Soft Computing

https://doi.org/10.1007/s00500-020-05133-x (0123456789().,-volV)(0123456789().,-volV)

METHODOLOGIES AND APPLICATION

Robust controller design for systems with probabilistic uncertain

parameters using multi-objective genetic programming
Rammohan Mallipeddi1 • Iman Gholaminezhad2 • Mohammad S. Saeedi2,3 • Hirad Assimi4 •

Ali Jamali2,3

Ó Springer-Verlag GmbH Germany, part of Springer Nature 2020

Abstract
Optimal design of controllers without considering uncertainty in the plant dynamics can induce feedback instabilities and
lead to obtaining infeasible controllers in practice. This paper presents a multi-objective evolutionary algorithm integrated
with Monte Carlo simulations (MCS) to perform the optimal stochastic design of robust controllers for uncertain time-
delay systems. Each potential optimal solution represents a controller in the form of a transfer function with the optimal
numerator and denominator polynomials. The proposed methodology uses genetic programming to evolve robust con-
trollers. Using GP enables the algorithm to optimize the structure of the controller and tune the parameters in a holistic
approach. The proposed methodology employs MCS to apply robust optimization and uses a new adaptive operator to
balance exploration and exploitation in the search space. The performance of controllers is assessed in the closed-loop
system with respect to three objective functions as (1) minimization of mean integral time absolute error (ITAE), (2)
minimization of the standard deviation of ITAE and (3) minimization of maximum control effort. The new methodology is
applied to the first-order and second-order systems with dead time. We evaluate the performance of obtained robust
controllers with respect to the upper and lower bounds of step responses and control variables. We also perform a post-
processing analysis considering load disturbance and external noise; we illustrate the robustness of the designed controllers
by cumulative distribution functions of objective functions for different uncertainty levels. We show how the proposed
methodology outperforms the state-of-the-art methods in the literature.

Keywords Genetic programming  Robust controller  Pareto  Monte Carlo simulation  Uncertainty

1 Introduction

Recently, intelligent controller design methods have

become propitious among researchers. Evolutionary algo-
rithms (EAs) have evolved as efficient techniques for
Communicated by V. Loia. tackling a variety of real-world optimization problems
(Mallipeddi et al. 2011; Jamali et al. 2013; Assimi et al.
& Ali Jamali
[email protected] 2017; Uyeh et al. 2018; Awad et al. 2019; Biswas et al.
2019; Kalat 2019). Among EAs, genetic programming
Rammohan Mallipeddi
[email protected] (GP) has been previously employed for controller design in
various industrial processes, because GP can optimize the
1
Department of Artificial Intelligence, College of IT, structure of a controller and tune its parameters in a holistic
Kyungpook National University, Daegu 41566, South Korea approach (Koza et al. 2003; Gomes et al. 2019; Chiou
2
Faculty of Mechanical Engineering, University of Guilan, 2018).
Rasht, Iran In the literature, GP has been used for the optimal design
3
Department of Mechanical Engineering, University Campus of controllers: In Fukunaga et al. (2012), a simulation-
2, University of Guilan, Rasht, Iran based GP was proposed to design a controller for a service
4
School of Computer Science, University of Adelaide, robot for managing the visiting traffic flow in an exhibition
Adelaide, SA 5005, Australia

123
 R. Mallipeddi et al.

space. In Kumaresan and Ratnavelu (2014), GP was cou- shortcoming of FORM and SORM lies in the calculation of
pled with a neuro-fuzzy system to solve the matrix Riccati most probable points, while there is derivative disconti-
differential equation for optimal control. Two GP approa- nuity on the limit state function (Smith et al. 2005; Field
ches (GP1 and GP2) to control linear continuous-time et al. 1996). H1 method considers the worst-case scenario
processes were proposed in Sekaj and Perkacz (2007). to design, which leads to over-conservative results (Di-
They observed the performance of GP1 and GP2 on two wekar and Kalagnanam 1997).
linear benchmarks in comparison with a genetic algorithm Robust design optimization (RDO) integrates sampling
(GA)-based PID controller. They reported that GP could techniques with optimization methods to obtain a robust
outperform GA with respect to the integral of absolute design (Diwekar and Kalagnanam 1997; Wang and Stengel
error. They showed that the most critical issue with GP- 2002; Kalos and Whitlock 1998). RDO aims to reduce the
based controllers is their high execution time. In Kadlic sensitivity of the system due to random variations of
et al. (2014), an approach based on Cartesian GP (CGP) uncertain parameters (Kang 2005; Nariman-Zadeh et al.
was proposed to reduce the execution time in Sekaj and 2007; Jamali et al. 2010). Therefore, RDO minimizes the
Perkacz (2007). CGP has been applied to the problem of mean and standard deviation of the desired objective
water turbine control design and nonlinear Duffing oscil- function.
lator (Balandina 2017). In Gholaminezhad et al. (2014), a From sampling techniques, Monte Carlo simulation
multi-objective GP was proposed for the deterministic (MCS) is commonly used for robust design to evaluate how
controller design problem in the time domain. In Hu et al. a control system responds to random inputs (Biwer et al.
(2018), an automatic crowd control framework to generate 2005; Bi et al. 2013). MCS generates random samples
control strategies based on Pareto multi-objective opti- according to the predefined probabilistic distribution of
mization of genetic programming was proposed. uncertain parameters. MCS simulates the system with each
Real-world engineering systems such as industrial con- randomly generated sample and minimizes the mean of the
trol systems inherently possess uncertainties that can be objective function of the samples; it also minimizes the
classified as aleatory and epistemic uncertainties (Bodla standard deviation of the objective function to achieve
et al. 2013; Pettersson et al. 2015; Darbra et al. 2008; robustness in the system.
Mohammadzadeh and Taghavifar 2020). Aleatory uncer- MCS and optimization algorithms have been coupled to
tainty refers to the uncertainty in the system parameters, design controllers: A numerical optimization approach for
which is typically handled by probabilistic methods. robust PI/PID controller design was proposed in Toscano
However, epistemic uncertainty arises due to the lack of (2005). In Jamali et al. (2013), Nariman-Zadeh et al.
knowledge and can be alleviated by improving our (2007), Jamali et al. (2010) and Hajiloo et al. (2008),
knowledge of the system. reliability-based design of controllers using genetic algo-
Neglecting uncertainties in the optimal design of a rithms and MCSs were investigated. Multi-objective
controller can lead to unstable solutions which are infea- genetic algorithms and GMDH neural networks were
sible in practice (Jamali et al. 2013, 2014). Therefore, we employed to estimate the reliability of settling time and
aim to design a robust controller to reduce performance overshoot in the controllers (Jamali et al. 2013; Hajiloo
variation of a controller in a noisy environment. In this et al. 2008). Later, the authors used the same approach for
study, we consider aleatory uncertainties which are asso- Pareto optimization of PI and PID controllers for a set of
ciated with the dynamic parameters of the plant. time-delay systems. In Hajiloo et al. (2008), robust and
Different methodologies have been proposed to reliability-based optimization techniques are integrated for
approximate the reliability of a control system including the stochastic optimal design of PI controllers.
polynomial chaos expansion (Eldred 2009; Pettersson et al. In current work, we aim to automatically detect the
2015), stochastic collocation method (Zhao et al. 2015; optimum structure of a transfer function for uncertain
Witteveen and Iaccarino 2012), dimension reduction (Li systems based on multi-objective GP (MOGP) integrated
and Ma 2013) and Taguchi method (Nejlaoui et al. 2013). with MCS. Each potential optimal robust controller is
Taguchi method designated the design with the most sig- represented in the form of a transfer function with the
nificant signal/uncertainty ration as an optimal robust optimal numerator and denominator polynomials. Our
design. methodology uses a new adaptive real-value operator to
Taguchi method has a shortcoming: It assumes that the make a balance between exploration and exploitation in the
design variables and uncertainty are independent. The first- search space. This feature enables GP to perform parameter
order reliability method (FORM) and second-order relia- tuning of controllers more efficiently. Time-domain simu-
bility method (SORM) estimate the reliability of objective lation of each controller evaluates the performance of each
functions by linear approximation and hyper-parabola controller. The simulation is carried out on the first- and
using surface response, respectively. The critical second-order systems with dead time in a closed-loop

123
Robust controller design for systems with probabilistic uncertain parameters

system. The objective functions considered are: (1) mean transfer function. Conversely, each expression tree (left
of integral time absolute error (ITAE), (2) standard devi- and right halves in Fig. 2) parses into one polynomial each
ation of ITAE and (3) maximum control effort. We depict corresponding to the polynomials of numerator and
the performance of obtained robust controllers with step denominator of K(s). The advantage of employing this
responses, control variables, load disturbances and external representation is that the order of polynomials in the
noise in a post-processing analysis. We also illustrate the transfer function (controller complexity) can be controlled
robustness of the obtained controllers with probability by specifying the length of tree structures. Hence, the
density functions and cumulative distribution functions of structures of the evolved transfer functions are automati-
objective functions for different uncertainty levels. Finally, cally tuned during the optimization process without any
we compare the obtained results with state-of-the-art predefined specifications. However, the symbolic repre-
stochastic optimization methods in the literature. sentation in stochastic robust controller design results in a
very complex searching space. Therefore, the design con-
straints described below should be carefully taken into
2 Methodology account so that the stochastic search space is bounded and
helps the optimization process.
2.1 Evolution of controllers using genetic In transfer function-based representation of controllers,
programming the degree of numerator polynomial should be less than or
equal to the degree of denominator polynomial. However,
Figure 1 shows a simple closed-loop feedback control the tree-based representation may result in invalid con-
system including a dynamic controller system (K(s)), trollers, where the degree of the numerator is greater than
dynamic continuous-time controlled model (P(s)), refer- the denominator, during the evolution process of the GP
ence signal (R(s)), noise input (N(s)), disturbance (d1(s)and algorithm. Hence, the invalid transfer functions should be
d2(s)), control error signal, controller variable and output penalized during the evolutionary optimization procedure.
signal (Y(s)). s refers to the complex variable which is In addition, transfer functions which break the stability
s ¼ r þ jx. condition based on the location of the poles of the closed-
A transfer function can represent a controller in con- loop system are also penalized. Finally, a steady state error
tinuous time domain which is constraint is specified such that the transfer functions with
bm sm þ bm1 sm1 þ . . . þ b1 s þ b0 small control efforts and that do not reach steady state are
K ðsÞ ¼ ð1Þ penalized in the evolutionary process.
an sn þ an1 sn1 þ . . . þ a1 s þ a0
The uncertain variables related to the plant are internal
Performance of the control system can be obtained by gain, time constant and dead time. The optimal controller or
closed-loop and open-loop simulation. Closed-loop simu- transfer function evolved considering these uncertain
lation is essential to calculate the objective functions pre- parameters should result in a control system where the
sented in Table 1 for controller design in the time domain. variation of plant conditions has the least effect on output of
GP can generate a tree-based structure to evolve the the closed-loop system. Therefore, the evolved controller
polynomials of K(s) without solving the complex differ- should guarantee stability of the system in the whole region
ential equation. Figure 2 depicts the tree structure of a of uncertain space and also assure efficient performance in
sample transfer function that corresponds to Eq. (2): terms of transient and steady state closed-loop indices.
sþ1 Table 1 lists various choices of objective functions for
K ðsÞ ¼ ð2Þ optimal design of controllers in time domain. Note that the
s2 þ 3s þ 2
choice of the objective function depends on the designer
Each syntax tree (Fig. 2) parses to an expression, which preference. Integral of the squared error (ISE) can elimi-
constitutes both numerator (left half in Fig. 2) and nate large errors quickly for fast response, but small errors
denominator (right half in Fig. 2) polynomials of the persist for a long time period. Integral of the absolute error
(IAE) responses are slower than ISE, but it can handle
small errors more efficiently. On the other hand, integral of
the time-weighted absolute error (ITAE) can help tune
systems to settle more quickly in comparison with other
methods and avoid sustained oscillations (Krishnan and
Karpagam 2013).
In this study, we consider ITAE for tuning the con-
trollers. We set the objective functions as the minimization
Fig. 1 A closed-loop feedback control system of mean of ITAE, standard deviation of ITAE and

123
 R. Mallipeddi et al.

Table 1 Various objective

Objective function Time-domain index
function choices for control
system design using multi- Integral of the absolute error (IAE) RTf
objective optimization JIAE ðhÞ ¼ jr ðtÞ  yðtÞjdt
t¼t0

Integral of the time-weighted absolute error (ITAE) RTf

JITAE ðhÞ ¼ tjr ðtÞ  yðtÞjdt
t¼t0

Integral of the squared error (ISE) RTf

JISE ðhÞ ¼ ðr ðtÞ  yðtÞÞ2 dt
t¼t0

Integral of the time-weighted squared error value (ITSE) RTf

JITSE ðhÞ ¼ tðr ðtÞ  yðtÞÞ2 dt
t¼t0

Total variation of control action RTf du

JTV ðhÞ ¼  dt 
t¼t 0

Integral of control action value RTf

JISU ðhÞ ¼ ðuðtÞÞ2 dt
t¼t0
 
Maximum value of control action JMaxU ðhÞ ¼ ðuðtÞÞ; t 2 t0 ; T f

design. The MOGP algorithm benefits from a new adaptive

real-value alteration operator in addition to the traditional
evolutionary mutation and crossover operators. The basic
crossover and mutation operator of GP is shown in Fig. 4.
As shown, crossover selects two parents based on tourna-
ment selection, cuts both parents at random points and
exchanges the segments between each other to generate
two new offsprings. Also, mutation operator cuts the tree
structure from a random cut-point and replace it with a
generated tree at the cut-point. From the implementation, it
is evident that both crossover and mutation operators in
Fig. 4 only change the tree structure and could not change
the elements (value) in its terminals. Therefore, GP basic
operators have a limitation which lead to significant
reduction in the number of solutions generated through the
basic crossover and mutation operators.
To address this drawback and improve GP exploration
ability, we propose a new adaptive operator for MOGP that
can change the terminal value of tree structures. The pro-
Fig. 2 Tree structure of a sample transfer function evolved by GP
posed operator referred to as real-value alteration operator
is described by the following equation:
maximum of control action are considered as the important n ¼ a þ F ðb  cÞ ð3Þ
goals of the optimization process. By minimization of
control action in the stochastic uncertain space, the mini- a, b and c represent three randomly selected values of
mum energy consumption of the control system will be terminals of a tree structure. n is the new integer corre-
satisfied. Figure 3 depicts the flowchart of our proposed sponding to the terminal of tree structure that is replaced.
approach. The proposed operator in Eq. (3) inspired from differential
evolution mutant operator (Jamali et al. 2014; Gho-
2.2 Multi-objective genetic programming laminezhad and Jamali 2016). The adaptive parameter F is
(MOGP) for robust controller design changed according to Eq. (4), and its variation versus g is
shown in Fig. (5). The decrease in the value of F from 0.9
In this study, a multi-objective genetic programming in early generations to 0.05 in the last generation enables
algorithm (MOGP) is applied to Pareto robust controller the algorithm to switch from exploration in the earlier

123
Robust controller design for systems with probabilistic uncertain parameters

Fig. 3 The flowchart of proposed method

Fig. 4 Basic crossover and mutation operators in GP

generations to exploitation in the later generations of the uses non-dominated sorting and crowding distance algo-
evolution process. As shown in Fig. (5), the rate of change rithms in the selection procedure of multi-objective opti-
of F is divided in two parts. In the first part (0:1  g  0:5), mization process (Jamali et al. 2008, 2013; Nariman-Zadeh
F decreases exponentially, while in the second part et al. 2010):
(0:5\g  1), the decrease is linear. In addition, MOGP

123
 R. Mallipeddi et al.

system parameters. Each sample controller is simulated, and

mean and variance of total samples can be calculated.

3.2 Multi-objective robust controller design

optimization problem formulation

We formulate the robust design optimization as follows:

Finding: controller structure
To minimize: F ðX; pÞ ¼ flðITAEÞ; rðITAEÞ; Maxðcf Þg
 
Subjected to : Pr Pr gj ðX; pÞ  0  Rj j ¼ 1; 2; . . .; m
ð8Þ
where X and p are represented as design variables (con-
troller structure) and uncertain parameters, respectively.
Fig. 5 Variation of adaptive parameter F versus g Also, cf is the control effort (control value) of a designed
controller. lðITAEÞ; rðITAEÞ andMaxðcf Þ are shown
2:07ð1:1  gÞ mean of ITAE, variation of ITAE and maximal control
F¼ ð4Þ
lnð100gÞ effort value, respectively. The constraint represents the
Current generation probability of failure, and gj ðX; pÞ is given by:
g¼
Maximum generation gj ðX; pÞ ¼ f15  ST  0MaxðrealðsÞÞ  0 ð9Þ
where ST and MaxðrealðsÞÞ represent the settling time of
step response and maximum real value of roots of the
3 Stochastic robustness and multi-objective denominator of closed-loop transfer function, respectively.
robust controller design optimization To calculate Maxðcf Þ, the system is simulated for each
problem formulation random sample generated by MCS and then the maximum
force that is applied to the plant is considered as the third
3.1 Stochastic robustness objective function. In this paper, an acceptable value of
Rj ¼ 0:001 is considered as the probability of failure.
Let X be a random variable, and a common model to
quantify uncertainty is the probability density function,
fX ð xÞ, or equivalently the cumulative distribution function, 4 Simulations and results
FX ð xÞ, which is given as
Zx We apply the proposed approach for Pareto optimal design
FX ð xÞ ¼ Pr PrðX  xÞ ¼ fX ð xÞdx ð5Þ of controllers with probabilistic constraints. We consider
two systems of the first order and second order. In account
1
for the uncertainty, MCS generates 100 random samples
where Pr(.) refers to the probability that a particular event with  5% Gaussian probabilistic distribution around each
ðX  xÞ has occurred. Using discrete sampling, mean value candidate controller. We use inverse cumulative distribu-
(lð X Þ) and variance (r2 ð X Þ) are given as follows: tion to generate samples. The produced samples are kept
unchanged during the optimization process. Table 2 shows
1X N
lð X Þ ffi xi ð6Þ the MOGP parameters used for optimizing the controllers.
N i¼1
4.1 Optimum robust controller design
1 X N
r2 ð X Þ ¼ Varð X Þ ffi ðxi  lð xÞÞ2 ð7Þ for the first-order system with time delay
N  1 i¼1

where xi and N denote the ith sample and total number of First-order plus dead-time systems are common in indus-
samples. trial system which can be modeled as follows (Toscano
We aim to minimize the mean and standard deviation 2005):
(rð X Þ) to achieve less variability in a stochastic environment. Ket0 s
P1 ðsÞ ¼ ð10Þ
Therefore, sample controllers are randomly generated with 1 þ ss
respect to the predefined statistical distribution of uncertain

123
Robust controller design for systems with probabilistic uncertain parameters

Table 2 Multi-objective genetic

Parameter Value
programming parameters
Population size 200
Number of iterations 250
Number of MCS samples 100
Maximum initial depth 6
Maximum depth 15
Probability of crossover 0.9
Probability of mutation 0.1
Probability of integer alteration operator 0.6
Fitness selection method Tournament selection
Tournament size 4
Stopping criteria Maximum generation
Original population Half full tree, Half grow tree
e-elimination threshold value 0.01
Function set *, ?, -, /, ^
Terminal set fS; randg

shown number of non-dominated points appearing in the

where K is the plant’s internal gain, s is the plant’s time
obtained Pareto front. In this study, 1-norm based on
constant and t0 is the dead time of the plant. In this case
Eq. (13) is used for such representation:
study, nominal values of plant parameters, K ¼ s ¼ t0 ¼ 1,
are considered for optimization. The simulation performed X
s

for 15 s based on an in-house developed MATLAB code. kJ ð hÞ k1 ¼ jJ i ðhÞj; 0  kJ ð hÞ k1  s ð13Þ

i¼1
Note that this setting is only used for simplified analysis.
However, we may observe multiple lags non-minimum This norm can be helpful in trade-off analysis between
phase zero in the actual process (Toscano 2005). We different objectives because it can provide valuable infor-
consider K, t0 and s as uncertain parameters with  5% mation such as non-convexity of Pareto fronts and the
uncertainty with Gaussian distribution in the multi-objec- worst objective for a specific point (Blasco et al. 2008).
tive optimization process. We carry out a three-objective Figure 6 depicts the level diagram of all three objective
optimization procedure for Pareto optimal robust controller functions based on 1-norm. It can be seen that lower values
design based on the method defined in Sect. 2.1. of 1-norm are located near the extremes (best values) of
We obtain 200 non-dominated solutions after 250 iter- objective functions. In addition, V layouts are obvious in
ations. In this research, level diagrams are employed in the figures, especially for mean ITAE and control efforts
order to better illustrate the Pareto fronts and the important which show the conflict between various objectives. For
design points based on different preferences. Level dia- control effort objective, it can be seen that most of the
grams are helpful to classify the non-dominated solutions points are concentrated under the value of 2.5.
in Pareto front with respect to the proximity to the extreme Based on these figures, the trade-off design point can be
points of Pareto front (Blasco et al. 2008; Zio and Bazzo chosen considering the distance of the points to the lowest
2011). For this purpose, we normalize every objective value. In this case, the lowest point in the all 1-norm fig-
considering its minimum and maximum values on the ures (point A) corresponds to a single individual with mean
Pareto front approximation as follows (Blasco et al. 2008): ITAE of 1.85, SD ITAE of 0.004 and control effort of 1.38.
The transfer function of this controller is of fifth order
JkM ¼ Ji ðhÞ; Jkm ¼ min Ji ðhÞ; i ¼ 1; . . .; l; k ¼ 1; 2; 3
h2Hp whose structure demonstrates in Eq. (14):
ð11Þ 6:802s5 þ 5:212e1s4 þ 7:721e2s3 þ 1:099e3s2 þ 5:072e2s þ 3:263e1
K ðsÞ ¼
8:686s5 þ 6:022e1s4 þ 8:144e2s3 þ 8:096e2s2 þ 5:712e1s
Ji ðhÞ  Jkm ð14Þ
J i ð hÞ ¼ ; 0  J i ðhÞ  1: ð12Þ
JiM  Jkm
We can designate another solution from the non-domi-
where Ji ðhÞ is the objective functions vector and J i ðhÞ is nated Pareto front with respect to the designer preferences.
the normalized objective functions vector, and also, l is the For instance, if better performance regarding control effort

123
 R. Mallipeddi et al.

Fig. 6 1-norm level diagrams representation for objective functions of the first-order system

is preferred, then another solution with better control effort - 0.0764 can be easily eliminated to reduce the controller
can be chosen. In this case, a simple structure PI controller to fourth order as follows:
with better control effort value of 0.98, mean ITAE of 2.17 6:802s4 þ 5:16e1s3 þ 7:682e2s2 þ 1:040e3s þ 4:27:9e2
and SD ITAE of 0.057 is selected from the Pareto set (point K ðsÞ ¼
8:686s4 þ 5:956e1s3 þ 8:099e2s2 þ 7:477e2s
B) whose structure depicts in Eq. (15): ð16Þ
0:469
K ðsÞ ¼ 0:530 þ ð15Þ Figure 7a illustrates the upper and lower bounds of step
s
responses for the two (Eqs. 15 and 16) designated optimum
It is worth to note that, in the selection of the trade-off robust controllers by MOGP. In addition, this figure pre-
design, other preferences such as controller structure can sents the same statistical performance of other suggested
also be taken into account. Such priorities can be consid- robust controllers in the literature for the same system.
ered as a post-processing step in the controller design Furthermore, Fig. 7b depicts the upper and lower bounds of
procedure as well as during the optimization process. For control variables for the same research works.
example, the structure of transfer functions can be con- As it can be seen, both suggested controllers have rel-
trolled by a constraint definition process. atively good performance compared to the other ones. It is
In addition, a zero/pole cancelation procedure can be shown that the proposed MOGP based on adaptive real-
applied to the fifth-order trade-off controller (Eq. 14) to value alteration operator has good performance in explo-
reduce its structure without affecting the closed-loop ration and exploitation of search space and finding the
response. This controller has zeros in [- 3.0718 ± 9.6972, optimum structure of transfer functions for uncertain
- 0.7212, - 0.2962, - 0.0763] and poles in [0, - 2.9361, systems.
± 8.8790, - 0.9843, - 0.0764]. From the set of zeros and The corresponding objective function values and trans-
poles, it can be seen that the zero - 0.0763 and pole of fer functions of the suggested controllers of this work and

123
Robust controller design for systems with probabilistic uncertain parameters

Fig. 7 a Upper and lower

bounds of time responses of the
controlled system output for
various robust controllers,
b upper and lower bounds of
control values of various robust
controllers for the first-order
system

of other works for the first-order system are tabulated in possesses much lower control effort with respect to the
Table 3. trade-off design. Overall, all the suggested controllers by
Table 3 shows that the mean and SD of ITAE for the various researches that are given in Table 3 are non-dom-
proposed trade-off controller is the lowest among all other inated to each other when considering all objective
controllers. However, the control effort of this point is functions.
much higher compared to others. On the other hand, the Furthermore, we apply MCS with different sizes of
suggested controller based on control effort preference samples including 1000, 2000 and 3000 samples for both

123
 R. Mallipeddi et al.

Table 3 Comparison of the proposed robust controllers with the state of the art on first-order system
MOGP MOGP Toscano (2005) Jamali (2009) Jamali et al. (2013) Hajiloo et al. (2008)
(trade-off) (control effort)

Structure Equation (16) Equation (15) K ðsÞ ¼ 0:646 þ 0:5712

s
K ðsÞ ¼ 0:34 þ 0:37
s K ðsÞ ¼ 0:525 þ 0:414
s K ðsÞ ¼ 0:301 þ 0:318
s
Mean ITAE 1.84 2.17 2 2.7 2.25 3.16
SD ITAE 0.004 0.057 0.02 0.07 0.06 0.09
Control effort 1.34 0.98 1.19 0.75 0.97 0.68

the proposed approach and the GA-based robust controller Ket0 s

design method by Jamali et al. (2013) for the first-order P2 ðsÞ ¼ ð16Þ
s2 þ a1 s þ a0
system. Table 4 lists the run time of MOGP compared with
GA. As it is expected, the run time of GP is a bit more than In this problem, Gaussian distribution around 5% of
of GA because of its symbolic nature. the nominal values of the plant parameters,
However, considering the unique advantage of the pro- K ¼ a1 ¼ a0 ¼ t0 ¼ 1, is considered for the simulation.
posed method in structural optimization of robust con- Simulations are performed for 35 s, and we use the same
trollers it can be considered as a computationally efficient parameters used in the first case study.
method in the open literature. The three objective functions, which provide the per-
In order to evaluate disturbance attenuation of the pro- formance measurements of each candidate controller in the
posed controllers, Fig. 8 shows the performance of all uncertain and probabilistic space, are also the same as the
controllers in Fig. 7 to a load disturbance. It is obvious that first benchmark.
the two controllers of this work and the one suggested by We obtained 200 non-dominated solutions after 250
Toscano (2005) damped the effect of disturbance almost in iterations. Figure 10 depicts the 1-norm and 1-norm level
the same time span. diagrams of various objective functions. From the 1-norm
Finally, we plot CDFs to show the robustness of the plots, we can observe lower values of both 1-norm and 1-
suggested trade-off controllers for various uncertainty norm are located near the extremes of mean and SD ITAE
levels. Figure 9 shows the CDF plot of mean ITAE for 1%, objective functions. Similar to the previous case study, the
3%, 5%, 7% and 10% uncertainties for the robust optimum trade-off design point depicted from the Pareto set and
controller. In the derivation of this figure, 5000 Monte shows by point C in Fig. 10. The obtained trade-off solu-
Carlo samples with Gaussian distribution are used. This tion has mean ITAE of 3.04, SD ITAE of 0.058 and control
figure suggests that the designated robust controller has effort of 0.633. This point has the lowest 1-norm value in
acceptable performance for the assumed range of uncer- this figure. The transfer function corresponds to the sug-
tainty levels. gested compromise solution given in Eq. (17):
2:629e2s4 þ 4:564e3s3 þ 5:346e3s2 þ 6:056e3s þ 5:805e2
K ðsÞ ¼
4.2 Optimum robust controller design 2:049s5 þ 1:036e2s4 þ 2:12e3s3 þ 1:599e4s2 þ 1:641e3s
for second-order system with time delay ð17Þ
The designated controller has zeros in [- 16.19,
Another common industrial process is characterized by
- 0.532, ± 1.009, - 0.1047] and poles in [0, - 17.0973,
non-aperiodic response. We can model a second-order plus
± 13.4459, - 16.2652, - 0.1040]. From the set of zeros
dead-time model to represent this category of processes as
and poles, it can be seen that the zero - 0.1047 and pole of
follows (Toscano 2005):

Table 4 Run time of MOGP

Number of samples MOGP (this work) MOGA Jamali et al. (2013)
algorithm with GAs for three
different MCS for the first- and First order 1000 72.56 64.5
second-order system with delay
time 2000 139.96 123.42
3000 210.67 185.74
Second order 1000 88.21 77.28
5000 383.49 321.2
10,000 743.98 668.9

123
Robust controller design for systems with probabilistic uncertain parameters

Fig. 8 The mean of

probabilistic load disturbance
responses of different robust
optimum design

Fig. 9 CDF plot of mean ITAE

for different uncertainty levels
for the robust trade-off
controller (Eq. 15) for the first-
order system

- 0.1040 can be easily eliminated to reduce the controller Figure 11 shows the upper and lower bounds of step
to fourth order as follows: responses correspond to optimum trade-off solution
2:629e2s3 þ 4:537e3s2 þ 4:772e3s þ 5:558e3 (Eq. 18) of this work as well as other researches in the
K ðsÞ ¼
2:049s4 þ 1:034e2s3 þ 2:109e3s2 þ 1:577e4s literature. In addition, Fig. 12 depicts the mean, upper and
ð18Þ lower bounds of control action for the compromise design.
The details of objective functions and controller structures

123
 R. Mallipeddi et al.

Fig. 10 Pareto fronts of multi-objective optimization for the second-order system

of all references for the second-order system are also tab- Jamali et al. (2013) settle the effect of disturbance more
ulated in Table 5. We can observe from Table 5 and quickly compared to other controllers. However, the
Fig. 12 that the obtained robust controller has better MOGP controller has relatively larger overshoot with
functionality in terms of some performance characteristics respect to the others.
such as mean and SD of ITAE and settling time. Again, Furthermore, a sinusoidal noise after the plant with
there are various choices for the designer to select among amplitude of unity and frequency of 1000 Hz is applied to
them for design purposes. The structure of the controller investigate noise rejection performance of the proposed
can be also seen as a design preference. Hence, a simpler controller as well as ones suggested by other methods. In
structure can be depicted from the non-dominated solu- this way, Fig. 14 illustrates step responses of all controllers
tions. It is also possible to restrict the order of transfer in the presence of noise. By comparison of Fig. 14a to b–d,
functions to a maximum specified value during the evolu- the superiority of the suggested robust controller in terms
tionary process. The run time comparison for the second- of noise rejection is apparent compared to other controllers.
order system obtained based on three different MCS of As the previous case, Fig. 15 shows the CDF plot of
1000, 5000 and 10,000 samples for both the proposed mean ITAE for various uncertainties (1–10%) with Gaus-
approach and the GA-based robust controller design sian distribution for the robust optimum controller C. This
method by Jamali et al. (2013) is also given in Table 4. figures suggests that the proposed controller based on the
Load disturbance rejection capabilities of all controllers stochastic approach of this work has acceptable robustness
are also shown in Fig. 13. It can be observed that the performance for around 5% uncertainties which is a rea-
obtained controller by this work and that suggested by sonable value in practice.

123
Robust controller design for systems with probabilistic uncertain parameters

Fig. 11 Upper and lower

bounds of time responses of the
controlled system output for
various robust controllers for
the second-order system

Fig. 12 Mean, upper and lower

bounds of control action of the
robust trade-off controller
obtained by MOGP for the
second-order system

Table 5 Comparison of the proposed robust controller by this work and by other references in the literature for the second-order system
MOGP Jamali (2009) Nariman-zadeh et al. (2007) Jamali et al. (2013)

Structure Equation (18) K ðsÞ ¼ 0:275 þ 0:408

s þ 0:939s K ðsÞ ¼ 0:138 þ 0:331
s þ 0:845s K ðsÞ ¼ 0:709 þ 0:412
s þ 0:32s
Mean ITAE 3.04 8.69 3.96 3.23
SD ITAE 0.058 1.4 0.06 0.06

123
 R. Mallipeddi et al.

Fig. 13 Load disturbance

rejection performance of robust
controllers suggested by
different references

Fig. 14 Noise rejection performance of different controllers for the second-order system

123
Robust controller design for systems with probabilistic uncertain parameters

Fig. 15 CDF plot of mean

ITAE for different uncertainty
levels for the robust trade-off
controller (Eq. 18) for the
second-order system

5 Conclusion Compliance with ethical standards

In this paper, a multi-objective approach based on genetic Conflict of interest Rammohan Mallipeddi has received research
grants from the Basic Science Research Program through the National
programming and Monte Carlo simulations deployed for Research Foundation of Korea (NRF) funded by the Ministry of
automatic synthesis of optimal robust controllers. In the Education, Science and Technology under the Grant
proposed methodology, the structure as well as parameters NRF2015R1C1A1A01055669. Iman Gholaminezhad declares that he
of the candidate controllers is simultaneously optimized in has no conflict of interest. Mohammad S. Saeedi declares that he has
no conflict of interest. Hirad Assimi declares that he has no conflict of
a transfer function representation. The advantage of the interest. Ali Jamali declares that he has no conflict of interest.
proposed multi-objective technique is that the optimum and
robust structure of each controller automatically evolved Ethical approval This article does not contain any studies with human
during the optimization process. Hence, the difficulty of participants or animals performed by any of the authors.
specifying controller structures as a predefined step in the
design procedure is overcome by this methodology. The
References
method is applied to a first-order and a second-order sys-
tem with dead time in a three-objective optimization pro- Assimi H, Jamali A, Nariman-zadeh N (2017) Sizing and topology
cess. The objective functions are the mean and standard optimization of truss structures using genetic programming.
deviation of integral time absolute error and control effort. Swarm Evolution Comput 37:90–103
Awad NH, Ali MZ, Mallipeddi R, Suganthan PN (2019) An efficient
The comparison of the designed controllers with that sug-
differential evolution algorithm for stochastic OPF based active–
gested by other stochastic numerical methods showed reactive power dispatch problem considering renewable gener-
based on step responses, disturbance attenuation and noise ators. Appl Soft Comput 76:445–458
rejection. And the efficiency of the proposed methodology Balandina GI (2017) Control system synthesis by means of cartesian
genetic programming. Proc Computer Sci 103:176–182
exhibited with respect to other stochastic optimization-
Bi S, Deng Z, Chen Z (2013) Stochastic validation of structural FE-
based methods in the literature. The robustness of the models based on hierarchical cluster analysis and advanced
suggested optimal controllers is also presented by CDF Monte Carlo simulation. Finite Elem Anal Des 67:22–33
plots of objective functions for various uncertainty levels. Biswas PP, Suganthan PN, Mallipeddi R, Amaratunga GAJ (2019)
Optimal reactive power dispatch with uncertainties in load
demand and renewable energy sources adopting scenario-based
Acknowledgement This research was supported by the Basic Science
approach. Appl Soft Comput 75:616–632
Research Program through the National Research Foundation of
Biwer A, Griffith S, Cooney C (2005) Uncertainty analysis of
Korea (NRF) funded by the Ministry of Education, Science and
Technology under the Grant NRF2015R1C1A1A01055669. penicillin V production using Monte Carlo simulation. Biotech-
nol Bioeng 90(2):167–179

123
 R. Mallipeddi et al.

Blasco X, Herrero JM, Sanchis J, Martı́nez M (2008) A new graphical manufacturing processes having probabilistic uncertainty in
visualization of n-dimensional Pareto front for decision-making experimental data. J Intell Manuf 28(1):149–163
in multiobjective optimization. Inf Sci 178(20):3908–3924 Jamali A, Khaleghi E, Gholaminezhad I, Nariman-zadeh N (2014b)
Bodla KK, Murthy JY, Garimella SV (2013) Optimization under Modelling and prediction of complex non-linear processes by
uncertainty applied to heat sink design. ASME J Heat Transf, vol using Pareto multi-objective genetic programming. Int J Syst Sci
135 47(7):1675–1688
Chiou S-W (2018) A data-driven bi-level program for knowledge- Kadlic B, Sekaj I, Pernecký D (2014) Design of continuous-time
based signal control system under uncertainty. Knowl-Based controllers using cartesian genetic programming. IFAC Proc Vol
Syst 160:210–227 47(3):6982–6987
Darbra RM, Eljarrat E, Barceló D (2008) How to measure Kalat AA (2019) A robust direct adaptive fuzzy control for a class of
uncertainties in environmental risk assessment. TrAC Trends uncertain nonlinear MIMO systems. Soft Comput
Anal Chem 27(4):377–385 23(19):9747–9759
Diwekar UM, Kalagnanam JR (1997) Efficient sampling technique Kalos MH, Whitlock PA (1998) Monte Carlo methods. Wiley, New
for optimization under uncertainty. AIChE J 43(2):440–447 York
Eldred M (2009) Recent advances in non-intrusive polynomial chaos Kang Z (2005) Robust design optimization of structures under
and stochastic collocation methods for uncertainty analysis and uncertainties. Institut fur Statik und Dynamik der Luft- und
design Raumfahrkonstruktionen, Universiẗat Stuttgart
Field RV, Voulgaris PG, Bergman LA (1996) Methods to compute Koza JR, Keane MA, Streeter MJ, Mydlowec W, Yu J, Lanaz G
probabilistic measures of robustness for structural systems. (2003) Genetic Programming IV: routine human-competitive
J Vibration Control 2(4):447–463 machine intelligence. Kluwer Academic Publishers, Berlin
Fukunaga A, Hiruma H, Komiya K, Iba H (2012) Evolving Krishnan K, Karpagam G (2013) Comparison of PID controller tuning
controllers for high-level applications on a service robot: a case techniques for a FOPDT system. Int J Current Eng Technol
study with exhibition visitor flow control. Genet Program 4:2667–2670
Evolvable Mach 13:239–263 Kumaresan N, Ratnavelu K (2014) Optimal control for stochastic
Gholaminezhad I, Jamali A (2016) A multi-objective differential linear quadratic singular neuro Takagi-Sugeno fuzzy system with
evolution approach based on e-elimination uniform-diversity for singular cost using genetic programming. Appl Soft Comput
mechanisms design. Struct Multidisciplinary Optim 24:1136–1144
52(5):861–877 Li HS, Ma C (2013) Hybrid dimension-reduction method for robust
Gholaminezhad I, Jamali A, Assimi H (2014) Automated synthesis of design optimization. AIAA J 51:138–144
optimal controller using multi-objective genetic programming Mallipeddi R, Suganthan PN, Pan QK, Tasgetiren MF (2011)
for two-mass-spring system. In: presented at the 2nd RSI/ISM Differential evolution algorithm with ensemble of parameters
international conference on robotics and mechatronics, ICRoM and mutation strategies. Appl Soft Comput 11(2):1679–1696
2014 Mohammadzadeh A, Taghavifar H (2020) A robust fuzzy control
Gomes FM, Pereira FM, Silva AF, Silva MB (2019) Multiple approach for path-following control of autonomous vehicles.
response optimization: analysis of genetic programming for Soft Comput 24(5):3223–3235
symbolic regression and assessment of desirability functions. Nariman-Zadeh N, Jamali A, Hajiloo A (2007) Frequency-based
Knowl-Based Syst 179:21–33 reliability Pareto optimum design of proportional-integral-
Hajiloo A, Nariman-Zadeh N, Jamali A, Bagheri A, Alasti A (2008) derivative controllers for systems with probabilistic uncertainty.
Pareto optimum design of robust PI controllers for systems with Proc Inst Mech Eng Part I J Syst Control Eng 221:1061–1075
parametric uncertainty. Int Rev Mech Eng 1(6):628–640 Nariman-Zadeh N, Salehpour M, Jamali A, Haghgoo E (2010) Pareto
Hu N, Zhong J, Zhou JT, Zhou S, Cai W, Monterola C (2018) Guide optimization of a five-degree of freedom vehicle vibration model
them through: an automatic crowd control framework using using a multi-objective uniform-diversity genetic algorithm
multi-objective genetic programming. Appl Soft Comput (MUGA). Eng Appl Artif Intell 23(4):543–551
66:90–103 Nejlaoui M, Houidi A, Affi Z, Romdhane L (2013) Multiobjective
Jamali A (2009) Pareto Robust design of controllers with probabilistic robust design optimization of rail vehicle moving in short radius
uncertainties using multi objective evolutionary algorithms. curved tracks based on the safety and comfort criteria. Simul
Ph.D. Thesis, University of Guilan Model Pract Theory 30:21–34
Jamali A, Nariman-Zadeh N, Atashkari K (2008) Multi-objective Pettersson MP, Iaccarino G, Nordstrom J (2015) Polynomial chaos
uniform diversity genetic algorithm (MUGA). In: Kosinski W methods for hyperbolic partial differential equations. In: Math-
(ed) In advanced in evolutionary algorithms. IN-TECH, Vienna ematical engineering, Springer
Jamali A, Hajiloo A, Nariman-zadeh N (2010) Reliability-based Pettersson MP, Iaccarino G, Nordstrom J (2015b) Polynomial chaos
robust Pareto design of linear state feedback controllers using a methods for hyperbolic partial differential equations. Springer,
multi-objective uniform-diversity genetic algorithm (MUGA). Berlin
Expert Syst Appl 37(1):401–413 Sekaj I, Perkacz J (2007) Genetic programming—based controller
Jamali A, Ghamati M, Ahmadi B, Nariman-zadeh N (2013a) design. In: 2007 IEEE congress on evolutionary computation,
Probability of failure for uncertain control systems using neural pp 1339–1343
networks and multi-objective uniform-diversity genetic algo- Smith BA, Kenny SP, Crespo LG (2005) Probabilistic parameter
rithms (MUGA). Eng Appl Artif Intell 26(2):714–723 uncertainty analysis of single input single output control
Jamali A, Salehpour M, Nariman-zadeh N (2013b) Robust Pareto systems. NASA
active suspension design for vehicle vibration model with Toscano R (2005) A simple robust PI/PID controller design via
probabilistic uncertain parameters. Multi-body Syst Dyn numerical optimization approach. J Process Control 15(1):81–88
30:265–285 Uyeh DD et al (2018) Interactive livestock feed ration optimization
Jamali A, Khaleghi E, Gholaminezhad I, Nariman-zadeh N, Gho- using evolutionary algorithms. Comput Electron Agric 155:1–11
laminia B, Jamal-Omidi A (2014a) Multi-objective genetic Wang Q, Stengel RF (2002) Robust control of nonlinear systems with
programming approach for robust modeling of complex parametric uncertainty. Automatica 38(9):1591–1599

123
Robust controller design for systems with probabilistic uncertain parameters

Witteveen J, Iaccarino G (2012) Simplex stochastic collocation with Zio E, Bazzo R (2011) Level diagrams analysis of Pareto Front for
random sampling and extrapolation for nonhypercube probabil- multiobjective system redundancy allocation. Reliab Eng Syst
ity spaces. SIAM J Sci Comput 34(2):A814–A838 Saf 96(5):569–580
Zhao Q, Chen X, Ma ZD, Lin Y (2015) Robust topology optimization
based on stochastic collocation methods under loading uncer- Publisher’s Note Springer Nature remains neutral with regard to
tainties. Math Probl Eng jurisdictional claims in published maps and institutional affiliations.

123