Baskar, 2009, 2010; Zhang, Zhuang,
⇑ Corresponding author. Tel.: +86 15821107474. Du, & Wang, 2009; Zuo, 1995) for both the single and multi-
E-mail address: ilyasminhas75@[Link] (M.I. Menhas). variable processes.

0957-4174/$ - see front matter Ó 2011 Elsevier Ltd. All rights reserved.
doi:10.1016/[Link].2011.09.152
 M.I. Menhas et al. / Expert Systems with Applications 39 (2012) 4390–4401 4391

AI-based evolutionary computational techniques can determine 1995). It works by having a group of m particles. Each particle
the most optimal sets of controller gains based on a given objective can be considered as a candidate solution to an optimization prob-
function in an iterative manner from thousands of possible alter- lem and it can be represented by a point or a position vector xij =
nate solutions that best fit the designer’s requirements. But the [xi1, . . ., xid] in a d dimensional search space which keeps on moving
performance of different methods may significantly vary in differ- toward new points in the search space with the addition of a veloc-
ent applications. In Willjuice and Baskar (2009) comparative per- ity vector vij = [vi1, . . ., vid] to further facilitate the search procedure.
formance analysis of various EAs such as real coded genetic The initial positions and velocities of particles are random from a
algorithm (RGA) with SBX crossover, differential evolution (DE), normal population u 2 [0, 1]. All particles move in the search space
modified particle swarm optimization (MPSO) and covariance ma- to optimize an objective function f(x). Each member of the group
trix adaptation evolution strategy (CMAES) was done and better gets a score after its evaluation on objective function f(x); the score
performance of CMAES and MPSO in comparison to BLT, RGA with is regarded as a fitness value. The member with the highest score is
SBX and multi-crossover approaches, was reported in the paper. called global best. Each particle memorizes its previous best posi-
The MPSO algorithm is a variant of real coded particle swarm opti- tions. During the search process all particles move toward the
mization algorithm. areas of potential solutions by utilizing the cognitive and social
The particle swarm optimization algorithm (PSO) was intro- learning components. The process is repeated until any prescribed
duced by Kennedy and Eberhart (1995) by simulation of swarms stopping criterion is reached. After any iteration, each particle up-
behavior in performing their tasks. The PSO has several advantages, dates its position and velocity to achieve better fitness values
it works by maintaining a population of solutions and hence allows according to the following Eqs. (1) and (2)
for parallel evaluations of several solutions, it does not require that
the optimization problem must be differentiable and comprises V ij ðt þ 1Þ ¼ w  V ij ðtÞ þ c1  r 1 ðpij ðtÞ  xij ðtÞÞ þ c2  r 2 ðg 1j ðtÞ  xij ðtÞÞ ð1Þ
very simple mathematics. The PSO’s simplicity and capability of xij ðt þ 1Þ ¼ V ij ðt þ 1Þ þ xij ðtÞ ð2Þ
solving very difficult problems have motivated many researchers
for its further development. Some recent developments can be where c1, c2 are two constants, called acceleration factors; w is iner-
seen in Baskar and Suganthan (2004), Zhao and Suganthan tia weight; V(t) is velocity of each particle during iteration t; x(t) is
(2009), Van den Bergh and Engelbrecht (2004) and Zhan et al. current position of particle at iteration t; p(t) is previous best posi-
(2009). tion of each particle till t; g(t) is the position of the best particle in
Although continuous PSO and most of its variants have been the group; r1, r2 are two quasi random numbers u 2 [0, 1]; t is the
successfully applied in many real-world engineering applications, current iteration time or index.
it is widely held that PSO gets trapped in local optima. Further- The pseudo code of PSO algorithm for the multivariable PID
more, many of proposed strategies to deal with the weaknesses controller design is provided in Fig. 1.
of continuous PSO algorithm have significantly increased the com-
putational costs. 2.2. Discrete binary PSO (DBPSO) algorithm
In 1997, Kennedy and Eberhart further extended the continuous
PSO algorithm to deal with the combinatorial optimization prob- Kennedy and Eberhart introduced a discrete version of the PSO
lems and proposed a discrete binary version of PSO. Unlike contin- aiming to deal with combinatorial optimization problems. In dis-
uous PSO, the discrete version of PSO (DBPSO) uses binary bits to crete binary PSO (DBPSO) (Kennedy & Eberhart, 1997), particle po-
represent each dimension of particle position vector. The binary sition vector is represented by a binary string with each
coded PSO algorithm can cover a wide range of applications as bin- component of this vector bounded to have a value either one or
ary sequences can be transformed to match the requirements of zero. In DBPSO, velocity update rule Eq. (1) of the PSO was pre-
any problem space. served, however, it was considered as a pseudo probability for
With the purpose of finding a suitable tuning technique, this pa- any component of position vector to take a value in the binary do-
per is focused on the performance evaluation of various binary
coded PSOs such as DBPSO, MBPSO, and PBPSO algorithms on opti-
mal design of multivariable PI and PID controllers with and with-
out decoupling of process described by Wood and Berry (1973),
Chang (2007) and Willjuice and Baskar (2009) which is considered
as a case study for the MIMO tuning problem.
In addition, the comparative performance analysis of computa-
tional techniques is useful in updating and integrating current
developments for further research and development.
The remaining paper is organized as follows: Section 2 de-
scribes various methods under consideration, Section 3 illustrate
implementation of proposed methods, Section 4 details experi-
ments and simulation results, finally conclusions are drawn in Sec-
tion 5.

2. Techniques

This section briefly explains the PSO and some binary PSO
variants.

2.1. Particle swarm optimization (PSO)

The PSO was introduced by Kennedy and Eberhart by simulat-

ing social behavior of birds flocks in 1995 (Kennedy & Eberhart, Fig. 1. Pseudo code of PSO for multivariable PID controller design.
4392 M.I. Menhas et al. / Expert Systems with Applications 39 (2012) 4390–4401

main as ‘0’ or ‘1’. The overall procedure of the DBPSO can be ex-
plained as follows:

Step 1. Initialize the swarm.

(i) Assign values for inertia weight W, acceleration factors C1
and C2
(ii) Set upper and lower limits of Vmax and Vmin = Vmax.
(iii) The initial position, initial velocity associated with ith parti-
cle can be given as
xbi ¼ ½xbi1 ; xbi2 ; . . . ; xbiD  ð3Þ
V i ¼ ½V i1 ; V i2 ; . . . ; V iD  ¼ ½randðÞ; randðÞ; . . . ; randðÞ ð4Þ
(iv) Generate an initial population of candidate solutions
fxbi ¼ ½xbi1 ; xbi2 ; :::; xbiD g comprising N particles in a D dimen-
sional search space subject to initial velocities {Vi = [Vi1, -
Vi2, . . ., ViD] = [rand(), rand(), . . ., rand()]} as
8
< 1 if
>
1
xbij ¼ randðÞ 6 Sij subject to Sij ¼ and f0 6 S 6 1g
>
: ð1 þ expðv ij ÞÞ
0 else
ð5Þ

Step 2. Evaluate each particle (candidate solution) on pre-

scribed fitness function f(x).
Step 3. Set the current position xb(t) of each particle as its histor-
ical best position pxb(t) based on their current fitness and choose
the best particle from the group based on its fitness as gxb(t).
Step 4. Update the velocity (pseudo probability) of each particle
as Fig. 2. Pseudo code of the DBPSO for multivariable PID controller design.

V ij ðt þ 1Þ ¼ w  V ij ðtÞ þ c1  r 1 ðpxbij ðtÞ  xbij ðtÞÞ þ c2

3. In place of sigmoid mapping it employs bounded linear trans-
 r2 ðgxb1j ðtÞ  xbij ðtÞÞ where fV min formations function to estimate the actual probability
6 V 6 V max g ð6Þ p 2 [0, 1] for a bit to assume a value.
4. At each time step a velocity component is added into past
Step 5. Generate new population of candidate solution accord- pseudo probability of any bit after that the pseudo probability
ing to relation (5). is transformed to the real probability by a linear transformation
Step 6. Evaluate each particle (candidate solution) on pre- function.
scribed fitness function f(x).
Step 7. Update history of each particle by comparing its current The PBPSO algorithm can be stated as follows:
position with its historical best position based on fitness.
Step 8. Renew the global best. Step 1. Initialize the swarm.
Step 9. If termination a criterion has been reached then output (i) Assign values for inertia weight W, acceleration factors C1
the global best solution. and C2.
(ii) Set upper and lower limits of Xmax and Xmin as well as Vmax
The discrete binary PSO (DBPSO) can cover a range of applica- and Vmin as Xmin = Xmax, Vmin = Vmax and Xmax = Vmax.
tions of diverse nature as the binary strings can be transformed (Note it is identical to setting bounds on the pseudo prob-
to match the requirements of any problem space. The algorithm ability Vmax in original discrete binary PSO (DBPSO)
is inherently capable of avoiding local optimum by making use of algorithm.)
bit flip mechanism. The pseudo code of the DBPSO algorithm for (iii) The initial position, initial velocity and initial pseudo
the multivariable PID controller design is provided in Fig. 2. probability vector associated with ith particle can be
given as
2.3. Probability based discrete binary PSO (PBPSO) algorithm
xbi ¼ ½xbi1 ; xbi2 ; :::; xbiD  ð7Þ
The probability based discrete binary PSO (PBPSO) algorithm V i ¼ ½V i1 ; V i2 ; :::; V iD  ¼ ½0; 0; :::; 0 ð8Þ
(Menhas et al., 2011; Wang, Wang, Fu, & Zhen, 2008) is a variant xi ¼ ½xi1 ; xi2 ; :::; xiD  ¼ ½0; 0; :::; 0 ð9Þ
of discrete binary PSO (DBPSO) algorithm. It is conceptually identi-
cal to the DBPSO algorithm except it differs in the following (iv) Generate an initial population of candidate solution vec-
aspects: tors fxbi ¼ ½xbi1 ; xbi2 ; :::; xbiD g comprising N particles in a D
dimensional binary search space subject to initial veloci-
1. It preserves both the velocity as well as position update rule of ties {Vi = [Vi1, Vi2, . . ., ViD] = [0, 0, . . ., 0]} as
8
PSO algorithm.
< 1 if
>
xij  X min
2. Instead of velocity (V) the particle position (X) is considered as a xbij ¼ rand 6 Pij subject to Pij ¼ and f0 6 P 6 1g
pseudo probability to determine the actual probability p 2 [0, 1] >
: X max  X min
0 else
for actual state of a bit (each component of the solution vector
in binary space). ð10Þ
 M.I. Menhas et al. / Expert Systems with Applications 39 (2012) 4390–4401 4393

8
Because, the initial value of x is zero, the initial bits are gener- >
> xi ðtÞ if 0 6 V 6 a
>
<
ated with P = 0.5.
Step 2. Evaluate each particle (candidate solution) on pre- xi ðt þ 1Þ ¼ pi ðtÞ if a 6 V 6 1=2ð1 þ aÞ ð13Þ
>
>
scribed fitness function f(x). >
:
g i ðtÞ if 1=2ð1 þ aÞ 6 V 6 1
Step 3. Set the current position xb(t) of each particle as its his-
torical best position pxb(t) based on their current fitness and
choose the best particle from the group based on its fitness as The modified approach replaced all the original equations of basic
gxb(t). DBPSO and proposed entirely new updating formulas. In MBPSO
Step 4. Update the velocity and the position (pseudo probabil- there are no acceleration coefficients or inertia weight. The initial
ity) of each particle as solutions are randomly generated. The velocity V is constrained in
the interval V 2 [0, 1]. The update is performed according to above
V ij ðt þ 1Þ ¼ w  V ij ðtÞ þ c1  r 1 ðpxbij ðtÞ  xbij ðtÞÞ þ c2  r 2 ðgxb1j ðtÞ  xbij ðtÞÞ equation (13). Where, a is a predefined static probability fixed as
a constant value in the range [0, 1] usually 0.5. Now, a comparison
where fV min 6 V 6 V max g
between V and a randomly produced value V 2 [0, 1] will determine
ð11Þ
any bit’s state as in Eq. (13).

xij ðt þ 1Þ ¼ xij ðtÞ þ V ij ðt þ 1Þ

ð12Þ
where fX min 6 x 6 X max g
3. Implementation
Step 5. Generate new population of candidate solution accord-
ing to relation (10). This section details the implementation of proposed methods.
Step 6. Evaluate each particle (candidate solution) on pre-
scribed fitness function f(x). 3.1. PID controller
Step7. Update history of each particle by comparing its current
position with its historical best position based on fitness. The three terms continuous PID controller transfer function can
Step 8. Renew the global best. be given as
Step 9. If a termination criterion has reached, output the global
CðsÞ ¼ K p þ K i =s þ K d s ð14Þ
best solution.
where Kp, Ki, and Kd represent proportional gain, integral gain and
The pseudo code of the PBPSO for the multivariable PID control- derivative gain respectively.
ler design is provided in Fig. 3.
 The proportional term provides a control action proportional to
the error and reduces rise time.
2.4. Modified discrete binary PSO (MBPSO) algorithm  The integral term reduces steady state error by performing an
integral control action based on past errors.
In Shen Qi (2004), a modified discrete binary PSO (MBPSO) was  The derivative term improves stability of the system and
proposed. In MBPSO, the updating formulas were given by follow- reduces overshoot with the prediction of future.
ing equations
3.2. Process model

Two inputs two outputs transfer function matrix of the binary

distillation column plant (Willjuice & Baskar, 2010; Wood & Berry,
1973) is given as below:

  " #
12:8es 18:9e3s
P11 ðsÞ P12 ðsÞ 16:7sþ1 21sþ1
PðsÞ ¼ ¼ ð15Þ
P21 ðsÞ P22 ðsÞ 6:6e7s 19:4e3s
16:9sþ1 14:4sþ1

The steady-state decoupling matrix of the above process model can

be given as

   
D11 D12 0:1570 0:1529
DðsÞ ¼ P1 ð0Þ ¼ ¼ ð16Þ
D21 D22 0:0534 0:1036

3.3. Matlab Simulink based control structures

The Matlab/Simulink based control structures adapted for com-

parative study of various methods are shown in the following Figs.
4–6 respectively.

3.4. Problem formulation

In the design of PID controller using optimization based tech-

niques, the performance criterion or objective function need to
be specified. Therefore, the following two objective functions were
Fig. 3. Pseudo Code of the PBPSO for multivariable PID controller design. considered based on the time-integral performance criterion.
4394 M.I. Menhas et al. / Expert Systems with Applications 39 (2012) 4390–4401

Fig. 4. Matlab Simulink based multivariable PID control structure of binary distillation column plant with a steady state decoupler.

Fig. 5. Matlab Simulink based multivariable PID control structure of binary distillation column plant without a decoupler.
 M.I. Menhas et al. / Expert Systems with Applications 39 (2012) 4390–4401 4395

Fig. 6. Matlab Simulink based multivariable PI control structure of binary distillation column plant without a decoupler.

The objective function F1 is based on minimizing the integral of

absolute error IAE and F2 is based on minimizing the integral of time
weighted absolute error ITAE.

3.5. PID controller parameter representation

Fig. 7. Particle representing PID parameters in binary space. The PID controller parameters [Kp, Ki, Kd] which represent
dimensions of particles were set as 16 bit binary string in the case
of binary coded PSOs comprises 96 bits for both controllers and
these bits were converted to corresponding real numbers before
simulating controllers response and fitness computing while in
the case of the PSO each particle dimension was a real number rep-
resenting one of the PID parameters as in the Figs. 7 and 8
Fig. 8. Particle representing PID parameters in real space. respectively.

0 1 3.6. Algorithmic parameter tuning

X2 BZ T C
F1 ¼ B jei ðtÞj dtC ð17Þ The parameter study is very important which plays significant
@ A
t¼0 role in the PSO performance and inappropriate parameter values
i¼1 |fflfflfflfflfflffl
ffl{zfflfflfflfflfflfflffl}
0
IAE
1 bring an undesirable effect on PSO performance. In PSO the balance
Z between the local and global exploration can be achieved by prop-
X2 B T C
F2 ¼ B t:jei ðtÞj dt C ð18Þ er adjustment of inertia weight W and by imposing limits on par-
@ A ticle velocities on each dimension to a maximum velocity Vmax so
t¼0
i¼1 |fflfflfflfflfflfflfflffl{zfflfflfflfflfflfflfflffl}
ITAE that particle do not jump over the optimal rejoins and can main-
tain balance between local and global exploration abilities. The ori-

Table 1
Parameter values of various methods.

Method Algorithmic parameters

W = inertia weight D = real dimension N = population size L = binary bits C1 C2 Wmax Wmin Vmax Tmax
PSO Linearly decreasing 6 40 – 1.0 1.0 0.9 0.2 0.5 200
DBPSO 0.8 6 40 16 1.0 1.0 0.8 0.8 5.0 200
PBPSO 0.8 6 40 16 2.0 2.0 0.8 0.8 50.0 200
4396 M.I. Menhas et al. / Expert Systems with Applications 39 (2012) 4390–4401

Table 2
PID controller parameters obtained using various methods.

Method PID parameters

F Kp1 Ki1 Kd1 Kp2 Ki2 Kd2 Fitness
MPSO-PID F1 2.0000 0.1038 0.9307 1.7503 0.1197 0.6528 213.4376
DBPSO-PID F1 1.8954 0.1293 1.4138 1.7357 0.1106 1.4375 258.8566
PBPSO-PID F1 1.9978 0.1121 0.5443 1.9990 0.1485 0.5619 159.5911
MPSO-PID F2 2.0000 0.1148 1.3810 2.0000 0.1264 1.4582 23.4101
DBPSO-PID F2 1.9879 0.0772 0.2271 1.8347 0.1800 0.1941 24.5121
PBPSO-PID F2 2.0000 0.1178 0.2493 1.9994 0.1250 0.2412 20.4109

ginal version of PSO did not have inertia weight. This improvement 4. Experiments and simulation results
was done by Shi and Eberhart (1998) hence a linearly decreasing
inertia weight was employed. Simulations were carried out using Core 2 Duo Processor
In addition the acceleration constants C1 and C2 control the bal- 2.2 GHz, 2 GB of RAM under identical conditions for each algorithm
ance between the social learning and cognitive learning of particles with same stopping criteria such as maximum number of functions
which also require cautious selection. In order to find good param- evaluations and Matlab software was used to simulate and verify
eters for each algorithm simulation trials were repeated with dif- the results.
ferent parameter combinations. The best parameters of each To study the comparative performance of various algorithms for
algorithm were chosen after extensive simulations and were set the multivariable PID controller design following three simulation
as listed in Table 1. experiments were performed.

Fig. 9. Output response Y1 and Y2 in experiment 1.

 M.I. Menhas et al. / Expert Systems with Applications 39 (2012) 4390–4401 4397

Table 3
PID controller parameters obtained using various methods.

Method PID parameters

F Kp1 Ki1 Kd1 Kp2 Ki2 Kd2 Fitness
PBPSO-PID F1 0.9976 0.0025 0.3450 0.0313 0.0078 0.0155 13.1976
DBPSO-PID F1 0.9892 0.0032 0.3238 0.0389 0.0071 0.0375 14.9125
MBPSO-PID F1 0.8508 0.0040 0.4061 0.0979 0.0157 0.0439 16.4461
RGA-SBX crossovera F2 1 0.0025 0.3860 0.0358 0.0073 0.0991 28.0441
MPSO-PIDa F2 1 0.0025 0.3872 0.0332 0.0073 0.0909 28.0478
DE-PIDa F2 0.9825 0.0026 0.4239 0.0563 0.0082 0.1563 22.8478
CMAESa F1 1 0.0025 0.3872 0.0332 0.0073 0.0909 28.0478
a
Data taken from Willjuice and Baskar (2009).

1. Design of multivariable PID controller with decoupler. Table 4

2. Design of multivariable PID controller without decoupler. PI controller parameters obtained using various methods.
3. Design of multivariable PI controller without decoupler.
Method PI controller parameters

In the design of multivariable PID controller with decoupler, ITAE F Kp1 Ki1 Kp2 Ki2 Fitness

and IAE were used as objective function, while in the design of PBPSO-PI F1 0.8261 0.0027 0.0117 0.0068 13.3833
multivariable PID and PI controller without decoupler minimiza- DBPSO-PI F1 0.8198 0.0019 0.0327 0.0085 14.2384
MBPSO-PI F1 0.8429 0.0024 0.0237 0.0079 13.8271
tion of IAE was considered as objective function. BLT-PIa F1 0.3750 0.0452 0.0750 0.0032 26.4023
RGA-multi- F1 0.9971 0.0031 0.0141 0.0071 13.5661
crossovera
RGA-SBX crossoverb F1 0.8622 0.0026 0.0135 0.0069 13.3834
4.1. Experimental results and discussion MPSO-PIb F1 0.8485 0.0026 0.0132 0.0069 13.3976
DEb F1 0.8485 0.0026 0.0132 0.0069 13.3936
CMAESb F1 0.8485 0.0026 0.0132 0.0069 13.3976
4.1.1. Experiment 1
a
In this experiment, design of the multivariable PID controller Data taken from Chang (2007).
b
Data taken from Willjuice and Baskar (2009).
with a steady state decoupler as in Fig. 4 was considered. Simula-
tion experiments were carried out using the PBPSO, DBPSO, and the
MPSO algorithm to minimize two objective functions i.e. the time
4.1.2. Experiment 2
integral of absolute error (ITAE) and integral of absolute error (IAE).
In this experiment, the design of multivariable PID controller to
The MBPSO algorithm being unable to find appropriate controller
minimize IAE as objective function without employing a decoupler
parameters over several trial runs was excluded in this experiment.
as in Fig. 5 was considered. Simulation experiments were repeated
The best simulation results among 20 independent trials with each
for 20 independent trials and PBPSO, DBPSO, and the MBPSO
algorithm such as the PID controller parameters and their corre-
algorithm were employed. The best simulation results of 20 con-
sponding ITAE and IAE values in both cases i.e. F1 and F2 are given
secutive independent trials such as PID parameters and corre-
in Table 2. The output responses Y1 and Y2 are shown in Fig. 9. From
sponding IAE values with each algorithm are given in Table 3.
Fig. 9, it is quite obvious that multivariable PBPSO-PID controller
For comparison purposes, already reported results (Willjuice &
exhibits better performance in comparison to other controllers.

Fig. 10. Output response Y1 and Y2 in experiment 2.

4398 M.I. Menhas et al. / Expert Systems with Applications 39 (2012) 4390–4401

Fig. 11. Output response Y1 and Y2 in experiment 3.

Baskar, 2009) in multivariable PID controller design for same

MIMO process employing MPSO, RGA with SBX crossover, DE,
and CMAES algorithm are directly taken, and these parameter val-
ues are also provided in Table 3. The controller output responses Y1
and Y2 are shown in Fig. 10. The IAE values of previously reported
results using MPSO, RGA with SBX crossover, DE, and CMAES algo-
rithm are provided accordingly.

4.1.3. Experiment 3
In this experiment, the PI control structure as in Fig. 6 was ta-
ken. The design of multivariable PI controller to minimize IAE as
objective function was conducted. Simulation experiments were
repeated for 20 independent trials using the PBPSO, DBPSO, and
the MBPSO algorithms. The best PI parameters and corresponding
IAE values obtained are given in Table 4. For comparison purposes,
already reported results in multivariable PI controller design for
same MIMO process employing MPSO, RGA with SBX crossover,
DE, and CMAES algorithm are also given in Table 4. The output re-
sponses Y1 and Y2 are shown in Fig. 11. From Fig. 11 multivariable
PBPSO-PI controller seems better in terms of transient response.
Fig. 12. Convergence behavior of various algorithms on objective function F1 in
experiment 1.

Table 5
Transient performance indices of various controllers.

Experiment Method Output response Y1 Output response Y2

F %Mp Tr 5%Ts %Mp Tr 5%Ts
Experiment 1 PBPSO F1 0.40 11.80 19.90 0.60 10.50 16.5
DBPSO F1 0.00 18.90 35.50 9.00 6.80 32.98
MPSO F1 1.38 13.8 23.70 1.49 10.95 19.10
PBPSO F2 0.40 12.11 20.80 5.10 7.90 21.50
DBPSO F2 3.90 13.90 22.00 1.00 12.50 21.80
MPSO F2 0.40 15.00 26.75 1.30 10.80 19.80
Experiment 2 PBPSO F1 1.40 1.60 5.00 12.00 2.70 18.00
DBPSO F1 1.00 2.85 5.80 14.70 3.10 16.50
CMAES F1 0.8059 2.90 11.90 8.70 4.85 22.50
MBPSO F1 10.40 2.60 12.80 63.50 4.00 23.50
Experiment 3 PBPSO F1 9.20 2.80 6.70 8.70 2.80 12.50
DBPSO F1 9.90 2.85 6.68 12.49 2.70 19.30
CMAES F1 10.8390 2.80 7.30 9.2414 2.90 12.60
MBPSO F1 9.80 2.90 6.65 18.40 2.80 19.5
 M.I. Menhas et al. / Expert Systems with Applications 39 (2012) 4390–4401 4399

requirements. However, the transient response performance of

the multivariable PID controllers for the decoupled process/pre-
compensated is much better. Additionally, the PBPSO-PID control-
ler outperforms other controllers.

4.3. Convergence analysis

The convergence curves and fitness comparisons of the PBPSO,

DBPSO, and MPSO algorithms on objective functions F1, and F2 dur-
ing 200 iterations in experiment 1 are shown in Figs. 12 and 13
respectively. In addition, the search and convergence behavior of
the PBPSO algorithm is also shown in Fig. 14. From Figs. 12 and
13 it can be seen that the binary coded PBPSO algorithm is consis-
tently avoiding local optima and it reaches the optimum solution
in significantly less number of functions evaluations. Fig. 14 shows
the exploration, exploitation, jumping out and convergence ability
of the PBPSO algorithm.

4.4. Statistical analysis

Fig. 13. Convergence behavior of various algorithms on objective function F2 in

In order to compare the performance of various methods and
experiment 1.
identify the most suitable method the statistical measures such as
mean fitness of solutions discovered by each method variance of fit-
ness, minimum computation time and success rate of various meth-
4.2. Comparison in terms of time response ods in obtaining a feasible solution during 20 independent trials of
each method in each experiment were computed. The data obtained
In order to compare performance of various controllers for time regarding statistical measures of various methods is summarized in
response, the transient performance indices such as rise time, over- Table 6. The binary coded PBPSO algorithm exhibits 100% success
shoot and settling time were computed in each experiment. The rate in all cases. The reliability index of the binary coded (PBPSO)
transient performance indices of various controllers are given in and real coded PSO (PSO) algorithms is much better than the DBPSO
Table 5. From Table 5 it is quite obvious that all of the controllers and the MBPSO algorithm in terms of mean fitness, variance of fit-
designed using various methods pass the transient response ness and success rate. In terms of CPU time, the binary coded PSOs

Fig. 14. Search and convergence behavior of PBPSO algorithm in experiment 2.

4400 M.I. Menhas et al. / Expert Systems with Applications 39 (2012) 4390–4401

Table 6
Performance measures of various methods.

Method Experiment Best fitness Statistical measures

Mean fitness Variance of fitness Minimum computation time (s) Success rate (%)
MPSO-F2 Experiment 1 213.4376 222.8706 32.9080 165.3423 100
DBPSO-F2 258.8566 277.4394 176.0583 179.2042 80
PBPSO-F2 159.5911 163.9377 9.6366 174.0134 100
MPSO-F1 23.4101 26.8276 2.4509 163.1367 100
DBPSO-F1 24.5121 28.7687 6.1593 178.6379 90
PBPSO-F1 20.4109 21.0613 0.1192 175.3215 100
PBPSO-F1 Experiment 2 13.1976 14.1286 0.2957 174.7921 100
DBPSO-F1 14.9125 17.6093 2.0958 177.9319 90
MBPSO-F1 16.4461 20.9938 8.5103 179.3013 80
PBPSO-F1 Experiment 3 13.3833 14.1457 0.1752 147.2212 100
DBPSO-F1 14.2384 16.5346 1.7616 151.0908 85
MBPSO-F1 13.8271 18.6057 8.9673 142.7143 80

require slightly higher computation time due to additional compu- Ang, K. H., Chang, G., & Li, Yun (2005). PID control system analysis, design and
technology. IEEE Transaction on Control System and Technology, 13(4), 559–577.
tations involved in the process of number conversions.
Astrom, K. J., & Hagglund, T. (1995). PID controllers: Theory, design and tuning (2nd
ed.). Instrument Society of America.
Astrom, K. J., & Wittenmark, B. (1995). Adaptive controls (2nd ed.). Addison Wesley.
5. Conclusions Baskar, S., & Suganthan, P. N. (2004). A novel concurrent particle swarm
optimization. In Congress on evolutionary computation, CEC2004 (Vol. 1, pp.
792–796).
In this paper, performance evaluation of various binary coded
Bingul, Z. (2004). A new PID tuning technique using differential evolution for
PSOs such as DBPSO, PBPSO, MBPSO, and MPSO in optimal design of unstable and integrating processes with time delay. In ICONIP, proceedings
multivariable PID and PI controller with and without decoupling of lecture notes in computer science (Vol. 3316, pp. 254–260).
the binary distillation column plant are performed. Multivariable Chang, W. D. (2007). A multi-crossover genetic approach to multivariable PID
controllers tuning. Expert Systems with Applications, 33, 620–626.
PID/PI controllers with and without employing decoupling are de- Chang, W. D. (2009). PID control for chaotic synchronization using particle swarm
signed considering minimization of ITAE and IAE. Simulations are car- optimization. Chaos, Solitons and Fractals, 39(2), 910–917.
ried out with identical stopping criteria in each case called maximum Chen, B. S., Cheng, Y. M., & Lee, C. H. (1995). A genetic approach to mixed H2/H1
optimal PID control. IEEE Control Systems, 15(5), 51–60.
number of fitness evaluations and results are compared with already Coelho, L. S., & Bernert, D. L. A. (2009). PID control design for chaotic
reported results. Simulation results can be summarized as follows: synchronization using tribes optimization approach. Chaos, Solitons and
Fractals, 42(1), 634–640.
Cohen, G. H., & Coon, G. A. (1953). Theoretical considerations of retarded control.
(i) Binary coded PBPSO has shown robustness in escaping from Transactions on ASME, 75, 827–834.
local optima as the bit flipping mechanism can help algo- Duan, H., Wang, D., & Yu, X. (2006). Novel approach to nonlinear PID parameter
rithm escape from local optima. optimization using ant colony optimization algorithm. Journal of Bionic
Engineering, 3(2), 73–78.
(ii) Significant control performance improvement is observed by Fukuda, T., & Shibata, T. (1992). Theory and application of neural networks for
employing steady state decoupling of the MIMO process. industrial control systems. IEEE Transactions on Industrial Electronics, 39(6),
(iii) PBPSO algorithm is identified as a robust method all in terms 472–489.
Gaing, Z. L. (2004). A particle swarm optimization approach for optimum design of
of its searching capability, convergence speed, consistency,
PID controller in AVR system. IEEE Transactions on Energy Conversion, 19(2),
and computational efficiency. 384–391.
(iv) All binary coded PSO’s are suitable for offline tuning. The Goshal, S. P. (2004). Optimization of PID gains by particle swarm optimization in
PBPSO being more consistent and fast algorithm is expected fuzzy based automatic generation control. Electric Power System Research, 72,
203–212.
to perform equally well in online tuning. Jan, R. M., Tseng, C. S., & Liu, R. J. (2008). Robust PID control design for permanent
(v) Both ITAE and IAE objective functions are suitable for the magnet synchronous motor: A genetic approach. Electric Power Systems
optimal design of PI or PID controller. Research, 78(7), 1161–1168.
Kennedy, J., & Eberhart, R. C. (1995). Particle swarm optimization. In Proceedings of
(vi) Better performance of PBPSO over DBPSO, MBPSO, MPSO, RGA the IEEE international conference neural networks (Vol. 4, pp. 1942–1948).
with SBX cross over, BLT, and CMAES algorithms is observed. Kennedy, J., & Eberhart, R. C. (1997). Discrete binary version of particle swarm
(vii) Further research should be focused on comparative perfor- optimization algorithm. In IEEE proceedings on computational cybernetics and
simulation (pp. 4104–4108).
mance analysis of various EAs over larger parameter spaces. Kim, T. H., Maruta, I., & Sugie, M. (2008). Robust PID controller tuning based on the
constrained particle swarm optimization. Automatica, 44(4), 1104–1110.
Lee, C. C. (1990). Fuzzy logic in control systems: Fuzzy logic controller – Part I and II.
IEEE Transactions on Systems, Man and Cybernetics, 20(2), 404–435.
Acknowledgements Marsh, P. (1998). Turn on tune in-Where can the PID controllers go next. New
Electron, 31(4), 31–32.
Menhas, M. I., Wang, L., Fei, M., & Ma, C. X. (2011). Coordinated controller tuning of
This work is supported by the Projects of Shanghai Science and boiler turbine unit with new binary particle swarm optimization algorithm.
Technology Community (10ZR1411800, 08160512100 and International Journal of Automation and Computing, 8(2), 185–192.
10JC1405000), National Natural Science Foundation of China Mukherjee, V., & Goshal, S. P. (2007). particle swarm optimized fuzzy PID controller
for AVR system. Electric Power System Research, 77(12), 1689–1698.
(Grant Nos. 60834002 and 61074032), Research Fund for Doctoral Shen Qi, J. J. (2004). Modified Particle Swarm optimization algorithm for variable
Program of Higher Education of China (20103108120008), Mecha- selection in MLR and PLS modeling. QSAR studies of antagonism of angoitensin
tronics Engineering Innovation Group Project from Shanghai Edu- II antagonists. European Journal of Pharmaceutical Sciences, 22, 145–150.
Shi, Y., & Eberhart, R. C. (1998). A modified particle swarm optimizer. In Proceedings
cation Commission and Graduate Innovation Fund of Shanghai of IEEE international conference on evolutionary computation, Anchorage, AK (pp.
University (SHUCX102218). 69–73).
Van den Bergh, F., & Engelbrecht, A. P. (2004). A Cooperative approach to particle
swarm optimization. IEEE Transactions on Evolutionary Computation, 8(3),
References 225–239.
Wang, L., Wang, X. T., Fu, J. Q., & Zhen, L. L. (2008). A novel probability binary
Åström, K. J., & Hägglund, T. (2001). The future of PID control. Control Engineering particle swarm optimization algorithm and its application. Journal of Software,
Practice, 9(11), 1163–1175. 9(3), 28–35.
 M.I. Menhas et al. / Expert Systems with Applications 39 (2012) 4390–4401 4401

Wang, J. S., Zhang, Y., & Wang, W. (2006). Optimal design of PI/PD controller for Zhan, Z.-H., Zhang, J., Li, Y., & Chung, H. S. (2009). Adaptive particle swarm
non-minimum phase system. Transactions of the Institute of Measurement and optimization. IEEE Transactions on System, Man and Cybernetics, 39(6),
Control, 28(1), 27–35. 1362–1381.
Willjuice, I. M., & Baskar, S. (2009). Evolutionary algorithm based design of Zhao, S. Z., & Suganthan, P. N. (2009). Diversity enhanced particle swarm optimizer
multivariable PID controller. Expert Systems with Applications, 36, 9159–9167. for global optimization of multimodal problems. IEEE Congress on Evolutionary
Willjuice, I. M., & Baskar, S. (2010). Covariance matrix adaptation evolution strategy Computation, 590–597.
based design of centralized PID controller. Expert Systems with Applications, 37, Ziegler, J. G., & Nichols, N. B. (1942). Optimum settings for automatic controllers.
5775–5781. Transactions on ASME, 64(8), 759–768.
Wood, R. K., & Berry, M. W. (1973). Terminal composition control of a binary Zuo, W. (1995). Multivariable adaptive control for a space station using genetic
distillation column. Chemical Engineering Science, 28(9), 1707–1717. algorithm. IEEE Proceedings Control Theory and Applications, 142(2), 81–87.
Zhang, J., Zhuang, J., Du, H., & Wang, S. (2009). Self-organizing genetic algorithm
based tuning of PID controllers. Information Sciences, 179(7), 1007–1018.