Swarm and Evolutionary Computation 55 (2020) 100670

Contents lists available at ScienceDirect

Swarm and Evolutionary Computation

journal homepage: www.elsevier.com/locate/swevo

A multi-objective differential evolutionary algorithm with angle-based

objective space division and parameter adaption for solving sodium
gluconate production process and benchmark problems
Zhan Guo a, b, Okan K. Ersoy b, Xuefeng Yan a, c, *
a
Key Laboratory of Advanced Control and Optimization for Chemical Processes of Ministry of Education, East China University of Science and Technology, Shanghai,
200237, PR China
b
School of Electrical and Computer Engineering, Purdue University, West Lafayette, IN, 47907, USA
c
Shanghai Institute of Intelligent Science and Technology, Tongji University, Shanghai, 200237, PR China

A R T I C L E I N F O A B S T R A C T

Keywords: Convergence and diversity are two main performance indicators in multi-objective evolutionary algorithms. The
Differential evolutionary algorithm fitness value in the objective space represents information which guides the evolution. To extract this useful
Angle-based information, a multi-objective differential evolutionary algorithm with angle-based objective space division and
Adaption
parameter adaption is proposed (MODE-ASP). In MODE-ASP, the objective space is split into several subspaces
Gaussian model
Multi-objective problem
based on angle, and the optimal direction in each subspace is extracted to accelerate the convergence. A prob-
Pareto front ability model is also built to achieve adaption of the parameters along with the evolution of the population.
Sodium gluconate production Compared with 5 state-of-the-art algorithms with 20 benchmark functions, MODE-ASP is shown to give a better
performance. Moreover, the operating conditions of the sodium gluconate fermentation process are optimized
with three proposed objective functions, to improve the utilization rates of equipment and conversion rates
effectively. The MODE-ASP is shown to obtain a better Pareto front in this application.

1. Introduction In the current study, new evolutionary strategies and new approaches
for differential evolutionary (DE) algorithm [7] are proposed to improve
Multi-objective optimization problems (MOPs) are complex. They the convergence and diversity of the algorithm. The result are discussed
usually contain two or more highly nonlinear functions, which often in comparison to other methods, such as the indicator based
conflict with each other. The mathematical formulation is as follows [1]: multi-objective EAs [8–10]; new dominate relationship based
multi-objective EAs [11,12], evolutionary strategy based multi-objective
min FðxÞ ¼ ff1 ðxÞ; f2 ðxÞ; …; fm ðxÞg EAs [13–15], and some studies are focus on increasing the search ability
gi ðxÞ  0; i ¼ 1; 2; …; q (1)
s:t: of solutions in decision space [16–18]. DE was first proposed by Storn
hj ðxÞ ¼ 0; j ¼ 1; 2; …; p
and Price in 1995. A generalized differential algorithm (GDE3) for global
where FðxÞ is the overall objective function; m indicates the mth objec- optimization was introduced by Kukkonen and Lampinen [19]. An
tive function; and gðxÞ， hðxÞ are constraints. Most of the time, these adaptive DE algorithm (ADEA) used a new self-adaptive mutation
functions cannot be optimized simultaneously. The solutions are not parameter, and comprehensively applied Pareto-based ranking and
single, but an optimal solution set called Pareto set (PS) is sought. crowding distance to make the parameters achieve real-time update [20].
Evolutionary algorithms (EAs) have the characteristics of parallelization MOSADE is a self-adaptive DE algorithm, which used the crowding en-
and robustness that make them successful to solve the MOPs. Some ex- tropy to measure the crowding degree [21]. Venske Sandra M. et al.
amples are non-dominated sorting in genetic algorithms (NSGAs) [2], combined the MOEA/D algorithm with a self-adaptive strategy selection
multiple objective genetic algorithms (MOGAs) [3], strength Pareto (adaptive pursuit and probability matching) and presented the adaptive
evolutionary algorithm (SPEA) [4], multi-objective particle swarm opti- DE algorithm for MOPs (ADEMO/D) [22]. In multi-objective differential
mization algorithm (MOPSO) [5], and multi-objective evolutionary al- evolution algorithm (MODEA) [23], the potential individuals in the
gorithm based on decomposition (MOEA/D) [6]. population were generated by utilizing the superiority of

* Corresponding author.P.O. BOX 293, MeiLong Road NO. 130, Shanghai, 200237, PR China.
E-mail address: [email protected] (X. Yan).

https://doi.org/10.1016/j.swevo.2020.100670
Received 3 April 2019; Received in revised form 12 February 2020; Accepted 24 February 2020
Available online 17 March 2020
2210-6502/© 2020 Elsevier B.V. All rights reserved.
Z. Guo et al. Swarm and Evolutionary Computation 55 (2020) 100670

opposition-based learning strategy and the random localization concept and the optimal results with the sodium gluconate fermentation process
in the mutation strategy. Meanwhile, some new benchmark test suits are are described in Section 5. Section 6 is the conclusions.
designed too. Li H, Deb K, Zhang Q et al. proposed a novel many and
multi-objective benchmark problems which contains complicated Pareto 2. Differential evolution algorithm
set, bias, disconnection or degeneracy [24]. Cheng R, Li M, Tian Y et al.
proposed a many-objective optimization test suite for CEC0 2018 [25]. The differential evolution algorithm is a metaheuristic algorithm. The
etc. evolution process is based on mutation, crossover and selection opera-
In the DE algorithm, the mutation and crossover parameters signifi- tions. The decision vector xi ¼ ½xi;1 ; xi;2 ; :::; xi;D  with D-dimensions can be
cantly affect the balance between global search and local search perfor- regarded as an individual. The population, contains NP individuals, and
mances. An optimal combination of the parameters not only improves the is denoted by X ¼ ½x1 ; x2 ; :::; xNP . The evolution process is as follows:
convergence of the algorithm, but also improves the diversity of the
population. The objective space contains the optimal information for the (1) Initialization: set the size and the members of the initial popula-
optimal direction and the optimal combination of the parameters. The tion, the maximum number of generations, the mutation param-
optimal direction guides the whole population towards the optimal re- eter, and the crossover parameter.
gion. Finding a proper way to extract the optimal direction is very (2) Mutation operation: the mutant individual is generated by one of
meaningful. In this article, an angle-based objective space partition and six strategies. The difference between them is the way of indi-
parameter adaption multi-objective DE algorithm (MODE-ASP) is pre- vidual selection (current individual, best individual and random
sented for this purpose. The objective space is split into several subspaces individual). The six strategies are given as follows: [32,33]
based on the angle between the reference points and the fitness values.
This is a preparation for selecting the superior individuals which contain }DE = rand = 1}½32 : vi ¼ xr1 þ F  ðxr2  xr3 Þ (2)
the optimal direction and the superior information. The differences be-


tween parents and offsprings indicate the quality of the corresponding }DE rand 2}½33 : vi ¼ xr1 þ F  ðxr2  xr3 Þ þ F  ðxr4  xr5 Þ (3)
parameters. A mixture-Gaussian model is built to estimate the optimal
parameters. This operation ensures the parameters adjust with the evo- }DE = best = 1}½33 : vi ¼ xbest þ F  ðxr2  xr3 Þ (4)
lution. Meanwhile, a mutation self-adaptive strategy is developed to
improve the diversity.


}DE best 2}½32 : vi ¼ xbest þ F  ðxr2  xr3 Þ þ F  ðxr4  xr5 Þ (5)
Furthermore, the MODE-ASP is used to solve the sodium gluconate
production process. Three objective functions are proposed to represent
}DE = current  to  best = 1}½32 : vi ¼ xi þ F  ðxbest  xi Þ þ F  ðxr1  xr2 Þ
the conversion rate, the utilization rate of equipment, and the remaining
(6)
D-glucose in the fermentation process of sodium gluconate. Sodium glu-
conate is a sodium salt widely used in medicine, leather, food, detergent,
and beverage industries [26]. It increases the strength and slows the }DE = current  to  best = 2}½33 : vi ¼ xbest þ F  ðxbest  xi Þ þ F  ðxr1  xr2 Þ
coagulation time of cement in the concrete industry [27] as water (7)
reducer and retarder. Moreover, sodium gluconate cleans glass as
where xbest is the superior individual; xr1 ; xr2 ; xr3 ; xr4 ; xr5 are
detergent and helps in water purification industry as stabilizer. The
randomly selected from NP individuals; F is the mutation parameter.
molecular formula of sodium gluconate is C6H11O7Na. It is non-toxic and
non-deliquescent. The main production methods are the immobilized cell
(3) Crossover operation: the individual is obtained probabilistically
and immobilized enzyme method and the biological fermentation pro-
by
cess. In batch fermentation [28] Aspergillus niger is used as fermenting
microbe, and D-glucose as substrate. The glucose oxidation (GOD) 
vi;j ; randð0; 1Þ  CRi
generated by Aspergillus niger results in D-gluconic acid and H2O2 after ui;j ¼ ; j ¼ 1; 2; :::; D (8)
xi;j ; otherwise
dehydrogenation. Adding sodium hydroxide helps maintain the pH and
results in the product sodium gluconate. Researchers have described this where CR is the crossover parameter.
process in many forms such as kinetic models [29], mathematical models
[30], and hybrid neural network models [31]. The output of sodium (4) Selection operation: if the functional value of the offspring is
gluconate is influenced by many factors such as the concentration of better than that of the parent, the offspring replaces the parent,
substrate, the concentration of biomass and the oxygen supply. It is very otherwise, the parent remains.
important to find an effective way to obtain the optimal combination of 
the factors to improve the production efficiency of sodium gluconate. ui ; f ðui Þ  f ðxi Þ
b
xi ¼ ; i ¼ 1; 2; …; NP (9)
Three major contributions are summarized as follows: xi ; f ðui Þ  f ðxi Þ

1) An angle-based objective space division method is designed. The

(5) End: if the evolution process satisfies the termination conditions,
whole space is divided into several sub-spaces, and the optimal di-
the process is terminated; else, step (2)–(4) are repeated.
rection in each subspace is extracted to accelerate the convergence
2) A probability model and a mutation strategy library are built to
3. Multi-objective differential evolutionary algorithm WITH
achieve adaption of the parameters along with the evolution of the
objective space division based on fitness-angle and parameter
population.
adaption
3) The proposed algorithm is test by 20 different type benchmark test
functions and applied to optimize the fermentation industry process.
3.1. The framework of MODE-ASP
Three designed evaluation indices has been used as a result to show
the effectiveness of MODE-ASP.
The proposed MODE-ASP has a similar framework as the traditional
DE algorithm, except for the space partition operation based on fitness-
The paper consists of six sections. Section 2 introduces the DE algo-
angle, parameter adaption strategy and mutation self-adaptive strategy.
rithm. The proposed MODE-ASP is described in Section 3. Section 4
In general, two main solution sets are included in MODE: population X
discusses the experimental results based on two quality indicators and
and archive A Population X contains the solutions that participate in the
the results of statistical analysis. The proposed three objective functions

2
Z. Guo et al. Swarm and Evolutionary Computation 55 (2020) 100670

Fig. 1. Fitness partitioning based on angle.

 
evolution process; archive A is a representative of the Pareto front. In Rpmax ¼ f1min ; …; ftmax ; …; fmmin ; t ¼ 1; 2; …; m and Rpmin
MODE-ASP, A is not used, all the elitist solutions are involved in the  min min 
¼ f1 ; f2 ; …; fmmin
evolution; and each solutions has its own parameter for the DE evolution
process, so a parameter population C is added it has the same size as X. Rpmax includes one function’s maximum value and other function’s
Algorithm 1 to be discussed below shows the framework of MODE-ASP. It minimum value. An example of bi-objective functions is shown in Fig. 1
comprises of the following steps. First, an initial population X and an 0
(a). We calculate the vector angle between f and Rpmax by
initial parameter population C are randomly generated in a limited range
with size N. The fitness values f of each solution in X are calculated. 0
θ ¼ arccosðf ðxÞ; Rpmax Þ (12)
Then, in the main loop, the space division operation based on fitness-
angle divides X into several sub-populations, i.e. S1 ; S2 ; :::; Sn . The 0
0 jf ðxi Þ  Rpmax j
parameter population C is also divided into sub-populations, i.e. C1 ;C2 ;:::; cosðf ðxi Þ; Rpmax Þ ¼ 0 ; i ¼ 1; 2; :::; N; (13)
Cn . A novel parameter adaption strategy is further proposed to improve kf ðxi ÞkkRpmax k
the performance. During the evolution process, not only the elitist solu- The β vector angle between reference points is given by
tions are reserved but also the non-dominated solutions are retained.
 
Non-dominated sorting [34] and crowding distance sorting [35] are used βðt;jÞ ¼ arccos Rpmax;t ; Rpmax;j ; t 2 f1; 2; :::; mg; j 2 f1; 2; :::; mg; t 6¼ j (14)
to maintain the population size.
If the number of objectives functions is more than two, any two
Algorithm 1 objective functions are used. ðt; jÞ denotes the chosen objective func-
Framework of MODE-ASP. tions. The individuals are given a 0–1 label based on θ and β by


1; θi > βðt;jÞ n
Sni ¼
; i ¼ 1; 2; :::; N; (15)
0; θi < βðt;jÞ n

where n is the subspace number. If the population size is small, too many
subspaces may lead to a small number of individuals in each subspace.
This slows down the convergence speed. Three subspaces are used in this
article. Fig. 1 (b) shows the corresponding details of θ and β. The small
triangles represent the reference points.
For the solution of the optimization problem, the individuals (Pbest )
that are closest to the Pareto front in the optimal direction include the
most superior information. After the operations above, the target space is
divided into several subspaces. Using equation (16) to find the center of
each subspace, the direction of the center (T k ; k ¼ 1; 2; :::; n) to the
In the next sections, the four main components of the proposed reference point is given by
MODE-ASP are described in more detail. P 0
f ðxi Þ
Tk ¼ ; i ¼ 1; 2; …; zk (16)
3.2. A space division method based on fitness-angle zk
The distance between all the individuals in the kth subspace and Rpmin
Algorithm 2 to be discussed details the procedure of the space divi- in the optimal direction are next calculated. The individual which has the
sion method based on the fitness-angle in MODE-ASP. To begin with, the minimum distance dmin;k is defined as the superior individual Pbest; k . zk is
fitness values are normalized and put in vector form as the solution number of the kth subspace (see Fig. 2).
fj ðxi Þ  fjmin  0 
di;k ¼ f ðxi;k Þcosðf ðxi;k Þ; T k Þ;
0
f j ðxi Þ ¼ i ¼ 1; 2; 3; :::; N; j ¼ 1; 2; :::; m (10) k ¼ 1; 2; …; n; ; i ¼ 1; 2; …; zk (17)
fjmax  fjmin

0
dmin;k ¼ minðd1;k ; d2;k ; d3;k ; :::; dzk ;k Þ (18)
f ðxi Þ ¼ ½f 1 ðxi Þ; f 2 ðxi Þ; …; f m ðxi ÞT (11)
Algorithm 2
The reference points are the boundary of space partition, generate by
Angle-Partitioning.
the minimum and maximum value of each objective functions’ fitness
value, defined as

3
Z. Guo et al. Swarm and Evolutionary Computation 55 (2020) 100670

Fig. 2. Visualization of angle partitioning.

perior individual are used as follows:

DE=best=2 : vi ¼ xbest;k þ F  ðx1  x2 Þ þ F  ðx3  x4 Þ (19)

DE=current  to  best=1 : vi ¼ xi þ F ðxbest;k  xi Þ þ F  ðx2  x3 Þ (20)

DE=rand  to  best=1 : vi ¼ x1 þ F ðxbest;k  x1 Þ þ F ðx2  x3 Þ (21)

where x1 ; x2 ; x3 ; x4 are randomly selected from S1 ; S2 ; :::; Sn . F is the

mutation parameter.
The crossover operation is given by

vi;j ; randð0; 1Þ  CRi
ui;j ¼ ; j ¼ 1; 2; …; D (22)
xi;j ; otherwise

where, bS k ¼ fu1 ; u2 ; :::; ui g; k ¼ 1; 2; :::; n represents the offspring of the

3.3. Update kth block. The whole offspring population is X b ¼S b 1 [ ::: [ S
b n . The par-
ents X and offspring X b are used by adopting non-dominated sorting and
In the update process, the classical mutation and crossover strategy of crowding distance sorting to select NP individuals to maintain the pop-
the DE algorithm are used. Every subspace is updated with three muta- ulation size.
tion operations. After the mutation and crossover operations, each indi-
vidual in the kth subspace is given a label matSk ; k ¼ f1; 2; :::; ng to
represent the category of the mutation strategy. This operation is a 3.4. A parameter adaption method
preparation for the mutation-adaption method. Algorithm 3 details the
procedure of the update process. In the DE algorithm, the mutation parameter determines the magni-
tude of the disturbance; the crossover parameter determines the proba-
Algorithm 3 bility of preserving the characteristics of the previous generation. Fixed
Update. set of parameters cannot satisfy the requirement of different evolution
phases. A dynamically adjusted parameter set based on the superior in-
dividuals has a great impact on the algorithm’s performance.
Algorithm 4 to be discussed presents the procedure of the proposed
parameter-adaption strategy. In this strategy, every individual has its
own mutation and crossover parameter in each generation. To start with,
the offsprings b S n are generated with fitness values bf 1 ;:::; b
S 1 ; :::; b f n . In the
case of all functions minimized, the sums of the differences between the
parents’ fitness value and the offspring’s fitness value are given by.
X  n o n o
Δfk;i ¼ fk;i  bf k;i ; k 2 1; 2; :::; n ; i 2 1; 2; …; zk (23)

where k is the number of blocks; zk is the number of individuals in the kth

DE/rand-to-best/1 block; and Δfk;i is the difference value. If Δfk;i is a positive real number, it
indicates the objectives are approaching to the optimal region; mean-
while, its corresponding individual’s parameter is beneficial for the
evolution. The positive Δfk;i is stored in V k ; k 2 f1; 2; :::; ng, and its cor-
responding individual’s parameter is stored in Rk ; k 2 f1; 2;:::; ng. Then,
The superior individuals Pbest in each subspace play an important role statistics-based distribution estimation algorithm is used by building a
for the evolution. They can provide superior direction to help the pop- probabilistic model to describe the distribution of parameters. The
ulation evolve to the true Pareto front. Here, three mutation strategies Gaussian distribution N ðμ; σ 2 Þ is adopted to describe the distribution
(DE/best/2, DE/current-to-best/1 and DE/rand-to-best/1) based on su- of optimal parameters and to generate a new parameter population by

4
Z. Guo et al. Swarm and Evolutionary Computation 55 (2020) 100670

modeling and random sampling. The Gaussian density function is given

3.5. A mutation-adaption method
by (24) in which μ is the mean, and σ 2 is the variance.

1 ðxμÞ2 In the mutation-adaption method every individual has its own mu-
ϕðxÞ ¼ pffiffiffiffiffiffiffiffie 2σ2 ; ∞ < x < þ∞ (24)
2πσ tation strategy and self-adjustment by using the feedback information,
based on the mat. The procedure of the mutation-self-adaption method is
Z þ∞ given in Algorithm 5.
μ¼ xϕðxÞdx (25) Equation (32) is used as a statistical percentage of different mutation
∞
strategy in the generated population.
Z þ∞ X
σ2 ¼ ðx  μÞ2 ϕðxÞdx (26) oii ¼ ðmat ¼ iiÞ; ii 2 f1; 2; 3g;
∞
(32)
proii ¼ oii =ðo1 þ o2 þ o3 Þ; ii 2 f1; 2; 3g;
In the parameter-adaption method, the individuals with positive Δfk;i
The percentage pro includes the optimal mutation strategy informa-
represent the optimal group in the kth block, and their corresponding
tion and can be used to generate the different ratios of mutation strate-
parameters are superior to other individuals’ parameters, Rk ; k 2
gies for the next generation.
f1; 2; :::; ng is used to store the optimal parameters. The mean b μ and
variance b σ 2 in Gaussian distribution are estimated by equations (27)- Algorithm 5
(28), respectively, based on Rk ; k 2 f1; 2; :::; ng: Mutation-adaption.
X
μk ¼
b Rk =δRk ; k 2 f1; 2; …; ng (27)

1 X 2
σ 2k ¼
b μk ;
Rk;j  b k 2 f1; 2; …; ng; j 2 f1; 2; …; δRk g (28)
δRk  1

There are n blocks; every block has its superior parameter distribution
μk; b
N ðb σ k Þ; k ¼ f1;2;…;ng. Different blocks contain different excellent
2

characteristics of parameters. The storage of Δfk;i in V k ; k 2 f1; 2; :::; ng is

used to judge the contribution of the kth block in the current generation. A 4. Experiments and discussions
dynamically adjusting weight factor w is proposed as
X . X X X The MODE-ASP was compared with five state-of-the-art algorithms: a
wk ¼ Vk V1 þ V2 þ … þ Vn Þ; k 2 f1; 2; …; ng (29) decomposition MOEA with covariance matrix adaption evolution strat-
egy (MOEA/D-CMA) [36], a clustering-based adaptive MOEA (CAMOEA)
The sum of V k ; k 2 f1; 2; :::; ng represents the entire changes between [37], inverse modeling multi-objective evolutionary algorithm
the parents and the offsprings in the optimal direction of the kth block. (IMMOEA) [38], MOEA based on decision variables (MOEA/DVA) [39],
Bigger wk ; k ¼ f1; 2; :::; ng indicates the offsprings have changed a lot non-dominated sorting and local search (NSLS) [40] and a decision space
compared with the parents in the optimal direction, namely, the changes based niching NSGA-II (DN-NSGA-II) [18] on the tests problems ZDT
of the kth block have a great contribution to the whole evolution in the [41], DTLZ [42] and CEC0 09 test suites [43]. These test functions have
current generation. Hence, the weight factor is used to combine the pa- two-objectives or three-objectives. They deal with multi-objective prob-
rameters distribution of each block, and new Gaussian distribution lems in different ways. The PlatEMO in Matlab is used to implement the
N ðμ; σ 2 Þ is generated to model the parameters’ distribution. above algorithms [44]. The performance indicators Inverted Genera-
X tional Distance (IGD) [45] and Hypervolume (HV) [46] are used for
μ¼ μk ;
wk b k 2 f1; 2; …; ng (30) comparison.
X
σ2 ¼ σk;
wk b
2
k 2 f1; 2; …; ng (31) 4.1. Performance metrics

Where random sampling is carried out with the new distribution to (1) HV: This indicator adopts a volume that dominates the solution
generate a new parameter population, which include the optimal set to measure the convergence and diversity. Larger HV means
information of parameter in current generation.Algorithm 4 the algorithm is better.
Parameter-adaption. (2) IGD: The obtained solution set must approximate the PF and be
distribute in a uniform manner. This indicator is based on the
minimum Euclidean distance between every individual and the
true PF. Smaller the IGD value indicates the better algorithm.
(3) The Friedman test [47] and Wilcoxon’s rank-sum test [48] are
adopted. They are two non-parametric tests with a significance
level chosen equal to 0.05.

4.2. Experimental setting

In all the experiments, the population size is chosen as NP ¼ 100. The

maximum number generations is different depending on the different
benchmark functions: bi-objective function Gmax ¼ 250; DTLZ1-2:
Gmax ¼ 300; DTLZ4-5: Gmax ¼ 200; UF8-10 and DTLZ3: Gmax ¼ 500.
The dimensions of the decision variable D are also different: CEC0 09
benchmark function and ZDT1-3: D ¼ 30; ZDT4 and ZDT6: D ¼ 10;
DTLZ1: D ¼ 7; DTLZ2-5: D ¼ 12. In MODE-ASP, the initial mutation
parameter and crossover parameter are generated randomly in the range

5
Z. Guo et al. Swarm and Evolutionary Computation 55 (2020) 100670

Table 1
Hv results of MOEADCMA, CAMOEA, IMMOEA, MOEADVA, NSLA and MODE-ASP.
MOEADCMA [36] CAMOEA [37] IMMOEA [38] MOEADVA [39] NSLS [40] MODE-ASP

ZDT1 6.4478E-01(4) 7.1900E-01(5) 5.9299E-01(3) 0.0000Eþ00(1) 0.0000Eþ00(1) 8.6533E-01(6)

ZDT2 3.4143E-01(4) 4.3077E-01(5) 1.6440E-01(3) 0.0000Eþ00(1) 0.0000Eþ00(1) 5.3284E-01(6)
ZDT3 5.2690E-01(3) 6.4632E-01(4) 6.8464E-01(5) 0.0000Eþ00(1) 0.0000Eþ00(1) 1.0164Eþ00(6)
ZDT4 6.0103E-02(3) 3.0057E-01(4) 8.6539E-01(6) 0.0000Eþ00(1) 1.6087E-03(2) 6.5990E-01(5)
ZDT6 3.8858E-01(5) 3.8694E-01(4) 0.0000Eþ00(1) 3.0340E-01(2) 4.2729E-01(6) 3.3469E-01(3)
DTLZ1 4.4958E-01(4) 8.3594E-01(6) 0.0000Eþ00(1) 5.0517E-01(5) 3.7201E-02(2) 1.2799E-01(3)
DTLZ2 5.5147E-01(3) 5.5097E-01(2) 6.2771E-01(4) 5.3315E-01(1) 7.3786E-01(6) 7.1002E-01(5)
DTLZ3 1.5684E-01(4) 5.3515E-01(6) 0.0000Eþ00(1) 1.7937E-03(3) 0.0000Eþ00(1) 4.1875E-01(5)
DTLZ4 5.0387E-01(4) 5.3789E-01(5) 1.0029E-02(1) 2.4655E-01(3) 1.5120E-02(2) 7.1745E-01(6)
DTLZ5 1.8998E-01(4) 1.9891E-01(6) 1.2124E-01(1) 1.9548E-01(5) 1.3155E-01(2) 1.3267E-01(3)
UF1 6.1105E-01(2) 5.8555E-01(1) 7.2504E-01(4) 6.1609E-01(3) 8.0694E-01(6) 7.3331E-01(5)
UF2 6.7659E-01(3) 6.6044E-01(2) 8.2151E-01(6) 6.5062E-01(1) 8.0020E-01(4) 8.0150E-01(5)
UF3 4.9038E-01(4) 4.1224E-01(3) 6.0550E-01(6) 2.3208E-01(1) 3.0630E-01(2) 4.9851E-01(5)
UF4 3.1741E-01(1) 3.6608E-01(3) 3.8356E-01(4) 3.6406E-01(2) 3.9444E-01(5) 4.3063E-01(6)
UF5 0.0000Eþ00(1) 1.7621E-01(5) 0.0000Eþ00(1) 7.7556E-06(3) 3.2625E-03(4) 2.3877E-01(6)
UF6 1.8195E-01(3) 3.0405E-01(5) 2.0639E-01(4) 3.1607E-02(1) 9.4248E-02(2) 3.7723E-01(6)
UF7 5.2576E-01(4) 4.6459E-01(2) 4.9801E-01(3) 3.8382E-01(1) 5.3102E-01(5) 5.8960E-01(6)
UF8 3.2053E-01(2) 3.0873E-01(1) 4.1483E-01(4) 4.4370E-01(5) 5.3907E-01(6) 3.7269E-01(3)
UF9 5.1050E-01(3) 4.8849E-01(2) 5.1318E-01(4) 7.0574E-01(5) 8.3185E-01(6) 4.3901E-01(1)
UF10 2.7229E-04(1) 1.5713E-01(4) 3.3109E-01(6) 5.4514E-03(3) 4.9623E-03(2) 1.7535E-01(5)
Average 3.1 3.75 3.4 2.4 3.3 4.8

Table 2
IGD results of MOEADCMA, CAMOEA, IMMOEA, MOEADVA, NSLS and MODE-ASP.
MOEADCMA [36] CAMOEA [37] IMMOEA [38] MOEADVA [39] NSLS [40] MODE-ASP

ZDT1 5.7850E-02(3) 4.5517E-03(1) 1.7800E-01(4) 6.8561Eþ00(5) 7.4200Eþ00(6) 6.4226E-03(2)

ZDT2 8.4259E-02(3) 1.4083E-02(2) 2.9000E-01(4) 9.3622Eþ00(5) 9.4800Eþ00(6) 5.7950E-03(1)
ZDT3 7.1110E-02(3) 3.9802E-02(2) 1.7400E-01(4) 8.1715Eþ00(6) 7.6300Eþ00(5) 6.2783E-03(1)
ZDT4 1.1172Eþ00(4) 4.4452E-01(3) 6.2400E-03(1) 8.5117Eþ00(6) 1.4400Eþ00(5) 1.5147E-01(2)
ZDT6 3.1228E-03(1) 3.9268E-03(2) 2.1900Eþ00(6) 1.1700E-01(5) 9.2800E-03(3) 7.3001E-02(4)
DTLZ1 4.0777E-01(5) 2.1707E-02(1) 3.5200Eþ00(6) 1.2561E-01(3) 3.3400E-01(4) 4.6131E-02(2)
DTLZ2 5.8786E-02(4) 5.7676E-02(2) 9.4600E-02(6) 5.8398E-02(3) 5.6700E-02(1) 6.7275E-02(5)
DTLZ3 9.1300Eþ00(5) 6.5218E-02(1) 4.3500Eþ01(6) 1.5988Eþ00(3) 5.7300Eþ00(4) 4.3530E-01(2)
DTLZ4 2.0011E-01(4) 8.9664E-02(3) 8.7800E-02(2) 4.7517E-01(6) 3.3700E-01(5) 6.6554E-02(1)
DTLZ5 2.2434E-02(5) 5.2179E-03(1) 2.5000E-02(6) 9.1520E-03(4) 7.0800E-03(3) 5.8389E-03(2)
UF1 7.2887E-02(3) 1.1219E-01(6) 8.8000E-02(4) 6.9153E-02(2) 3.6900E-02(1) 8.9175E-02(5)
UF2 3.9832E-02(1) 5.2768E-02(6) 4.1100E-02(2) 4.5799E-02(4) 4.4600E-02(3) 4.9348E-02(5)
UF3 1.7584E-01(2) 2.3551E-01(3) 1.5300E-01(1) 4.0933E-01(6) 3.7100E-01(5) 2.8534E-01(4)
UF4 9.3318E-02(6) 6.0821E-02(2) 9.2800E-02(5) 5.6172E-02(1) 8.5200E-02(4) 6.4252E-02(3)
UF5 1.0560Eþ00(5) 3.5385E-01(2) 1.2200Eþ00(6) 9.5184E-01(4) 7.9700E-01(3) 3.4580E-01(1)
UF6 3.1768E-01(3) 1.9005E-01(2) 3.5800E-01(4) 5.7378E-01(6) 4.5800E-01(5) 1.7586E-01(1)
UF7 3.8767E-02(1) 1.1555E-01(4) 1.5700E-01(6) 1.4633E-01(5) 1.0200E-01(3) 8.9660E-02(2)
UF8 2.6591E-01(3) 2.7696E-01(6) 2.7500E-01(5) 1.1078E-01(1) 1.9200E-01(2) 2.7113E-01(4)
UF9 2.7887E-01(4) 2.7749E-01(3) 3.5900E-01(5) 9.3860E-02(1) 1.5200E-01(2) 3.9024E-01(6)
UF10 1.1442Eþ00(6) 3.9745E-01(3) 3.0800E-01(1) 6.4381E-01(4) 6.5200E-01(5) 3.9500E-01(2)
Average 3.55 2.75 4.2 4 3.75 2.75

[0.2, 0.7] and [0.1, 0.4], respectively; the initial three mutation strategies higher value means the solutions have better convergence and diversity.
are pro1 ¼ 0:35; pro2 ¼ 0:30; pro3 ¼ 0:35; the number of blocks is The proposed algorithm obtained 8 times in the first place and 7 times in
chosen as n ¼ 3; the label set is chosen as mat ¼ ∅. All the tests are run the second place. CAMOEA and NSLS are competitive. CAMOEA has
30 times independently. archived 8 times in the first place and in the second place. NSLS has
archived 7 times in the first place and in the second place. The CAMOEA
ranked in the second place. The score of HV metric equal to 3.75. It shows
4.3. Comparisons and analysis
that the clustering-based method is a very effective method when dealing
with MOPs. The NSLS shows a good performance when solving three-
4.3.1. Performance with ZDT,DTLZ and CEC0 09 benchmark functions
objective problems. MOEADCMA obtain the second place only 1 time.
This section presents the comparison results and statistical analysis
It has excellent performance when solving the MOPs with bias. The
results on HV and IGD, with all 20 benchmark functions. Tables 1 and 2
average ranking of MODE-ASP is 4.80. Comparing with other five algo-
give mean values and statistical results (in brackets) for the five com-
rithms, MODE-ASP demonstrates a better performance.
parison algorithms. The distribution of standard deviations is displayed
Table 2 presents the results of IGD metric obtained with five MOEAs
in Figs. 3–4. Tables 3 and 4 are obtained by testing the hypothesis that
algorithms and the proposed MODE-ASP. The average ranking of MODE-
there is no significant difference between HV and IGD values from
ASP is still in the first place with the score 2.75. It has archived 5 times in
MODE-ASP and other algorithms. The p-value is used to evaluate whether
the first place and 7 times in the second place. Smaller value of IGD in-
the null hypothesis is rejected or not with, the significance level equal to
dicates better performance. CAMOEA is competitive, with a score of IGD
0.05.
metric equal to 2.75. It has archived 3 times in the first place and 7 times
Table 1 presents the results of HV metric obtained with MOEADCMA,
in the second place. The remaining four compared algorithm has
CAMOEA, IMMOEA, MOEADVA, NSLA, and the proposed MODE-ASP.
archived no more than 3 times in the first place and no more than 2 times
The ranks of each algorithm are shown in brackets. For HV metric,

6
Z. Guo et al. Swarm and Evolutionary Computation 55 (2020) 100670

Fig. 3. Box plot of MOEADCMA, CAMOEA, IMMOEA, MOEADVA, NSLS and MODE-ASP for the index of HV (CEC0 09 UF).

Table 3 finding a good distribution solutions set on ZDT, DTLZ CEC0 09. The
IGD and HV ranking of the ZDT,DTLZ and CEC’09 benchmarks functions. performance of MODE-ASP to obtain an accurate solutions set approxi-
HV IGD mation to the Pareto front is better than the other five popular algo-
rithms. Simultaneously, through the comparison of HV and IGD value,
MOEADCMA [36] 3.13 3.55
CAMOEA [37] 3.75 2.75 the obtained solution by proposed MODE-ASP has both better diversity
IMMOEA [38] 3.45 4.20 and convergence.
MOEADVA [39] 2.48 4.00
NSLS [40] 3.40 3.75 4.3.2. Number of partitions
MODE-ASP 4.80 2.75
The number of partitions of MODE-ASP affects the algorithm’s per-
formance. To investigate the influence of number of partitions, self-
Table 4 adaptive mutation strategy and self-adaptive parameter strategy, we
P-values obtained by the IGD and HV of ZDT,DTLZ and CEC’09 benchmark chose the partition number as 1, 2, 3 with CEC0 09, DTLZ and ZDT
functions. benchmark functions. The symbols ‘  = þ = ’ means the experimental
MODE-ASP vs HV IGD result is significant worse/better/statistically than the comparison re-
sults. Tables 5 and 6 show the result of IGD and HV values with different
p-value Sign (0.05) p-value Sign (0.05)
partition numbers, with one/two/three sub-spaces as MODE-ASP-1/
MOEADCMA [36] 0.00573399 ● 0.04785751 ● MODE-ASP-2/MODE-ASP-3 respectively.
CAMOEA [37] 0.02063344 ● 0.88129271 –
In Tables 5 and 6, the larger the number of partitions, the better the
IMMOEA [38] 0.02762101 ● 0.02509351 ●
MOEADVA [39] 0.00642460 ● 0.00999639 ● performance with 3 being a suitable selection. MODE-ASP-3 obtained
NSLS [40] 0.01524006 ● 0.00999639 ● better results 15, 11 times on IGD, and 14, 10 times on HV better than
MODE-ASP-1 and MODE-ASP-2 respectively. However, the performance
of MODE-ASP-2 is good as well, especially with bi-objective problems. It
in the second place. Average performance of CAMOEA is quite same as is better 6 times on IGD and 5 times on HV than MODE-ASP-3. The
MODE-ASP. IGD values of those two algorithms are the smallest one
number of partitions is not fixed. It can be changed depending on the
compared with the other four algorithms and demonstrate the competi- complexity of functions. Fig. 5 shows the result in different subspaces,
tiveness in solving general MOPs. Figs. 3–4 show the distribution of HV
using DTLZ1, DTLZ3, ZDT6 and UF4. It is observed that the number of
and IGD in ZDT, DTLZ and CEC0 09. MODE-ASP is stable when solving individuals in each subspace changes along with the evolution
MOPs.
generation.
Apparently, the five other algorithms all have a good performance in

Fig. 4. Box plot of MOEADCMA, CAMOEA, IMMOEA, MOEADVA, NSLS, and MODE-ASP for the index of IGD (ZDT and DTLZ).

7
Z. Guo et al. Swarm and Evolutionary Computation 55 (2020) 100670

Table 5 Table 6
Influence of different partition number on MODE-ASP by IGD. Influence discussion of different partition number on MODE-ASP by HV.
MODE-ASP-1 MODE-ASP-2 MODE-ASP-3 MODE-ASP-1 MODE-ASP-2 MODE-ASP-3

ZDT1 2.7604E-02þ 4.7323E-03 6.4226E-03 ZDT1 8.2972E-01þ 8.6900E-01 8.6533E-01

ZDT2 3.8137E-02þ 4.9209E-03 5.7950E-03 ZDT2 4.7360E-01þ 5.3494E-01 5.3284E-01
ZDT3 2.9098E-02þ 5.5372E-03 6.2783E-03 ZDT3 9.5871E-01þ 1.0208Eþ00 1.0164Eþ00
ZDT4 1.7185E-01þ 1.9902E-01þ 1.5147E-01 ZDT4 6.3832E-01þ 5.9488E-01þ 6.5990E-01
ZDT6 2.9193E-01þ 8.9235E-02þ 7.3001E-02 ZDT6 1.2516E-01þ 3.1234E-01þ 3.3469E-01
DTLZ1 4.8659E-02þ 6.0561E-02þ 4.6131E-02 DTLZ1 1.2796E-01 1.2163E-01 1.2799E-01
DTLZ2 6.6877E-02 6.6803E-02- 6.7275E-02- DTLZ2 6.9450E-01þ 6.9956E-01þ 7.1002E-01
DTLZ3 3.9857E-01 7.2973E-01þ 4.3530E-01 DTLZ3 4.7813E-01 2.6064E-01þ 4.1875E-01
DTLZ4 6.8800E-02þ 6.9397E-02þ 6.6554E-02 DTLZ4 7.0730E-01þ 7.1479E-01 7.1745E-01
DTLZ5 5.7188E-03 5.8591E-03þ 5.8389E-03 DTLZ5 1.3236E-01 1.3237E-01 1.3267E-01
UF1 9.5452E-02þ 9.1443E-02þ 8.9175E-02 UF1 7.1963E-01þ 7.2715E-01þ 7.3331E-01
UF2 4.2213E-02 4.1262E-02 4.9348E-02 UF2 8.1087E-01 8.1427E-01 8.0150E-01
UF3 3.3197E-01þ 2.1319E-01 2.8534E-01 UF3 3.8092E-01þ 5.3402E-01 4.9851E-01
UF4 8.9990E-02þ 7.2529E-02þ 6.4252E-02 UF4 3.8784E-01þ 4.1701E-01þ 4.3063E-01
UF5 4.0128E-01þ 2.8315E-01- 3.4580E-01- UF5 1.8506E-01þ 2.7463E-01 2.3877E-01
UF6 2.4021E-01þ 1.7920E-01þ 1.7586E-01 UF6 2.7843E-01þ 3.6453E-01þ 3.7723E-01
UF7 2.2186E-01þ 1.0472E-01þ 8.9660E-02 UF7 4.6616E-01 5.8012E-01þ 5.8960E-01
UF8 2.8178E-01þ 2.7734E-01þ 2.7113E-01 UF8 3.5138E-01þ 3.4882E-01þ 3.7269E-01
UF9 3.1828E-01- 3.4557E-01 3.9024E-01 UF9 5.5801E-01 5.1608E-01 4.3901E-01
UF10 7.2296E-01þ 4.8906E-01þ 3.9500E-01 UF10 5.7246E-02þ 9.3131E-02þ 1.7535E-01
Kþ 15 11 Kþ 14 10
K 5 8 K 4 6
K 0 1 K 2 4

5. Optimization of the fermentation process of sodium gluconate equations are given as follows:
using MODE-ASP
r1
S þ X
!X;
5.1. Modeling of sodium gluconate fermentation process r2 (35)
C þ S þ X
!H2 O þ CO2 ;
r3
C þ S þ X
!P;
Sodium gluconate is a widely used raw material of industry. It is a
crystalline powder with white or yellowish color. It is salty, mild, and S; X; P represent the concentration of substrate (glucose, g/L), fer-
non-toxic. It has been widely used in construction, chemical food and menting microbe (Aspergillus niger, g/L), and product (sodium gluconate,
medical industries. The general method of sodium gluconate production g/L), respectively; C is the dissolved oxygen tension (%); r1 ; r2 ; r3
is fermentation. It uses Aspergillus niger as the major fermenting strain, represent the reaction rate of growth of bacteria, consumption of glucose,
glucose as carbon donor; and sodium hydroxide as neutralizer. Aspergillus and production of sodium gluconate.
niger is a species of aspergillus. It can produce various zymin and organic
acids. It effectively metabolizes the various monosaccharides in the 5.2. Construction of MOP in the fermentation process of sodium gluconate
environment. In the process of sodium gluconate production, GOD is
produced by Aspergillus niger. GOD results in the oxidative dehydroge- In the fermentation process, the more sodium gluconate obtained, the
nation of glucose, which increases the acidity of fermentation broth, better, the less consumption during the process, the better, the shorter
allowing sodium hydroxide to neutralize fermentation broth to maintain the reaction time, the better, and the less remaining glucose in the ter-
pH value. The whole reaction is complex. The following two reaction minal time, the better. The maximum output of sodium gluconate, the
equations are used to describe the main reaction: shortest reaction time, and the minimum remaining glucose are the best
C6H12O6þ1/2O2→C6H12O7 (33) combination. Three objective functions representing the requirements of
the main reaction are proposed. The ratio between sodium gluconate in
C6H12O7þNaOH→C6H12O6Na (34) the terminal time and the difference of glucose in the initial time and the
end time are defined as the conversion rate f1 . The amount of glucose in
The process of sodium gluconate production in the laboratory in-
terminal time is defined as f2 . The ratio between sodium gluconate in the
cludes two stages: the breeding process of strain and the production of
end time and the whole reaction time is defined as the utilization rate of
sodium gluconate. A model that contains the biological mechanism in-
equipment f3 . These three objective functions are all conflicting and
formation, was developed by Wang et al. [29]. Three main processes in
complying with each other, are generate a multi-objective optimization
fermentation process are taken into consideration: thallus growth, thallus
problem. The mathematical description is as follows:
metabolism, and product formation. The generalized stoichiometric

8
>
> FðxÞ ¼ fmaxðf1 Þ; minðf2 Þ; maxðf3 Þg;
>
>
>
> P
>
>
< f1 ¼ ðS  S Þ ;
0 e
Pmin < P < Pmax ; Smin < S < Smax ; tmin < t < tmax ; (36)
>
> f2 ¼ S e ;
>
>
>
>
>
>
: f3 ¼ P ;
t

8
Z. Guo et al. Swarm and Evolutionary Computation 55 (2020) 100670

Fig. 5. The distribution of individuals in different partition of DTLZ3,DTLZ1,ZDT6,UF4.

9
Z. Guo et al. Swarm and Evolutionary Computation 55 (2020) 100670

follows: the maximum generation Gmax equal to 100; the size of popu-
lation NP equal to 100; the block number n equal to 3; the range of
mutation control parameter F and crossover control parameter CR equal
to [0.2, 0.7], [0.1, 0.4], respectively. When the dissolved oxygen tension
C is below 45%, the aeration increases while the agitation speeds up to
maintain dissolved oxygen C.
Fig. 6 presents the Pareto front of sodium gluconate fermentation
process obtained by MODE-ASP. An optimal solution in the obtained
Pareto front is selected to compare with the experiment data. Details are
shown in Table 7.
In Table 7, all the compared algorithms are applied to fermentation
process. The output sodium gluconate in the fermentation process of
sodium gluconate is increased from 23.03 g/L to 26.35 g/L, increasing by
approximately 3.32 g/L. The amount of substrate concentration of
glucose is increased from 31.00 g/L to 33.63 g/L, enhancing by 2.63 g/L.
The remaining glucose and reaction time (28 h) had no significant dif-
ference. The indicators represented by the three objective functions ob-
tained by MODE-ASP improved the conversion rate and utilization rate of
equipment compared with the experimental data, from 91.13% to
Fig. 6. Pareto front of the production of sodium gluconate via the fermenta- 95.81%, and from 79.25% to 92.43%, respectively. Conversion rate ob-
tion method. tained by NSLS is 96.56% which is better than MODE-ASP. While the
remaining glucose and utilization rate of equipment are 8.64 g/L and
where t; S0 ; Se represent the reaction time, the concentration of 87.56%. They are worse than the results obtained by MODE-ASP. The
glucose in the initial time (g/L) and the concentration of glucose in the remaining four algorithms can successfully optimal the production of
terminal time, respectively. sodium gluconate compared with the experiment data. They have no
Dissolved oxygen is an important variable. It can directly affects the obvious advantage compared with MODE-ASP. This performance shows
bacterial growth and sodium gluconate production. Initially, the dis- that the MODE-ASP can optimize the fermentation process of sodium
solved oxygen should be big enough to ensure the growth of bacterial. gluconate effectively and efficiently.
The original kinetic model built by Wang et al. treats the dissolved ox- Table 8 shows the Friedman test results of all comparison algorithm
ygen tension as a continuous uncontrollable variable, decreasing with for gluconate production process. The value of objectives f1 ; f3 is the
time. In the laboratory, the dissolved oxygen tension can be controlled by larger the better. The value of objective f2 is the smaller the better. The
adjusting the aeration rate and agitation speed. Here, discretizing the standard of evaluation is that if the ranking is larger the result is better.
dissolved oxygen tension is treated as a control variable to ensure it can The ranking of MODE-ASP is 6.00 the largest one which shows the
be dynamically adjusted at different times. advantage of the proposed algorithm.
Simultaneously, the Pareto front obtained by MODE-ASP not only
5.3. Optimization of the fermentation process of sodium gluconate by provides one solution but a solution set, thus providing more choice for
MODE-ASP experimenters to select a suitable combination. The details of time-
varying concentration of sodium gluconate and glucose generated by
The operating conditions of the fermentation process of sodium glu- the selected optimal solution is shown in Figs. 7–8.
conate are optimized by MODE-ASP using the experimental data ob-
tained in the laboratory. In MODE-ASP the parameters are chosen as

Table 7
Optimization results of fermentation process of sodium gluconate.
the biomass the concentration of the substrate time conversion the remaining utilization rate of
concentration (g/L) sodium gluconate (g/L) concentration of glucose (h) ratef1 s (%) glucosef2 (g/L) equipmentf3 (%)
(g/L)

Experiment Initial 0.01 0.84 31.00 0 – 31.00 –

data End – 23.03 6.65 28 91.13 6.65 79.25
MOEADCMA Initial 0.014 0.9 27.59 0 – 27.59 –
[36] End 21.78 5.93 28 92.26 5.93 74.56
CAMOEA [37] Initial 0.014 0.81 31.5 0 – 31.5 –
End – 24.25 7.80 28 95.47 7.80 83.7
IMMOEA [38] Initial 0.014 0.82 29.23 0 – 29.23 –
End – 22.74 6.70 28 93.67 6.70 78.29
MOEADVA Initial 0.014 0.57 30.78 0 – 30.78 –
[39] End 23.28 7.45 28 94.92 7.45 81.12
NSLS [40] Initial 0.014 0.9 33 0 – 33 –
End – 25.41 8.64 28 96.56 8.64 87.56
MODE-ASP Initial 0.014 0.47 33.63 0 – 33.63 –
End – 26.35 6.62 28 95.81 6.62 92.43

Table 8
Ranking of Friedman test for gluconate production process.
Experiment data MOEADCMA [36] CAMOEA [37] IMMOEA [38] MOEADVA [39] NSLS [40] MODE-ASP

ranking 3.67 3.00 4.00 3.00 3.67 4.67 6.00

10
Z. Guo et al. Swarm and Evolutionary Computation 55 (2020) 100670

Declaration of competing interest

None.

Acknowledgments

The author are grateful for the support of National Natural Science
Foundation of China (21878081), Fundamental Research Funds for the
Central Universities under Grant of China (222201917006), and the
Program of introducing Talents of Discipline to Universities (the 111
Project) under Grant B17017.

References

[1] S. Mirjalili, P. Jangir, S. Saremi, Multi-objective ant lion optimizer: a multi-

objective optimization algorithm for solving engineering problems, Appl. Intell. 46
(2017) 79–95.
[2] L.A. Pereira, S. Haffner, G. Nicol, T.F. Dias, Multiobjective optimization of five-
phase induction machines based on NSGA-II, IEEE Trans. Ind. Electron. 64 (2017)
Fig. 7. Time-varying concentration of sodium gluconate. 9844–9853.
[3] R. Datta, S. Pradhan, B. Bhattacharya, Analysis and design optimization of a robotic
gripper using multiobjective genetic algorithm, IEEE Trans. Syst. Man Cybernet.:
Systems 46 (2016) 16–26.
[4] E. Zitzler, L. Thiele, Multiobjective evolutionary algorithms: a comparative case
study and the strength Pareto approach, IEEE Trans. Evol. Comput. 3 (1999)
257–271.
[5] H. Han, W. Lu, J. Qiao, An adaptive multiobjective particle swarm optimization
based on multiple adaptive methods, IEEE Trans. Cybern. 47 (2017) 2754–2767.
[6] S. Jiang, S. Yang, An improved multiobjective optimization evolutionary algorithm
based on decomposition for complex pareto fronts, IEEE Trans. Cybern. 46 (2016)
421–437.
[7] R. Storn, K. Price, Differential Evolution-A Simple and Efficient Adaptive Scheme
for Global Optimization over Continuous Spaces, ICSI Berkeley, 1995.
[8] J. Bader, E. Zitzler, HypE: an algorithm for fast hypervolume-based many-objective
optimization, Evol. Comput. 19 (2011) 45–76.
[9] S. Jiang, J. Zhang, Y.-S. Ong, A.N. Zhang, P.S. Tan, A simple and fast hypervolume
indicator-based multiobjective evolutionary algorithm, IEEE Trans. Cybern. 45
(2015) 2202–2213.
[10] Y. Tian, R. Cheng, X. Zhang, F. Cheng, Y. Jin, An indicator-based multiobjective
evolutionary algorithm with reference point adaptation for better versatility, IEEE
Trans. Evol. Comput. 22 (2017) 609–622.
[11] Y. Yuan, H. Xu, B. Wang, X. Yao, A new dominance relation-based evolutionary
algorithm for many-objective optimization, IEEE Trans. Evol. Comput. 20 (2016)
16–37.
[12] J. Molina, L.V. Santana, A.G. Hernandez-Díaz, C.A.C. Coello, R. Caballero, g-
Fig. 8. Changing concentration of glucose with time. dominance, Reference point based dominance for multiobjective metaheuristics,
Eur. J. Oper. Res. 197 (2009) 685–692.
[13] A. Mousa, M. El-Shorbagy, W. Abd-El-Wahed, Local search based hybrid particle
6. Conclusions swarm optimization algorithm for multiobjective optimization, Swarm Evol.
Comput. 3 (2012) 1–14.
In the current study, an objective space partitioning and parameter [14] W. Hu, G.G. Yen, Adaptive multiobjective particle swarm optimization based on
parallel cell coordinate system, IEEE Trans. Evol. Comput. 19 (2015) 1–18.
adaption multi-objective DE algorithm based on angle (MODE-ASP) is [15] L.C.T. Bezerra, M. L opez-Iban~ ez, T. Stützle, Automatic component-wise design of
proposed. In MODE-ASP, the objective space is divided into several multiobjective evolutionary algorithms, IEEE Trans. Evol. Comput. 20 (2016)
subspaces based on the vector angle between chosen reference points and 403–417.
[16] Z. Li, S. Li, C. Yue, Z. Shang, B. Qu, Differential evolution based on reinforcement
fitness values in each sub-population. The cosine distance is used to select learning with fitness ranking for solving multimodal multiobjective problems,
the superior individuals, and changing of fitness values between parents Swarm Evol. Comput. 49 (2019) 234–244.
and offsprings expresses the contribution of its corresponding parame- [17] W. Zhang, G. Li, W. Zhang, J. Liang, G.G. Yen, A Cluster Based PSO with Leader
Updating Mechanism and Ring-Topology for Multimodal Multi-Objective
ters. A self-adaptive parameter strategy is proposed to extract the optimal
Optimization, Swarm and Evolutionary Computation, 2019, p. 100569.
parameter information in each subspace and to generate the optimal [18] J. Liang, C. Yue, B. Qu, Multimodal multi-objective optimization: a preliminary
parameters in order to guide the whole population towards the true study, in: 2016 IEEE Congress on Evolutionary Computation (CEC), IEEE, 2016,
Pareto front. In this process, every individual has its own parameters. The pp. 2454–2461.
[19] S. Kukkonen, J. Lampinen, GDE3: the third evolution step of generalized differential
results of IGD and HV obtained with 20 benchmark functions showed evolution, in: Evolutionary Computation, 2005. The 2005 IEEE Congress on, IEEE,
that the proposed MODE-ASP outperformed the five other multi- 2005, pp. 443–450.
objective evolutionary algorithms. The proposed algorithm is optimal [20] W. Qian, Adaptive differential evolution algorithm for multiobjective optimization
problems, Appl. Math. Comput. 201 (2008) 431–440.
in both convergence and diversity. Meanwhile, as compared to the [21] Y.-N. Wang, L.-H. Wu, X.-F. Yuan, Multi-objective self-adaptive differential
experimental data of the fermertation process of sodium gluconate, the evolution with elitist archive and crowding entropy-based diversity measure, Soft
MODE-ASP is able to get higher conversion rate and utilization rates of Comput. 14 (2010) 193.
[22] S.M. Venske, R.A. Gonçalves, M.R. Delgado, ADEMO/D: multiobjective
equipment with the same remaining glucose. In this process, this shows optimization by an adaptive differential evolution algorithm, Neurocomputing 127
MODE-ASP is able to optimize the fermentation process of sodium glu- (2014) 65–77.
conate efficiently. The acquired Pareto front provides not only a single [23] M. Ali, P. Siarry, M. Pant, An efficient differential evolution based algorithm for
solving multi-objective optimization problems, Eur. J. Oper. Res. 217 (2012)
combination of the experimental conditions but also a solution set. Thus, 404–416.
the experimenters can select a proper one to guide production, and [24] H. Li, K. Deb, Q. Zhang, P. Suganthan, L. Chen, Comparison between MOEA/D and
achieve high production efficiency. NSGA-III on a set of novel many and multi-objective benchmark problems with
challenging difficulties, Swarm Evol. Comput. 46 (2019) 104–117.

11
Z. Guo et al. Swarm and Evolutionary Computation 55 (2020) 100670

[25] R. Cheng, M. Li, Y. Tian, X. Xiang, X. Zhang, S. Yang, Y. Jin, X. Yao, in: Benchmark [37] Y. Hua, Y. Jin, K. Hao, A clustering-based adaptive evolutionary algorithm for
Functions for the Cec’2018 Competition on Many-Objective Optimization, 2018. multiobjective optimization with irregular pareto fronts, IEEE Trans. Cybern. 49
[26] X. Han, G. Liu, W. Song, Y. Qu, Production of sodium gluconate from delignified (2018) 2758–2770.
corn cob residue by on-site produced cellulase and co-immobilized glucose oxidase [38] R. Cheng, Y. Jin, K. Narukawa, B. Sendhoff, A multiobjective evolutionary
and catalase, Bioresour. Technol. 248 (2018) 248–257. algorithm using Gaussian process-based inverse modeling, IEEE Trans. Evol.
[27] J. Plank, Applications of biopolymers and other biotechnological products in Comput. 19 (2015) 838–856.
building materials, Appl. Microbiol. Biotechnol. 66 (2004) 1–9. [39] X. Ma, F. Liu, Y. Qi, X. Wang, L. Li, L. Jiao, M. Yin, M. Gong, A multiobjective
[28] F. Lu, Z. Wang, W. Zhao, J. Chu, Y. Zhuang, A simple novel approach for real-time evolutionary algorithm based on decision variable analyses for multiobjective
monitoring of sodium gluconate production by on-line physiological parameters in optimization problems with large-scale variables, IEEE Trans. Evol. Comput. 20
batch fermentation by Aspergillus Niger, Bioresour. Technol. 202 (2016) 133–141. (2015) 275–298.
[29] X. Wang, X. Yan, F. Lu, M. Guo, Y. Zhuang, Power series kinetic model based on [40] B. Chen, W. Zeng, Y. Lin, D. Zhang, A new local search-based multiobjective
generalized stoichiometric equations for microbial production of sodium gluconate, optimization algorithm, IEEE Trans. Evol. Comput. 19 (2015) 50–73.
Biochem. Eng. J. 113 (2016) 30–36. [41] E. Zitzler, K. Deb, L. Thiele, Comparison of multiobjective evolutionary algorithms:
[30] J.-Z. Liu, L.-P. Weng, Q.-L. Zhang, H. Xu, L.-N. Ji, A mathematical model for empirical results, Evol. Comput. 8 (2000) 173–195.
gluconic acid fermentation by Aspergillus Niger, Biochem. Eng. J. 14 (2003) [42] K. Deb, L. Thiele, M. Laumanns, E. Zitzler, Scalable multi-objective optimization test
137–141. problems, in: Evolutionary Computation, 2002. CEC’02. Proceedings of the 2002
[31] Y. Dong, Q. Fan, X. Yan, M. Guo, F. Lu, Development of a hybrid model for sodium Congress on, IEEE, 2002, pp. 825–830.
gluconate fermentation by Aspergillus Niger, J. Chem. Technol. Biotechnol. 89 [43] Q. Zhang, A. Zhou, S. Zhao, P.N. Suganthan, W. Liu, S. Tiwari, Multiobjective
(2014) 1875–1882. Optimization Test Instances for the CEC 2009 Special Session and Competition,
[32] R. Storn, On the usage of differential evolution for function optimization, in: Fuzzy University of Essex, 2008.
Information Processing Society, 1996. NAFIPS., 1996 Biennial Conference of the [44] Y. Tian, R. Cheng, X. Zhang, Y. Jin, PlatEMO: a MATLAB platform for evolutionary
North American, IEEE, 1996, pp. 519–523. multi-objective optimization [educational forum], IEEE Comput. Intell. Mag. 12
[33] A.K. Qin, V.L. Huang, P.N. Suganthan, Differential evolution algorithm with (2017) 73–87.
strategy adaptation for global numerical optimization, IEEE Trans. Evol. Comput. [45] H. Li, Q. Zhang, Multiobjective optimization problems with complicated Pareto sets,
13 (2009) 398–417. MOEA/D and NSGA-II, IEEE Trans. Evol. Comput. 13 (2009) 284–302.
[34] A. Kishor, P.K. Singh, J. Prakash, NSABC: non-dominated sorting based multi- [46] E. Zitzler, D. Brockhoff, L. Thiele, The hypervolume indicator revisited: on the
objective artificial bee colony algorithm and its application in data clustering, design of pareto-compliant indicators via weighted integration, in: International
Neurocomputing 216 (2016) 514–533. Conference on Evolutionary Multi-Criterion Optimization, 2007, pp. 862–876.
[35] K. Deb, A. Pratap, S. Agarwal, T. Meyarivan, A fast and elitist multiobjective genetic [47] M. Friedman, The use of ranks to avoid the assumption of normality implicit in the
algorithm: NSGA-II, IEEE Trans. Evol. Comput. 6 (2002) 182–197. analysis of variance, Publ. Am. Stat. Assoc. 32 (1939) 675–701.
[36] H. Li, Q. Zhang, J. Deng, Biased multiobjective optimization and decomposition [48] F. Wilcoxon, Individual comparisons by ranking methods, Proc. Biometr. Bull. 1
algorithm, IEEE Trans. Cybern. 47 (2016) 52–66. (1945) 80–83.

12