Advances in Engineering Research, volume 127

3rd International Conference on Electrical, Automation and Mechanical Engineering (EAME 2018)

Improved NSGA-II Algorithm for Optimization of

Constrained Functions
Yun Zhang* and Bin Jiao
Shanghai Dianji University Shanghai, China
*
Corresponding author

Abstract—In order to solve the constrained multi-objective Wang [6] proposed a partial order relation in the improved
optimization problem, an improved NSGA-II algorithm is algorithm and used Cauchy distribution to do crossover
proposed. On the basis of NSGA, the cross operation of the operations. Then, according to the partial order relation, the
feasible and unfeasible solution is implemented in order to give individuals are sorted and generated.. W. Zhang [7] proposed
full play to the role of the infeasible solution in the optimization an improved NSGA-II algorithm based on differential
process. In addition, the external preservation set is updated on evolution mechanism. Z. Han [8] proposed a multi-objective
the basis of the obtained dominant individual to preserve the dynamic adaptive differential evolution algorithm (MODADE)
optimal solution of the problem. The improved algorithm is to improve the search efficiency of general differential
applied to typical test functions and compared with NSGA-II.
evolution operations. I. O. Essiet [9] proposed that the
The experimental results show that the algorithm is superior.
combination of reference points and adaptive crossover rate can
Keywords—multi-objective optimization; improved non- improve the performance of NSGA- II.
dominated sorting genetic algorithm; infeasible solutions; external
II. NSGA-II ALGORITHM
save set
NSGA-II is proposed by Deb et al [10]. This algorithm is
I. INTRODUCTION one of the most excellent multi-objective optimization
In production management, job scheduling, metallurgical evolutionary algorithms, and has been widely applied [11-13].
process, and other engineering fields, there are a lot of The core of NSGA-II evolutionary algorithm lies in two
optimization problems. Among them, most of the optimization aspects: the fast non-dominated sorting and elitist selection
problems have a variety of constraints. Evolutionary strategy for individuals. The fast non-dominated sorting is
optimization method has been widely used to solve a variety of based on non dominated rank and crowding distance index.
optimization problems, but the use of this method to solve The Pareto control relation between the optimal energy
constrained optimization problems, it is necessary to melt the functions determines the rank of the non-dominated function.
necessary constraint processing technology [1]. In recent years, Take a multi-objective optimization problem with two
human penalty function technology is used in evolutionary performance indicators as an example. Pareto domination is
optimization methods to solve constrained optimization defined as follows: for the solution vector x1 and x 2 in the
problems. However, this method has a reasonable penalty
feasible field, x1 is called Pareto domination or x1 dominates
factor setting problem: the penalty factor is too large, although
it can make the evolutionary population converge quickly to x 2 (donate x1  x 2 ), when the following condition holds:
the feasible domain of the problem, but neglects the value of
the population to be of no value. By using the feasible solution,
it is difficult to find the optimal solution of the problem; on the
f i ( x1 )  f i ( x2 ) , i  (1,2) 
contrary, if the penalty factor is set too small, then the f j ( x1 )  f j ( x2 ) j  (1,2)
evolutionary population will stay in the unfeasible region [2][3], , (1)
and the other way to deal with the constraints is that it is
difficult to get the feasible solution of the problem. The The crowd distance index of each individual is a
problem of optimization is transformed into a multi-objective representation of the distribution uniformity of the current
optimization problem without constraints, which is solved by population.
using the existing unconstrained optimization methods.
However, most methods require the infeasible solution to be The evolution of the NSGA-II algorithm is shown in
inferior to the feasible solution when comparing the Figure I.
performance of the individual, so it is difficult to fill in the Individual choice can be described as follows: first, take
valuable information provided by the infeasible solution at the the second generation evolution as an example, the whole
early stage of evolution. In addition, the effectiveness of the individual comes from two parts: Parent individuals (Pt) and
method also depends on the specific multi-objective Offspring individuals (Qt) from the last generation; secondly,
optimization method [4][5]. calculate the non-dominated rank F and crowding distance
In order to solve constrained multi-objective optimization index for each individual in the new population of (Pt, Qt).
problems, many scholars propose improvement methods. M.

Copyright © 2018, the Authors. Published by Atlantis Press.

This is an open access article under the CC BY-NC license (http://creativecommons.org/licenses/by-nc/4.0/). 316
 Advances in Engineering Research, volume 127

IV. MULTI-OBJECTIVE CONSTRAINED OPTIMIZATION

Based on the traditional NSGA algorithm, this paper
proposes an improved algorithm. In the processing of the
constrained optimization problem of NSGA algorithm, the
constraint conditions are converted to the target of optimization.
In the search process, the comparison rule of Deb is used to
compare the feasible solution with the infeasible solution. In
practical applications, the real optimal solution of most
constrained optimization problems is often located in the
vicinity of the constrained boundary, and the probability of the
objective function of the unfeasible solution located in the
constrained boundary is superior to that of the feasible solution
of the target function. Therefore, the utilization of these highly
infeasible solutions can enhance the search speed closer to the
feasible domain.
FIGURE I. EVOLUTION PROCESS OF NSGA-II ALGORITHM [10]
Considering the difference between feasible solution and
According to the level of each individual (small rank is infeasible solution, it is proposed that genetic operators can be
priority), excellent individuals are selected to the elite group. selected from every evolutionary number of generations with
For individuals with the same rank, individuals with crowding better feasible solution set and infeasible solution set. The
distance index and large crowding distance have priority. Then, concept of evolutionary optimization is proposed: for a two
the TTH generation evolved until the population size reached N, objective optimization problem, assuming that the parent
forming a fresh parent population Pt 1 individual P1 , P2 , sub-generation individual C1 , C 2 , parent
individual and progeny corresponding target function values
III. CONSTRAINT OPTIMIZATION PROBLEM AND ITS 1 2 1 2
are f p , f p , f c , f c respectively. Two of the nine possible
RELATED DESCRIPTION
relationships for individuals to compare, the first four cases
Generally speaking, a constrained optimization problem can
belong to the evolutionary preferred range.
be described as follows:

min f ( x), x  D  f c1  f p1  f c1  f p1  f c1  f p1  f c1  f p1
  2 2  2 2  2 
 g i ( x)  0, i  1,2,  , q  f c  f p  f c  f p  f c  f p2  f c2  f p2
, , ,
h( x)  0, j  q  1,  , m
 (2)  f c1  f p1  f c1  f p1  f c1  f p1  f c1  f p1
 2 2  2 2  2 
In the formula, f (x ) is the objective function, and x is a  f c  f p  f c  f p  f c  f p2  f c2  f p2
, , ,
decision variable. g i ( x)  0, i  1,2,  , q is the inequality  f c1  f p1
constraint condition of the problem,  2
  f c  f p
2
D  x  R n li  xi  ui , li , ui  R, i  1,2,  , n is a (4)
search space, all the individual solutions satisfying the
constraint condition in the search space D form feasible Genetic manipulation is effective to compare offspring and
solutions. h ( x )  0, j  q  1,  , m is an equation constraint parental individuals.

condition. h j ( x )    0 ,  is an equality constraint that c1  p1 , c1  p2 , c2  p1 , c2  p2 (5)

violates function thresholds and is generally set to a smaller
positive number. In the evolutionary process. Define It is desirable to have at least one generation of individuals.
This operation can ignore the poorly performing evolutionary
max0, g i ( x ),1  i  q process, exclude individuals with poor characters and ensure
Gi ( x)  
max0, hi ( x )   , q  1  i  m
that the search is faster and more efficient to get closer to the
optimal solution area, and the optimization performance of the
(3)
algorithm is improved. After generation, the unfeasible solution
and feasible solution are executed and reorganized, and an
It represents the violation degree of a solution x to the adaptive adjustment mechanism is set up for this link.
i constraint in a group, and reflects the distance order of  Considering that the search is close to the feasible domain and
from the feasible domain. the optimal solution direction in the global optimization
process, the number of feasible solutions is gradually increased
as far as the whole process of evolution is concerned. If the

317
 Advances in Engineering Research, volume 127

genetic operation that executes the feasible solution and the 3) Non dominated sorting, crowding distance calculation
infeasible solution too much in the later stage of evolution, it and feasible selection operators are used to determine the first
may counteract the effect of the algorithm in the feasible parent.
domain. Therefore, we should gradually reduce the frequency 4) Perform advantage set archiving and replace parental
of genetic operations between infeasible solutions and feasible
individuals.
solutions in search process. In this paper, the evolutionary
algebra of the feasible solution and the intersecting link of the 5) Performing genetic manipulation
infeasible solution are adaptively adjusted in the process of 6) To perform intermediate selection operations for
executing the genetic operation between the feasible solution parent-child mixed sets to determine the next parent
and the infeasible solution, that is, when the evolutionary generation and determine whether evolution reaches the
algebra is  , the cross variation of the infeasible solution and maximum evolutionary algebra, if it is the end of the algorithm;
the feasible solution is performed. otherwise, jump to step 7)
7) To determine whether evolution can reach the
 K  roud (k ) conditions to carry out genetic operations between infeasible

 
solutions and feasible solutions, if so, turn step 8); otherwise,
k  k  T k  e , t  [0,1,0.1,0.2,0.3,  ]
t
turn to step 4).
(6) 8) Performing genetic operations on infeasible solutions
and feasible solutions.
In the formula, K is the evolutionary number of the feasible 9) The next generation of feasible solutions update the
and infeasible solutions to perform crossover and mutation
next parent, and the infeasible solution is retained as the next
operations. T is the total population evolution algebra.
parent, then, turn to step 3).
In order to effectively retain the dominant individual in the
process of population evolution to better play its guiding role in V. EXPERIMENTS
the direction of evolution, this paper preserves the first frontier This article tests the typical test functions (ZDT1-2) listed
feasible solution archive as an advantage set in the optimization in [10]. Using NSGA-II algorithm and INSGA-II algorithm as
process, and ensures that the replacement link is effective in the comparison algorithms. Converting equality constraint
process of replacing the dominant set .Compared with
h( x)  0 in optimization function to h( x)    0
individuals a and elite individuals b , b has better
performance than a and individual substitution. In view of the inequality constraint,   0.0001 , Population size is 500,
serious overlap of dominant individuals in the feasible domain Evolutionary algebra is 200, The crossover probability is 0.9,
of search, this paper proposes half performing marginal The probability of mutation is 0.1. Experimental results shown
mutation operation on the dominant individual size. In the in Figure II, Figure III.
process of alienation from x to x , if the value range of point
C is [U min , U max ] , Then xk can be obtained by the following:

U min
k
, rand (0,1)  0

xk   k
U max , rand (0,1)  1
(7)

rand (0,1) indicates that 0 or 1 are selected with the same

probability. If the range of decision variables is large and
unconstrained, marginal variation will probably not improve
the effectiveness of the algorithm. For the case where the
optimal solution is at the boundary of the feasible domain, the
marginal mutation operation focuses on the global. The key
FIGURE II. COMPARISON OF EXPERIMENTAL RESULTS OF TWO
point is that the individual solution is still located in the ALGORITHMS ON ZDT1
feasible domain of the target space after the mutation, and the
global searching ability of the algorithm is improved and the
performance of the algorithm is improved.
The process steps for improving the algorithm NSGA-II are
as follows:
1) Set algorithm parameters and initialize population.
2) To distinguish the population set of feasible solutions
and infeasible solution set.

318
 Advances in Engineering Research, volume 127

[8] Han Z, Wang S, Dong X, et al. Improved NSGA-II algorithm for multi-
objective scheduling problem in hybrid flow shop[C]// International
Conference on Modelling, Identification and Control. IEEE, 2018.5.10.
[9] I. O. Essiet, Y. Sun and Z. Wang, "Enhanced NSGA based on adaptive
crossover rate and reference points," 2017 2nd International Conference
on Robotics and Automation Engineering (ICRAE), Shanghai, 2017, pp.
295-300.
[10] Deb K, Pratap A, Agarwal S, et al. A fast and elitist multiobjective
genetic algorithm: NSGA-II[J]. IEEE Transactions on Evolutionary
Computation, 2002, 6(2):182-197.
[11] Zeinalzadeh A, Mohammadi Y, Moradi M H. Optimal multi objective
placement and sizing of multiple DGs and shunt capacitor banks
simultaneously considering load uncertainty via MOPSO approach[J].
International Journal of Electrical Power & Energy Systems, 2015,
67:336-349.
[12] Li Y, Lu X, Kar N C. Rule-Based Control Strategy With Novel
FIGURE III. COMPARISON OF EXPERIMENTAL RESULTS OF TWO Parameters Optimization Using NSGA-II for Power-Split PHEV
ALGORITHMS ON ZDT2 Operation Cost Minimization[J]. IEEE Transactions on Vehicular
Technology, 2014, 63(7):3051-3061.
[13] Zamanifar M, Fani B, Golshan M E H, et al. Dynamic modeling and
From the experimental results, it can be seen that both optimal control of DFIG wind energy systems using DFT and NSGA-
algorithms can obtain a Pareto frontier with better convergence, II[J]. Electric Power Systems Research, 2014, 108(3):50-58.
and the Pareto frontier is very similar to the real frontier,
however, the Pareto frontier obtained by the improved
algorithm is uniformly distributed, the Pareto frontier obtained
by the NSGA-II algorithm has local leakage, and the
uniformity is slightly weaker. On the other hand, the number of
solutions in the Pareto frontier obtained by the improved
algorithm is more than the number of Pareto frontier solutions
obtained by the original algorithm.
VI. CONCLUSION
In this paper, based on the classical NSGA-II algorithm, an
improved algorithm NSGA-II algorithm for solving
constrained multi-objective optimization problems is proposed.
Based on the NSGA-II algorithm, this algorithm performs
genetic operations on infeasible solutions and feasible solutions
every certain evolutionary algebra and adaptively adjusts the
frequency of this genetic operation. Set restrictions to ensure
the validity of genetic operations. Solving the problem of
overlap of late dominant individuals. Experimental results
show that the algorithm has superiority.
REFERENCES
[1] B. Tessema and G. G. Yen, "An Adaptive Penalty Formulation for
Constrained Evolutionary Optimization," in IEEE Transactions on
Systems, Man, and Cybernetics - Part A: Systems and Humans, vol. 39,
no. 3, pp. 565-578, May 2009.
[2] Correia A, Mestre P, Serodio C, et al. Adaptive Penalty and Barrier
function based on Fuzzy Logic[J]. Expert Systems with Applications,
2015, 42(19):6777-6783.
[3] Lemonge A C C, Barbosa H J C, Bernardino H S. Variants of an
adaptive penalty scheme for steady-state genetic algorithms in
engineering optimization[J]. Engineering Computations, 2015,
32(8):2182-2215.
[4] Deb K. An efficient constraint handling method for genetic algorithms[J].
Computer Methods in Applied Mechanics & Engineering, 2000,
186(2):311-338.
[5] Sarker R A, Elsayed S M, Ray T. Differential Evolution With Dynamic
Parameters Selection for Optimization Problems[J]. IEEE Transactions
on Evolutionary Computation, 2014, 18(5):689-707.
[6] Wang M, Dai G, Hu H. Improved NSGA-II Algorithm for Optimization
of Constrained Functions[C]// International Conference on Machine
Vision and Human-Machine Interface. IEEE, 2010:673-675.
[7] Zhang W, Zhang J L. Improved NSGA-II algorithm based on differential
evolution mechanism[C]// Chinese Control Conference. 2017:4334-4338.

319