You are on page 1of 14

Measurement 46 (2013) 16161629

Contents lists available at SciVerse ScienceDirect

Measurement
journal homepage: www.elsevier.com/locate/measurement

Fuzzy model predictive control of non-linear processes using


convolution models and foraging algorithms
Ankush Chakrabarty a,, Suvadeep Banerjee b, Sayan Maity c, Amitava Chatterjee c
a
Purdue University, Electrical and Computer Engineering Department, IN, USA
b
Georgia Insitute of Technology, Electrical and Computer Engineering Department, GA, USA
c
Jadavpur University, Electrical Engineering Department, West Bengal, India

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

Article history: In this paper, a fuzzy model predictive control (FMPC) approach is introduced to design a
Received 27 March 2012 control system for nonlinear processes. The proposed control strategy has been success-
Received in revised form 20 November 2012 fully employed for representative, benchmark chemical processes. Each nonlinear process
Accepted 23 November 2012
system is described by fuzzy convolution models, which comprise a number of quasi-linear
Available online 13 December 2012
fuzzy implications (FIs). Each FI is employed to describe a fuzzy-set based relation between
control input and model output. A quadratic optimization problem is then formulated,
Keywords:
which minimizes the difference between the model predictions and the desired trajectory
Fuzzy convolution models
Fuzzy model predictive control
over a predened predictive horizon and the requirement of control energy over a shorter
Bacterial Foraging algorithms control horizon. The present work proposes to solve this optimization problem by employ-
Direct adaptive control ing a contemporary population-based evolutionary optimization strategy, called the Bacte-
rial Foraging Optimization (BFO) algorithm. The solution of this optimization problem is
utilized to determine optimal controller parameters. The utility of the proposed controller
is demonstrated by applying it to two non-linear chemical processes, where this controller
could achieve better performances than those achieved by similar competing controller,
under various operating conditions and design considerations. Further comparisons
between various stochastic optimization algorithms have been reported and the efcacy
of the proposed approach over similar optimization based algorithms has been concluded
employing suitable performance indices.
2012 Elsevier Ltd. All rights reserved.

1. Introduction over a nite time-horizon to determine the optimal inputs


required to minimize such errors. A major reason for the
Model predictive control (MPC) is a family of advanced success of MPC techniques in numerous industrial applica-
discrete-time control algorithms, which have steadily tions is due to its ability to handle constraints in a system-
gained popularity in previous years due to its simplicity atic way during the design and implementation of the
of implementation and its high levels of accuracy reported controller [2,4,5]. This makes model predictive control
in the technical literature [28]. The philosophy of the MPC optimal for applications in multi-variable systems, non-
scheme is based on the on-line use of an explicit process minimum phase and dead-time systems, as well as situa-
model, also known as a design model, to predict the behav- tions involving plant-model mismatch in order to perform
ior of an actual physical system. This predicted trajectory is in a robust and efcient manner [69,26,39].
compared to the actual measured data at each time sam- The performance of a model-based predictive control-
ple, and thereby an optimization procedure is formulated ler, also known as a receding horizon controller, is highly
sensitive to the accuracy of the design model, making the
Corresponding author. design of the controller a non-trivial engineering problem.
E-mail address: chakraa@purdue.edu (A. Chakrabarty). Furthermore, the goals of the control system must be

0263-2241/$ - see front matter 2012 Elsevier Ltd. All rights reserved.
http://dx.doi.org/10.1016/j.measurement.2012.11.046
A. Chakrabarty et al. / Measurement 46 (2013) 16161629 1617

mathematically reduced to a single all-encompassing However the chief drawback of this proposed methodology
objective function to avoid lengthy multi-objective optimi- was that it required frequent activation of update laws for
zation procedures and use the prevalent active-set, interior tuning the controller parameters. Another approach was
point or sequential quadratic programming approaches proposed in [10] for designing a fuzzy model-based state
[28]. Most traditional MPC algorithms essentially assume space feedback controller, which treated their obtained
that the plant can be described by a linear model, which TS model as a conglomeration of conventional piecewise
is a deviation from real life situations where most indus- linear models. However, this resulted in the deterioration
trial plants exhibit inherently non-linear dynamics. Hence, of the closed-loop dynamic performance of the system
in many practical situations, this renders linear MPC and loss of the soft transitions in varying operating regions,
strategies inefcient unless the nonlinear model can be lin- which is usually considered the inherent characteristic or
earized about an equilibrium state without adversely sac- strength of any TS model.
ricing modeling accuracy due to approximation. In order In the present paper, we propose a new methodology
to facilitate the utilization of the MPC philosophy with re- for fuzzy model-based predictive control which has been
quired degree of satisfaction in those real world situations, successfully employed to control two benchmark process
many researchers have later focused their attention on control systems, very popularly considered in the chemical
developing a new family of MPC algorithms, called non- engineering industry. These chemical processes are de-
linear MPC (NMPC) methodologies [3]. scribed using fuzzy convolution models, and are consid-
Fuzzy control has the inherent advantage of combining ered to accurately describe the plant dynamics as
heuristic logic with analytical functions, thereby seam- reported in [1]. In this work, these fuzzy convolution mod-
lessly integrating qualitative knowledge with highly non- els are utilized to predict the temporal variations in the
linear systems and providing superior control performance output obtainable from these design models, and these
on many occasions [1317,4147]. Over the years, fuzzy outputs are utilized to perform a comparison with a refer-
control has evolved as an extremely popular and viable ence trajectory and a pre-dened set-point. This error,
control alternative for the purpose of modeling and control along with the change in control energy is then formulated
in a variety of applications, ranging from robotics and in the form of a quadratic cost function, which is mini-
mechanical systems, to electrical drives, in process control mized to obtain some optimal weighting matrices. These
of highly non-linear chemical processes and in other elds matrices comprise the major controller parameters re-
of engineering [3034]. Similarly stochastic optimization quired to design the FMPC. The present work proposes to
techniques, specially biologically-inspired optimization solve this minimization procedure utilizing a global, sto-
algorithms in particular, have also been employed success- chastic optimization strategy and it has successfully ap-
fully to solve the adaptive control problems and other re- plied one such modern technique developed in the last
lated problems such as robotic navigation problems, decade simulating the behavior of E. coli in the human
communication resource allocation problems, power intestinal tract, known as Bacterial Foraging Optimization
systems control, power electronics and drives related (BFO) algorithm. This algorithm, proposed in [20], and used
problems [3538], etc. In the last decade, such biologi- successfully in a wide range of engineering applications
cally-inspired optimization algorithms and swarm intelli- [21,29] mimics the foraging behavior of a living organism
gence based algorithms have been successfully employed (represented by a vector of decision variables in the
in conjunction with several direct and indirect adaptive optimization algorithm) in a nutrient concentration (repre-
fuzzy control, model predictive control, fuzzy sliding mode sented by the objective function surface in the optimiza-
control, fuzzy robust control [18,19,29,33,40]. In this tion algorithm), where it tries to maximize nutrient
paper, we explore a specic scheme of non-linear MPCs, intake in minimum time with the minimum expenditure
known as fuzzy MPCs (FMPCs) which combine the heuris- of energy. These computed optimal weighting matrices
tic strategy of fuzzy control and the accuracy of model- are then used to compute the actual output trajectory of
based predictive control and demonstrate how such fuzzy the plant under the inuence of the optimally-weighted
controllers can be effectively designed using a contempo- FMPC, and its performance is compared with the work pre-
rary bio-inspired stochastic optimization algorithm called sented in [1], as well as identical FMPCs designed with
Bacterial Foraging Optimization (BFO) algorithm. other contemporary evolutionary optimization algorithms
It is well known that a non-linear system can be divided to demonstrate the superiority of the proposed approach
into a number of linear or nearly linear subsystems. It is with the help of performance indices such as rise time, set-
possible to develop quasi-linear empirical models utilizing tling time, peak overshoot and other error metrics.
the concept of fuzzy logic for each such individual subsys- The rest of the paper is organized as follows. Section 2
tem [12]. Each model, hence developed, can be called a discusses an overview of the fuzzy convolution model
rule-based fuzzy implication (FI). The overall process and fuzzy model predictive control paradigm. Section 3
behavior can then be characterized by a fuzzy weighted discusses the Bacterial Foraging Optimization (BFO)
sum of the outputs from all quasi-linear FIs. This method- algorithm in detail and how it is implemented here in
ology facilitates the development of a non-linear model connection with the optimal design of the FMPC. Section
that is essentially a collection of a number of quasi-linear 4 presents the performance evaluation, when the control-
models regulated by fuzzy logic. It also provides an oppor- ler is applied to two process control systems described
tunity to simplify the design of non-linear model predic- by their respective fuzzy convolution models. Finally, the
tive controllers. In [11] an NMPC algorithm was conclusions inferred from the results are presented in
developed using Takagi Sugeno (TS) type fuzzy models. Section 5.
1618 A. Chakrabarty et al. / Measurement 46 (2013) 16161629

2. Fuzzy model predictive control employing fuzzy be minimized within a proper number of predictive hori-
convolution models zon steps (Ny) through the selection of optimal control ac-
tions for the future control horizon (Nu) steps. Hence this
In chemical and petrochemical plants, the continuous optimization problem can be congured as [28],
and batch processes involved are inherently very non-lin-
min Jn; 4
ear. In [12], the authors proposed a novel fuzzy logic-based Dun;Dun1;;DunNu 
modeling methodology for handling such highly non-linear
where J[n] is the objective function, dened as,
systems. In this approach, they de-construct the original
non-linear system into a number of linear or quasi-linear Ny
X X
Nu
subsystems, that is, a single-input single-output (SISO) Jn li ky^n i  yr n ik2 mi kDun ik2 ; 5
highly non-linear system S was decomposed into p linear i1 i1

or quasi-linear subsystems. An integral part of this design where li and mi are weighting factors for the prediction er-
algorithm is the formulation of rule-based fuzzy-implica- ror and control energy respectively. Ny and Nu are the pre-
tions (FIs) for each subsystem. An FI is a rule based on the dictive and control horizons of the model-predictive
classical IF-THEN structure of fuzzy logic approach, consist- ^n i and yr[n + i] are the predicted and refer-
controller, y
ing of symbolic antecedents as the premise (IF. . .) and a ence controller outputs respectively, and Du[n + i] is the
linear algebraic expression as the consequence (THEN. . .). future change in control action which is to be derived using
Each FI is generated based on a system response to an im- the minimization procedure discussed above. The control
pulse signal [23,24]. Hence, analogous to signal theory, such policy is so developed that for each jth fuzzy convolution
fuzzy models are called fuzzy convolution sub-models subsystem there is a local FMPC controller that develops
where each rule relating the system output y with control the ith future step control strategy as Duj[n + 1]. Then for
input u is given as, p such local controllers, all their local control policies are
Ri : IF mixed in a fuzzy manner to generate the overall control
policy,
yn is Ai0 ; yn  1 is Ai1 ;    ; yn  m 1 is Aim1 ; Pp j j
AND j1 w Du n i
Dun i Pp j
: 6
un is Bi0 ; un  1 is Bi1 ;    ; un  l 1 is Bil1 j1 w

THEN It is quite common to choose the same weight wj for


X
T both the jth local controller and the corresponding jth
i
yi n 1 yn hj Dun 1  j; sub-convolution model. A substitution of (2) and (6) into
j1 (5) yields a computation heavy objective function which
1 can be approximated, using Cauchys inequality, as shown
in [1],
where y[n  j] is the measured output y at j instances be-
fore the present sampling instant, u[n  j] is the control in- X
p
min eJn min xj 2eJ j n;
put at instant n  j; Aij the fuzzy set for the output y[n  j] Dun;Dun1;;DunNu  Dun;Dun1;;DunN u 
j1
in the ith FI, Bij the fuzzy set for the control action u[n  j]
i
in the ith FI, hj is the impulse response coefcient in the ith 7
FI, T is the model horizon for the impulse-response model where,
and Du[n] the difference between consecutive control in-
Ny
X X
Nu
puts, dened as Du[n] = u[n]  u[n  1].
eJ j li ky^j n i  yr n ik2 mi kDuj n ik2 ; 8
If p such FIs constitute a complete fuzzy convolution
i1 i1
model for the system S, then the system output y[n + 1]
is computed as, and
Pp j j
i1 w y n 1 wj
yn 1 Pp j
: 2 xj Pp : 9
i1 w
j
j1 w

Here, wj represents the composite membership value for However, even such a relatively simpler form of the
the jth FI and it can be evaluated based on the activation objective function is quite difcult to solve for suitable
strength of all fuzzy sets in the antecedent parts of rule determination of controller gains using classical optimiza-
Ri, that is, tion strategies. Hence, in our opinion, this problem
formulation presents a suitable opportunity for utilization
wj ^i Aji ^k Bjk : 3
of a stochastic, derivative-free optimization algorithm to
Here, ^ denotes the AND operation on fuzzy sets. In this determine the relevant controller parameters. As men-
paper, we use this fuzzy convolution model as the design tioned before, this work proposes the use of the Bacterial
model for implementation of our optimally-weighted fuz- Foraging Optimization (BFO) algorithm, a relatively recent
zy model predictive control (FMPC) strategy. The develop- population-based global optimization algorithm, to solve
ment of such an FMPC will attempt to optimize an this controller design problem.
objective function so that the predictive error between The predictive output of the jth FI can be derived using
the process output and a given reference trajectory can the structure of the fuzzy convolution model presented in
A. Chakrabarty et al. / Measurement 46 (2013) 16161629 1619

(1). Expressed in a matrix form, the predictive output can Similarly, large values of control energy must be penalized,
be written as, resulting in the second weighting matrix, of the form,
2 3
Y j n Aj DU n Yn P j n; 10 mj1 0  0
6 7
where, j
6
6 0 mj2    0 7
7
W2 6 .. .. 7: 23
2 3 6 7
aj1 0 0  0 4 .   . 5
6 7 0 0  mjNu
6
6 aj2 aj1 0  0 7
7
6 7
j
A 6 aj3 aj2 aj1  0 7; 11 The overall objective function to be optimized is then
6 7
6
6 .. .. .. .. .. 7
7 derived by substituting the individual objectives ~J j n in
4 . . . . . 5 (21) to Eq. (7) as the control aim is to minimize deviation
ajNy ajNy 1 ajNy 2    ajNy Nu 1 for all p subsystems. Algorithm 1 describes the implemen-
tation of the proposed afore-mentioned optimal-feedback
X
i fuzzy model predictive controller (FMPC). We call this
j
aj1 hk ; 12 strategy optimal-feedback because the optimization proce-
k1 dure derives optimal values of the feedback gain matrix K.
Hence, we can reformulate the objective function to in-
Yn yn yn yn    yn ; 13 clude the feedback matrix as shown,
h i>
P j n Pj1 n Pj2 n    PjNy n ; 14 X
p
mineJn min xj 2eJ j n;
K K
j1
X
i X
T    
j eJ j n Y j n  Y r n > W j Y j n  Y r n
Pji n hi Dun l  k; 15 1
k1 lk1 K j Z j n> W j2 K j Z j n: 24
 >
Y j n y
^j n 1 y
^j n 2    y
^j n Ny  : 16
Algorithm 1. Computation of the cost function in pseudo-
Hence, the control law for the localized fuzzy model pre-
code format.
dictive controller, corresponding to the jth FI, can be ex-
pressed in the form of a local linear state-feedback
controller employing the control law, Require: Nmh (model horizon),
DU j n K j Z j n; 17 Require: Nu (control horizon),
where,
  Require: Ny (output predictive horizon),
U j Duj n 1 Duj n 2    Duj n Nu ; ; 18
Require: p (number of subsystems),
Z j n Yn  Y r n P j n; 19
Require: DT (sampling time in accordance with the
 > open-loop settling time restriction),
Y r n y ^r n 2    y
^r n 1 y ^r n Ny  ; 20
Require: Tref (complete simulation time),
Kj is the feedback gain matrix for the jth subsystem. In our
proposed control strategy, the optimal controller parame- Require: W1 and W2,
ters are determined using BFO algorithm. The objective
function to be solved for determining this solution, refor- Require: Impulse response matrix h.
mulated in (8), can be expressed in a standard quadratic Calculate the feedback gain matrix K from the
form as, healthiest controller candidate from the previous
   
~J j n Y j n  Y r n > W j Y j n  Y r n iteration of the BFO algorithm.
1
Initialize all the counter values to zero.
DU n> W j2 DU n: 21 Calculate model output response values yr for
Here, W j1
is the vector of weighting factors penalizing large entire time of simulation.
deviations from the reference trajectory Yr[n], and has the for n = 1 to Tref do
form, Compute P using (14) and (15).
2 3 Compute Y[n], Yr[n], Zj[n] for the jth subsystem
lj1 0  0
as in (13), (19) and (20).
6 7
j
6 0
6 lj2    0 7
7 Compute the jth FIs control law U j n using
W1 6 . .. 7: 22
6 . 7 (17).
4 .   . 5
Compute Aj as in (11) and (12).
0 0  ljNy
(continued on next page)
1620 A. Chakrabarty et al. / Measurement 46 (2013) 16161629

Compute the predictive horizon values the chemotactic step, k for the reproduction step, and l for
j the elimination-dispersal event. Also let p be the dimension
Y Aj DU j n Yn P j n.
n
of the search space, S the total number of bacteria in the
Calculate unfuzzied objective function using
population, Nc the number of chemotactic steps, and Ns
Eq. (21).
the swimming length of each bacterium. Nre is dened as
Using fuzzy operations, calculate fuzzy ring
the number of reproduction steps, Ned the number of elim-
rules x = [x1,    ,xp]>.
ination-dispersal events, and Ped the Elimination-dispersal
Calculate fuzzied overall objective function
P probability per cycle. C[i] is the size of the step taken in
Jn pi1 x2i J i n. the random direction specied by the tumble.
Calculate overall control action U+[n] as shown The BFO algorithm is given in algorithm form in [20,21],
P
in (18), where, Dun pk1 xk DU n. whose main modules are described as follows.
if n < Nmh then
Calculate output response by fuzzy 3.1.1. Chemotaxis
convolution model of the system enforcing Chemotaxis [22] is the tendency of each bacterium to
causality move in various directions in search for nutrients. Alter-
else nate rotations of the bacterial agella in clockwise and
Calculate output response by fuzzy counter-clockwise direction help the bacterium in two
convolution model of the system up to model kinds of actions: swimming and tumbling. A tumble is rep-
horizon resented by a unit length random direction of movement,
end if which is kept xed throughout the optimization: denoted
end for as /[j]. Thus, each subsequent position reached by the bac-
Plot control performances. terium after chemotaxis is given as,
hi j 1; k;  hi j; k;  Ci/j: 25

3. Bacterial Foraging Optimization (BFO) for 3.1.2. Swarming


determining control gains If a bacterium stumbles onto a nutrient rich environ-
ment, it can signal nearby bacteria to swarm around
3.1. The BFO algorithm that nutrient-rich space. If however, the bacteria come
too close, repellant signals can be exuded. Thus, these
Foraging strategies are methods for locating, handling phenomena are modeled in the swarming step. The com-
and ingestion of nutrients [20,21]. In accordance with bined cell-to-cell attraction and repulsion can be ex-
the axiomatic survival of the ttest, organisms are nat- pressed as,
urally selected to survive if they have the most efcient
X
S
foraging strategies. Foragers aim to maximize the energy J cc h; Pj; k;  J cc h; hi j; k; 
intake per unit time spent foraging in lieu of all possible i1
2 !
environmental constraints. It was in this respect that p 
X 2
Passino [20] mathematically modeled the behavior of XS 6 watt hm  him
6
the micro-organism Escherichia coli and formulated this 6datt exp m1
4
strategy as an optimization tool. The entire foraging i1

strategy can be classied into four sub-groups: (a) che- !


motaxis, (b) swarming, (c) reproduction and (d) elimina- Xp  2 3
i
tion-dispersal. wrep hm  hm 7
7
The BFO algorithm mimics these four steps in order to hrep exp m1 7: 26
5
acquire a global optimum (if any). Each virtual bacterium
is a trial solution on the nutrient (functional) surface.
Here the objective is to determine the minimum of the In the above equation, watt, hrep, wrep and datt are appropri-
cost function J(h) where h denotes the coordinates of a ate coefcients that have to be chosen according to the
bacterium in a p-dimensional search space, i.e. h 2 Rp , problem under consideration.
where no analytical description of the gradient rJ(h) is
available. In the context of a real-world optimization 3.1.3. Reproduction
function, p denotes the number of variables for optimiza- In this step, among the entire population of S bacteria, Sr
tion. Positive, zero and negative values of J(h) represent bacteria die due to poor health and the Sr healthiest bacte-
inhibitory, neutral and supportive environments respec- ria enjoy reproductive success and each of them splits up
tively. Locomotion in bacteria takes place in two modes: into two co-located bacteria, so that the population re-
by tumbling (longer distances covered) and running mains constant. The health of a bacterium is determined
(smaller steps). by its tness value or the objective function, that is, for a
A chemotactic step is dened as a tumble followed by a minimization problem, the lower the tness value, the bet-
tumble or a tumble followed by a run. Let j be the index for ter the health of the bacterium.
A. Chakrabarty et al. / Measurement 46 (2013) 16161629 1621

Fig. 1. Application of BFO in optimal selection of the FMPC parameters.

3.1.4. Elimination and dispersal Algorithm 2. The Bacterial Foraging Optimization (BFO)
This step usually occurs after the reproduction step. algorithm in fuzzy model predictive control design
The bacteria in this step are selected from a uniform dis-
tribution and subsequently destroyed if they are too far
from any local minima. New bacteria replace them in ran- Require: Choose population size S, i.e. the total
dom new locations simulating real-world conditions such number of controller parameters required for each
as wind dispersal, thereby leveraging their chances of dis- candidate controller.
covering a wayward optimal point previously missed by
Require: Initialize the parameters, C[i],i = 1, 2, . . . , S.
the rest of the population. The total population S how-
ever, is always maintained constant to prevent overuse Also initialize all the counter values to zero.
of computational effort and alleviate dwindling bacterial Require: Select dimension of search space p = Ncon.
populations. The control scheme proposed in this work
loop
uses only the chemotactic movement of the BFO algo-
for = 1 to Ned do
rithm since online controller candidate selection methods
for k = 1 to Nre do
require swift responses and limiting the BFO algorithm to
for j = 1 to Nc do
only the chemotactic step improves the response time of
for i = 1 to NS do
the system. Fig. 1 depicts the block diagram of the
Compute J[i,j,k,] using the procedure
DAFMPC.
detailed in Algorithm 1.
Replace Jlast = J[i,j,k,].
Generate a random vector Di 2 Rp
3.2. Application of BFO to direct adaptive fuzzy model
(Tumbling).
predictive control (DAFMPC)
Tumble to the new position determined
by the vector D[i] in the following manner,
For each of the p subsystems, the feedback gain matrix
is K j 2 RNu Ny . Thus, for each sub-system there are N/ num- hi j; k 1;  hi j; k;  Ci kDDi
ik.

ber of controller parameters where N/ = Nu  Ny. Thus, for Compute the objective function at the
p number of subsystems there are a total of Ncon number new location, J[i,j + 1,k,].
of controller parameters for the control of all the quasi- Set m = 0.
linear subsystems describing the performance of the highly while m < NS do
non-linear system S, where Ncon = pN/. Update counter: m = m + 1.
In the application of the BFO, each bacterium repre- if J[i,j + 1,k,] < Jlast then
sents a controller candidate h 2 RNcon and the total bac- J[i,j + 1,k,] = Jlast.
terial population is held xed at S. After each foraging Tumble to the new position
iteration, the 2S healthiest bacterial population are selected determined by the vector D[i] in the following
and the feedback gain matrix Kj, where j = 1,2,    ,p is manner, hi j 1; k;  hi j 1; k;  Ci kDDi
ik.
reformulated from the nal population set. Using this up- else
dated feedback gain matrix, we calculate the objective m = Ns.
function for the next iteration of the BFO algorithm as re- end if
ported in Algorithm 1. Algorithm 2 demonstrates this end while
application of BFO to nd the optimum controller end for
parameters.
(continued on next page)
1622 A. Chakrabarty et al. / Measurement 46 (2013) 16161629

for i = 1 to S do such that TDt P the open-loop settling time, i.e. the time
PNc 1 taken for open loop response to reach 99% of its steady-
Compute J ihealth r1 Ji; r; k; l.
state value. All simulations were performed on MATLAB
end for
R2012a. Parity amongst the different optimization algo-
Sort bacteria in order of cost values of
rithms used were maintained as far as possible in order
health Jhealth.
to compare results fairly, by equalizing population sizes,
Destroy Sr bacteria with the highest values
generation numbers and percentages of population repro-
of health Jhealth (i.e. least healthy bacteria).
duced as well as step sizes. In particular, for the Bacterial
Split each of the Sr bacteria with the lowest
Foraging Optimization, Table 1 shows the parameter val-
values of health Jhealth into two and each such pair
ues used during simulation.
resides in the same original location of the parent.
end for
for i = 1 to S do 4.1. Case study I
Eliminate and disperse each bacterium with
probability ped, keeping total population of bacteria In this system, we have considered the following fuzzy
constant. convolution model, for two FIs (R1 and R2 ),s
end for
R1 : IF yn is A1 THEN y1 n 1
end for
end for X
70
1
yn hi un 1  i; 27
end loop (termination criteria is reached) i1
Select healthiest bacteria, i.e. controller candidate
and formulate optimal state feedback matrix K for
R2 : IF yn is A2 THEN y2 n 1
output response simulation.
X
70
2
yn hi un 1  i; 28
i1

where the fuzzy sets A1 and A2 are presented in Fig. 2.


4. Performance evaluation Considering U = 2, V = 3 and T = 70, the impulse re-
sponse coefcients of the system have been shown in
In this section, the proposed optimal fuzzy model pre- Fig. 1. The sampling time of the controller is taken to be
dictive controller (FMPC) has been applied to two bench- 10 s.
mark, non-linear chemical systems to study the response The weighting factors we use for each convolution mod-
of the designed system for a regulatory problem. The rst el to be employed in Eq. (24) are as follows,
system is modeled using two FIs and the second system 2 3
20 0 0
is described by three FIs and has its input and output con- 6 7
taining random noise for robustness analysis of the pro- W j1 4 0 20 0 5; 29
posed method [1]. 0 0 10
The proposed controller is applied for closed-loop adap-

tive control, maintaining some constraints in order to de- 25 0
W j2 : 30
crease computational complexity and computation time. 0 25
Like all MPCs, the parameters of the FMPC to be tuned
manually are the Model Horizon (T), Control Horizon (U), We consider  the reference model to be discrete samples
t

and Predictive Horizon (V), and U and V are chosen suf- of yref t ysp 1  es , where the time constant s is se-
ciently small to avoid large computational delays. An in- lected as 0.5 and 2.5 min, and the set point is xed at
crease in V results in an increased accuracy of the control ysp = 6.0 for the entire simulation. Fig. 3 shows the closed
and increase in U results in enhanced robustness of the loop controlled response of the process output with our
controller. Hence, the choice of U and V should be made designed BFO-FMPC controller in comparison to similar
as a trade-off between system stability and computational
effort. Furthermore, the selection of sampling time Dt is
Fuzzy Sets for Case 1
1.5
1
A
Table 1
A2
Parameter set used in Bacterial Foraging Optimization algorithm for FMPC
design. The code for the implementation of BFO is freely available at [27]. 1
(y)

Parameter Variable name Value


Population size S 20
0.5
Number of reproductive steps Nre 10
Number of chemotactic steps Nc 4
Number of elimination-dispersal steps Ned 2
Probablity of elimination-dispersal 0
0 2 4 6 8 10 12 14 16 18
in total bacteria population ped 0.25
Fraction of population reduced Sr 0.5
y (t)
Chemotactic step size  0.1
Fig. 2. Fuzzy sets chosen for the system model in case study I.
A. Chakrabarty et al. / Measurement 46 (2013) 16161629 1623

0.2 System 1 Impulse Response of the Convolution Model 1

0.15

h1 (t)
0.1

0.05

0
0 5 10 15 20 25 30 35 40 45 50 55 60 65 70
Sample (N)

2 System 1 Impulse Response of the Convolution Model 2

1.5
h (t)

1
1

0.5

0
0 5 10 15 20 25 30 35 40 45 50 55 60 65 70
Sample (N)

Fig. 3. Impulse responses of the fuzzy convolution models used to model system I.

FMPC controllers designed employing two other popular


competing evolutionary algorithm techniques, called ge-
netic algorithm (GA) and particle swarm optimization Comparison of Performances in System I with =0.5 min
(PSO). These two controllers are named as GA-FMPC and (a) 8
PSO-FMPC respectively. It is expected that the performance
7
will become sluggish as the time constant s increases,
which is clearly visible in the performances achieved in 6
Fig. 4. An interesting observation in Fig. 4 is the relatively 5
large overshoot and sustained oscillations produced by
out

4
y

the PSO-FMPC. One of the marked possibilities of such


behavior can be attributed to the fact that the PSO algo- 3 Reference
BFOFMPC
rithm may not be performing that well with a relatively GAFMPC
2
small swarm size. Since our BFO-FMPC was able to achieve PSOFMPC

much better performances than PSO-FMPC using a small 1


population size, then it can be concluded that the proposed
0
BFO-based controller is not only accurate, but also requires 0 0.5 1 1.5 2 2.5 3
reduced memory consumption and lesser computational Time (t) [mins]
burden, and hence evolves as a clear winner for integration
Performance Comparison in System I with =2.5 min
into low-capacity devices. At rst glance, there is no distin- (b) 7
guishing characteristic in the responses of the GA- and
6
BFO-FMPC, but a closer inspection of the performance met-
ric indicates that the GA-FMPC produces a larger SSE in
5
comparison to our proposed approach. Hence, for a regula-
tory control problem, the BFO-FMPC could demonstrate a 4
greater effectiveness than its competitors.
out
y

The performance indices considered in this study are 3


chosen carefully due to the lack of quantitative perfor-
2
mance indices in [1], which we compare our proposed Reference
BFOFMPC
techniques against. By eye inspection, the rise time, 98%
1 PSOFMPC
settling time and peak overshoot percentages are reported GAFMPC

for our proposed scheme, its competitor schemes, as well 0


0 0.5 1 1.5 2 2.5 3
as the scheme proposed in [1]. In addition, we calculate
Time (t) [mins]
the sum of the squared-errors (SSEs) for the three stochas-
tic-optimized feedback controllers, and compare them in Fig. 4. Output response for case study I with different choices of time
Table 2. The SSE is dened by, constant s.
1624 A. Chakrabarty et al. / Measurement 46 (2013) 16161629

Table 2
Quantitative performance comparison among competing controllers for case study I with two different choices of time constant (s). The bold values are
indicative of the controller which outperforms the other controllers for each category. It is quite easily discernible from these bold values that the proposed
FMPC-BFO approach is the best candidate for this controller design.

Variation of s Performance FMPC with iterative design Proposed FMPC with Proposed FMPC with GA- Proposed FMPC with
for system I indices based controller [1] PSO-based approach based approach BFO-based approach
s = 2.5 min Rise time (s) 235 17 19 18
Peak 20 6 0 0
overshoot
(%)
Settling time 480 135 129 129
(s)
SSE (e) N/A 2528.8 2498.2 2488.7
s = 0.5 min Rise time (s) 60 6 10 7
Peak 75 36 8 3
overshoot
(%)
Settling time 230 140 115 92
(s)
SSE (e) N/A 2550.5 2567.9 2520.6

Performance comparison Performance comparison Performance comparison


when W=diag(25) when W=diag(10) when W=diag(8)
8 8 8
Reference Reference Reference
BFOFMPC BFOFMPC BFOFMPC
7 PSOFMPC 7 PSOFMPC 7 PSOFMPC
GAFMPC GAFMPC GAFMPC

6 6 6

5 5 5
out

out
out

4 4 4
y

y
y

3 3 3

2 2 2

1 1 1

0 0 0
0 1 2 3 0 1 2 3 0 1 2 3
Time (t) [mins] Time (t) [mins] Time (t) [mins]

Fig. 5. Output response comparison with variation in W j2 . (a) W = diag(25), (b) W = diag(10), (c) W = diag(8).

X
T advantage to the controller design problem. As known
e kyi  yref ik2 : 31 from the design of MPC controllers [28], there is no empir-
i0
ical or analytical relation to obtain the weighting matrices
Here, i denotes the time index and y is the output of the in the cost function described in (21), as a heuristic ap-
controller. yref is the reference signal, from which the devi- proach may be used effectively for the design of the
ation of y should be minimized, and hence this variation is weighting matrices without expecting a great degradation
computed as an error metric. in performance.
In order to verify the sensitivity of the controller to the An odd behavior is noticed in the GA-FMPC with a de-
weighting matrix W j2 , the elements of this diagonal matrix crease in W j2 . For the rst case, where W j2 is high, the con-
are varied from 8 to 25 and three such elements are chosen troller produces a satisfactory output, and is comparable to
to create different W j2 matrices. The performance of our the BFO-FMPC. However, as the relative importance given
proposed FMPC is also compared with the FMPC developed to the control input energy lowers, the GA-FMPC shows
in [1] where the controller design was accomplished using an oscillatory behavior in Fig. 5b and gradually diverges
a two-layer iterative design process, using the basic con- from the reference in Fig. 5c.
cept of decompositioncoordination in a large-scale sys- Thus, the GAFMPC shows adverse effects due to the
tem theory [25]. It is quite apparent from the subplots of lowering of the W-matrix values, and hence it can be
Fig. 5 that our proposed BFO-FMPC (blue line) controller deemed inferior to the BFOFMPC in terms of stability and
is relatively insensitive to the W j2 matrix. This offers a large robustness of parameter settings. The PSO-FMPC performs
A. Chakrabarty et al. / Measurement 46 (2013) 16161629 1625

Table 3
Quantitative performance comparison among competing controllers for case study I when the weighting factor is changed.

Variation of W Performance FMPC with iterative design Proposed FMPC with Proposed FMPC with Proposed FMPC with
matrix for system I indices based controller [1] PSO-based approach GA-based approach BFO-based approach

W j2 diagf25; 25g Rise time (s) 66 17 19 17


Peak 80 7 0 0
overshoot
(%)
Settling time 238 443 110 110
(s)
SSE (e) N/A 2528.8 2498.2 2488.7

W j2 diagf10; 10g Rise time (s) 60 22 20 19


Peak 45 0 0 0
overshoot
(%)
Settling time 192 160 150 142
(s)
SSE (e) N/A 2519.4 2501.5 2490.5

W j2 diagf8; 8g Rise time (s) 60 21 19 19


Peak 27 0 0 0
overshoot
(%)
Settling time 162 144 140 140
(s)
SSE (e) N/A 2494.8 2495.7 2487.3

Table 4 Fuzzy Sets for Case 2


Statistical experimentation on the performance of the controllers on system 1.5
I. B1
2
B
Iteration number SSE BFO- SSE PSO- SSE GA- B
3

FMPC FMPC FMPC 1


(y)

1 2486.29 2490.10 2491.26


2 2487.41 2490.90 2491.75
3 2489.06 2491.90 2491.94 0.5
4 2489.64 2493.93 2492.18
5 2489.66 2493.93 2492.20
6 2490.28 2493.93 2492.32
7 2491.12 2493.93 2493.25 0
8 2491.62 2493.93 2495.90 0 2 4 6 8 10 12 14 16 18
9 2491.68 2502.13 2498.79 y (t)
10 2492.37 2502.34 2498.80
Mean 2489.91 2494.70 2493.84 Fig. 6. Fuzzy sets chosen for the convolution model for system in case II.
S.D. 1.94 4.22 2.91
p-Value 0.022 0.011 improvements in rise time depict the versatility of the
Statistically signicant Yes Yes
designed FMPC controller. The peak overshoot being tre-
(p > 0.05)?
mendously low in most cases indicate that the controller
is safe to use in an industrial environment as the overow
of chemicals is reduced by the controllers accuracy. In this
respect, all three of the global optimization schemes per-
worse than either of the other controllers, showing highly form admirably, although the PSO algorithm tends to
oscillatory behavior in all three cases. The results of this sometimes waver from its desired goal, and hence is not
simulation is shown in Table 3. recommended for integration with the MPC for systems
It can very well be appreciated from Tables 2 and 3 that where overshoots could be deemed critical.
the proposed BFO-FMPC controller has consistently Finally, in order to determine the statistical signicance
outperformed the controller in [1], and also proved of our results, we tabulated the errors over 10 runs and
superior to the PSO-FMPC and GA-FMPC, with different calculated the mean, standard deviation as well as the
choices of s and W j2 and it is reected in all but one of p-values1 and t-test values in a second order two-sample t-
the performance indices. An important observation is the test on a 95% condence interval using MATLABs standard
drastic lowering of the rise times and the peak overshoots, ttest2 function. The p-value is an important metric in
which indicate that the controller performs very quickly to
reach a value very near the reference value ref y and then 1
p-Value indicates that each t-test was performed between BFO-FMPC
takes time to settle within a 2% error. Highly non-linear
errors and other respective control schemes, e.g. p-value in column 2 is the
chemical plants tend to be sluggish in response and the result of the test between BFO-FMPC and PSO-FMPC.
1626 A. Chakrabarty et al. / Measurement 46 (2013) 16161629

Impulse response of the convolution models used in modeling System II [1]


0.2

h1 (n)
0
0 10 20 30 40 50 60 70 80
Sample (N)

2
1.5
h (n)

1
2

0.5

0
0 10 20 30 40 50 60 70 80
Sample (N)

0.8

0.6
h (n)

0.4
3

0.2

0
0 10 20 30 40 50 60 70 80
Sample (N)

Fig. 7. Impulse responses of the fuzzy convolution models used to model system II.

statistical hypothesis testing, measuring how probable the R3 : IF yn is B3 THEN y3 n 1


test statistic is in comparison to some observed statistic. An X
80
3
inherent assumption is that the null hypothesis is true, and yn hi un 1  i; 34
the decision regarding whether a null-hypothesis is to be re- i1
jected depends on whether the p-value is less than a selected
small number a. In our study, this a is 0.05, indicating a 95% where the three fuzzy sets Bi,i 2 [1,3] are dened in Fig. 6.
condence interval in our test. The results of the ten runs are The impulse responses of the three convolution models are
shown and the statistical parameters reported in Table 4. We also shown in Fig. 7.
can see that the means, standard deviations as well as the p- Considering U = 2, V = 3 and T = 80, the impulse re-
values support the statistical signicance of the results and sponse coefcients shown in Fig. 7 are used to simulate
prove conclusively that the BFO-FMPC is an overall better the design model. The sampling time of the controller is ta-
control scheme than the others for this system. ken to be 10 s, as before. The weighting factors are given by
the same matrices as before. The reference model is also
4.2. Case study II chosen of the same form as in system I, but s = 0.88 min
in this model, and ysp = 6. The behavior of the controllers
After achieving encouraging results for the rst nonlin-
ear system, we selected a second system from [1] in order Direct Adaptive Controller Output
to conrm our observations. The second system selected 6
for simulation and performance evaluation contains a large
amount of noise, and hence the controller response to this 5
system proves its ability to handle noisy measurements
and perform satisfactorily in spite of model mismatch. 4
The fuzzy convolution model for this system is considered
yout

to have three FIs, that is, 3

R1 : IF yn is B1 THEN y1 n 1
2
X
80
1
yn hi un 1  i; 32 1
Reference
BFOFMPC
i1 PSOFMPC
GAFMPC
0
R2 : IF yn is B2 THEN y2 n 1 0 0.5 1 1.5 2 2.5 3
Time (t) [mins]
X
80
2
yn hi un 1  i; 33 Fig. 8. Comparison of controller performance on system II in the presence
i1 of disturbances.
A. Chakrabarty et al. / Measurement 46 (2013) 16161629 1627

Performance comparison when W=diag(25) Performance comparison when W=diag(15)


10 10
Reference Reference
BFOFMPC BFOFMPC
9 9
PSOFMPC PSOFMPC
GAFMPC GAFMPC
8 8

7 7

6 6
out

out
5 5
y

y
4 4

3 3

2 2

1 1

0 0
0 0.5 1 1.5 2 2.5 3 0 0.5 1 1.5 2 2.5 3
Time (t) [mins] Time (t) [mins]

Fig. 9. Closed loop dynamic response of system II with variation of W j2 , (a) W = diag(25), (b) W = diag(15).

Table 5
Quantitative performance comparison among competing controllers for case study II when the weighting factor is changed.

Variation of W Performance FMPC with iterative design Proposed FMPC with Proposed FMPC with Proposed FMPC with
matrix for system II indices based controller [1] PSO-based approach GA-based approach BFO-based approach

W j2 diagf25; 25g Rise time (s) 84 7 10 7


Peak 20 5 0 4
overshoot
(%)
1st Settling 300 37 136 28
time (s)
SSE (e) N/A 3104.15 3491.58 2860.25

W j2 diagf15; 15g Rise time (s) 54 17 7 7


Peak 33 52 4 3
overshoot
(%)
1st Settling 410 380 32 31
time (s)
SSE (e) N/A 3256.7 2889.6 2858.8

is plotted and shown in Fig. 8. All three controllers perform the case of the PSO-FMPC which shows sustained oscilla-
admirably on this system, but closer inspection shows that tions, proving its inefciency in this fuzzy paradigm. In
the BFO-MPC has advantages over the rest. The kinks pres- the presence of a noisy system, the BFO-FMPC clearly
ent in the output trajectory of the controller is due to the shows its superiority to the GA-FMPC by attaining the set
sudden variation in impulse value as seen in Fig. 6 at the point much earlier than the GA-FMPC as well as by show-
ninth sample of h1[n] and the rst sample of the others. ing relative noisy immunity whereas the GA-FMPC shows
This disturbance helps us to determine the controllers ex- minor oscillations throughout the simulation time in
ibility and reaction speed, and in both cases the BFO-FMPC Fig. 9b.
shows excellent performance by returning to the reference Next, we show the performance of the proposed BFO-
trajectory within 3045 s, whereas the others take closer FMPC controller with variation in set-point. Intuitively,
to 60 s to reduce the deviation between the control output this is the single most important performance test as the
and the reference to zero. Furthermore, the rise time of the controller is required to behave satisfactorily for any set
BFO-FMPC is smaller than the others. This evidence of its point and not just at ysp = 6.0 as shown till this stage.
swiftness is a great advantage, as speed is a trait which is Fig. 10 shows the performance of the BFO-FMPC at varying
crucial for noisy systems. set points from a range of 3.512.0. We also noted its per-
With variation of W j2 , with s = 0.88 min, the closed loop formance parameters with respect to the rise time, the 98%
dynamic response is plotted in Fig. 9 and metrics are com- settling time and the peak overshoot.
pared in Table 5. As before, the system performance re- From Table 6, we prove that for this case study too, our
mains largely unaffected with variations in W j2 , except in proposed controller has signicantly outperformed the
1628 A. Chakrabarty et al. / Measurement 46 (2013) 16161629

Performance Evaluation of BFOFMPC with variation of SetPoint Table 7


Statistical experimentation on the performance of the controllers on system
12 II.

Iteration number SSE BFO- SSE PSO- SSE GA-


10 FMPC FMPC FMPC
1 2860.25 3104.15 3491.58
8 2 2869.63 2871.82 2861.39
3 2860.45 3510.17 3614.22
yout

4 2864.30 3266.28 5728.32


6
5 2863.52 3299.48 2526.30
6 2863.87 3266.28 3959.49
4 7 2867.24 3321.10 3482.12
8 2858.29 3266.28 2861.82
9 2867.06 3096.36 3537.12
2
10 2862.19 3265.45 2863.12
Mean 2863.68 3226.74 3492.55
0 S.D. 3.55 169.54 904.29
0 0.5 1 1.5 2 2.5 3
p-Value 1.2  106 0.0206
Time (t) [mins] Statistically signicant Yes Yes
(p > 0.05)?
Fig. 10. Closed loop dynamic response under varying set points.

Table 6 important when it is kept in mind that the system II con-


Comparison of performance parameters between the two methods for tains large noise in its open-loop dynamic response [1].
system II when the set point is changed.
As in the previous case study, we have computed SSE val-
Variation of Performance FMPC with Proposed ues for 10 iterations and performed statistical tests upon
set points indices iterative design FMPC with them, using MATLABs test22 function upon a 95% con-
(ysp) for based controller BFO-based
dence interval. The results are shown in Table 7. Thus, it
system II [1] approach
can be said that in both case studies, the BFO-FMPC provides
ysp = 3.5 Rise time (s) 102 8
signicantly better control performances as compared to the
Peak 23 0
overshoot other control schemes implemented in this work. This con-
(%) clusively demonstrates the efcacy of our proposed FMPC
1st Settling 300 36 control paradigm in the control of highly non-linear noisy
time (s)
chemical plants with fuzzy convolution modeling.
ysp = 6.0 Rise time (s) 84 7
Peak 20 4
overshoot 5. Conclusion
(%)
1st Settling 288 33 The present paper proposed a novel methodology of
time (s)
designing fuzzy model predictive controllers employing
ysp = 10.0 Rise time (s) 66 6 modern stochastic optimization algorithms. The system
Peak 20 4
utilizes fuzzy convolution models which can describe a
overshoot
(%)
highly nonlinear system utilizing several quasi-linear fuzzy
1st Settling 245 22 implications. This proposed methodology utilizes a mod-
time (s) ern bio-inspired optimization algorithm, called Bacterial
ysp = 12.0 Rise time (s) 54 12 Foraging Optimization algorithm, for controller design,
Peak 8 4 which can efciently determine the optimal controller
overshoot structure and plant parameters. The performance of the
(%)
proposed BFO-FMPC controller is compared vis--vis the
1st Settling 210 37
time (s) FMPC controller proposed in [1] by Huang et al. and
GA-FMPC and PSO-FMPC controllers for two benchmark,
challenging, highly non-linear chemical processes and the
superiority of the proposed BFO-FMPC controller is rmly
controller in [1], with respect to all ascertained perfor- established.
mance indices. Here also, for our controller, there is less
than 5% peak percent overshoot in the proposed approach References
for all four set-points varied over a wide range. It was
noted that the performance deteriorates greatly if the run [1] Y.L. Huang, H.H. Lou, J.P. Gong, T.F. Edgar, Fuzzy model predictive
control, IEEE Trans. Fuzzy Syst. 8 (6) (2000) 665677.
length of the bacteria C[i] as described in [24] is greater
[2] R.A. Abou-Jeyab, Y.P. Gupta, J.R. Gervais, P.A. Branchi, S.S. Woo,
than 0.05 for the higher set-point values. This is because Constrained multivariable control of a distillation column using a
a very small change in any element in the feedback matrix
K causes a large variation in the response for that run of 2
p-Value indicates that each t-test was performed between BFO-FMPC
the BFO-FMPC. Also, the vast improvement achieved over errors and other respective control scheme, e.g. p-value in column 2 is the
settling time as evident from Table 6 becomes all the more result of the test between BFO-FMPC and PSO-FMPC.
A. Chakrabarty et al. / Measurement 46 (2013) 16161629 1629

simplied model predictive control algorithm, J. Process Contr. 11 [29] S. Banerjee, A. Chakrabarty, S. Maity, A. Chatterjee, Feedback
(5) (2001) 509517. linearizing indirect adaptive fuzzy control with foraging based on-
[3] R.M. Ansari, M.O. Tade, Non-Linear Model Based Process Control: line plant model estimation, Appl. Soft Comput. 11 (2011) 3441
Applications in Petroleum Rening, vol. 11, Springer, London, 3450.
England, 2000. 12. [30] S. Kamalasadan, A.A. Ghandakly, A neural network parallel adaptive
[4] J.G. Bekker, I.K. Craig, P.C. Pistorius, Model predictive control of an controller for dynamic system control, IEEE Trans. Instrum. Meas. 56
electric arc furnace oilgas process, Contr. Eng. Pract. 8 (2000) 445 (5) (2007) 17861796.
455. [31] A. Jnifene, W. Andrews, Experimental study on active vibration
[5] W.V. Brempt, T. Backx, J. Ludlage, P.V. Overschee, B.D. Moor, R. control of a single-link exible manipulator using tools of fuzzy logic
Tousain, A high performance model predictive controller: and neural networks, IEEE Trans. Instrum. Meas. 54 (3) (2005) 1200
application on a polyethylene gas phase reactor, Control Eng. Pract. 1208.
9 (8) (2001) 829835. [32] N. Karlsson, B. Karlsson, P. Wide, A glove equipped with nger
[6] R.K. Mehra, R. Rouhani, Theoretical considerations on model exion sensors as a command generator used in a fuzzy control
algorithmic control for non-minimum phase systems, in: Joint system, IEEE Trans. Instrum. Meas. 47 (5) (1998) 13301334.
Automatic Control Conf., San Francisco, CA, 1980. [33] S. Kamalasadan, A.A. Ghandakly, Multiple fuzzy reference model
[7] C. Onnen, R. Babuska, U. Kaymak, J.M. Sousa, H.B. Verbruggen, R. adaptive controller design for pitch-rate tracking, IEEE Trans.
Isermann, Genetic algorithms for optimization in predictive control, Instrum. Meas. 56 (5) (2007) 17971808.
Contr. Eng. Pract. 5 (10) (1997) 13631372. [34] A. Ferrero, A. Federici, S. Salicone, Instrumental uncertainty and
[8] D.M. Prett, R.D. Gilette, Optimization and constrained multivariable model uncertainty unied in a modied fuzzy inference system, IEEE
control of a catalytic cracking unit, in: AIChE Annual Meeting, Instrum. Meas. 59 (5) (2010) 11491157.
Houston, TX, 1979. [35] S. Ma, T. Feng, H. Zhang, J. Xue, Optimization of preventive
[9] L.X. Wang, Adaptive Fuzzy Systems and Control, Prentice-Hall, maintenance period based on hybrid swarm intelligence, IEEE
Englewood Cliffs, NJ, 1994. Trans. Instrum. Meas. 5 (2010) 26562659.
[10] S.W. Kim, E.T. Kim, M. Park, A new adaptive fuzzy controller using [36] B. Bhushan, M. Singh, Adaptive control of DC motor using bacterial
the parallel structure of fuzzy controller and its application, Fuzzy foraging algorithm, Appl. Soft Comput. 11 (8) (2011) 4913
Sets Syst. 81 (2) (1996) 205226. 4920.
[11] Y. Nakamori, K. Suzuki, T. Yamanaka, Model predictive control using [37] M. Kiani, H.N. Soloklo, M.A. Mohammadi, M.M. Farsangi, Design of a
fuzzy dynamic models, in: Proc. IFSA 91 Brussels, vol. 135, Brussels, supplementary controller for power system stabilizer using bacterial
Belgium, vol. Engineering, July 1991, pp. 138. foraging optimization algorithm, Comput. Intell. Inform. Technol.
[12] T. Takagi, M. Sugeno, Fuzzy identication of systems and its 250 (2) (2011) 405410.
application to modeling and control, IEEE Trans. Syst., Man, [38] M. Turduev, M. Kirtay, P. Sousa, V. Gazi, L. Marquez, Chemical
Cybern. SMC-15 (1985) 116132. concentration map building through bacterial foraging optimization
[13] K.M. Passino, S. Yurkovich, Fuzzy Control, Addison Wesley Longman, based search algorithm by mobile robots, IEEE Int. Conf. Syst. Man,
Inc, 1998. Cybern. (2010) 32423249.
[14] C. Lee, Fuzzy logic in control systems: fuzzy logic controller Parts [39] J. Nanda, S. Mishra, L.C. Saikia, Maiden application of bacterial
III, IEEE Trans. Syst., Man, Cybern. 24 (5) (1994) 736744. foraging-based optimization technique in multiarea automatic
[15] T. Ross, Fuzzy Logic in Engineering Applications, McGraw-Hill, New generation control, IEEE Trans. Power Syst. 24 (2) (2009) 602609.
York, 1995. [40] J. Zhang, P. Shi, Y. Xia, Robust adaptive sliding-mode control for
[16] L.X. Wang, A Course in Fuzzy Systems and Control, Prentice-Hall, fuzzy systems with mismatched uncertainties, IEEE Trans. Fuzzy
Englewood Cliffs, NJ, 1997. Syst. Vol. 18 (4) (2010) 700711.
[17] D. Driankov, H. Hellendoorn, M. Reinfrank, An Introduction to Fuzzy [41] K. Das Sharma, A. Chatterjee, A. Rakshit, A hybrid approach for
Control, Springer-Verlag, New York, 1993. design of stable adaptive fuzzy controllers employing Lyapunov
[18] J.R. Layne, K.M. Passino, Fuzzy model reference adaptive control for theory and particle swarm optimization, IEEE Trans. Fuzzy Syst. 17
cargo ship steering, IEEE Contr. Syst. Mag. 13 (6) (1993) 2334. (2) (2009) 329342.
[19] W.A. Kwong, K.M. Passino, Dynamically focused fuzzy learning [42] K. Das Sharma, A. Chatterjee, A. Rakshit, Design of a hybrid stable
control, IEEE Trans Syst., Man, Cybern. 26 (1) (1996) 5374. adaptive fuzzy controller employing Lyapunov theory and harmony
[20] K.M. Passino, Biomimicry of bacterial foraging for distributed search algorithm, IEEE Trans. Contr. Syst. Technol. 18 (6) (2010)
optimization and control, IEEE Contr. Syst. Mag. 22 (3) (2002) 5267. 14401447.
[21] M. Maitra, A. Chatterjee, A novel technique for multilevel optimal [43] K. Das Sharma, A. Chatterjee, A. Rakshit, A random spatial lbest PSO-
magnetic resonance brain image thresholding using bacterial based hybrid strategy for designing adaptive fuzzy controllers for a
foraging, Meas., Sci. Direct 41 (10) (2008) 11241134. class of nonlinear systems, IEEE Trans. Instrum. Meas. 61 (6) (2012)
[22] S. Dasgupta, S. Das, A. Abraham, A. Biswas, Adaptive Comput. 16051612.
Chemotaxis Bacterial Foraging Opt.: An Anal. 13 (4) (2009) 919. [44] S. Bhattacharya, A. Chatterjee, S. Munshi, An improved PID-type
[23] Z.P. Liu, Y.L. Huang, Fuzzy model-based optimal dispatching for NO fuzzy controller employing individual fuzzy P, fuzzy I and fuzzy D
reduction in power plants, Int. J. Electr. Power Energy Syst., vol. 20, controllers, Trans. Inst. Meas. Contr. 25 (4) (2003) 352372.
Elsevier, 1998, pp. 169176. 3. [45] S. Bhattacharya, A. Chatterjee, S. Munshi, A new self-tuned PID-type
[24] H.H. Lou, Y.L. Huang, Fuzzy logic based process modeling using fuzzy controller as a combination of two-term controllers, ISA Trans.
limited experimental data, Int. J. Eng. Appl. Artif. Intell. 13 (2) (2000) 43 (3) (2004) 413426.
121135. [46] A. Chatterjee, K. Watanabe, An adaptive fuzzy strategy for motion
[25] M. Jamshidi, Large-Scale Systems: Modeling and Control, vol. 9, control of robot manipulators, Soft Computing, vol. 9, Springer
Elsevier Science, New York, 1983. Publishers, 2005, pp. 185193. 3.
[26] A. Zhe Song Kusiak, Opt. Temp. Process.: A Model Predictive Contr. [47] A. Chatterjee, R. Chatterjee, F. Matsuno, T. Endo, Augmented stable
Approach 13 (1) (2009) 169. fuzzy control for exible robotic arm using LMI approach and neuro-
[27] http://www.ece.osu.edu/passino/ICbook/Code/foragingadaptiveind.m. fuzzy state space modeling, IEEE Trans. Ind. Electr. 55 (3) (2008)
[28] J.M. Maciejowski, Predictive Control with Constraints, Prentice Hall, 12561270.
Englewood Cliffs, NJ, 2002.

You might also like