Sie sind auf Seite 1von 12

Chemometrics and Intelligent Laboratory Systems 136 (2014) 85–96

Contents lists available at ScienceDirect

Chemometrics and Intelligent Laboratory Systems


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

Multi-objective differential evolution with ranking-based mutation


operator and its application in chemical process optimization
Xu Chen, Wenli Du ⁎, Feng Qian
Key Laboratory of Advanced Control and Optimization for Chemical Processes, Ministry of Education, East China University of Science and Technology, Shanghai 200237, China

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

Article history: Dynamic optimization problems in chemical processes are often quite challenging because these problems often
Received 14 January 2014 involve multiple and conflicting objectives. To solve the multi-objective dynamic optimization problems
Received in revised form 24 April 2014 (MDOPs), in this paper, we propose a new multi-objective differential evolution (MODE) variant, named
Accepted 7 May 2014
MODE-RMO for short, inspired by the phenomenon that good individuals which contain good information
Available online 17 May 2014
often have more chance to be utilized to guide other individuals. In MODE-RMO, the ranking-based mutation op-
Keywords:
erator is integrated into the MODE algorithm to accelerate the convergence speed, and thus enhance the perfor-
Dynamic optimization mance. The performance of our proposed algorithm is firstly evaluated in ten test functions and compared with
Multi-objective optimization other MOEAs. The results demonstrate that MODE-RMO can generate Pareto optimal fronts with satisfactory con-
Differential evolution vergence and diversity. Finally, MODE-RMO is applied to solve three MDOPs taken from literature using control
Ranking-based mutation operator vector parameterization. The obtained results indicate that MODE-RMO is an effective and efficient approach for
MDOPs.
© 2014 Elsevier B.V. All rights reserved.

1. Introduction only find one solution at each run; therefore, it has to be run many times
to obtain the Pareto optimal sets.
Most of optimization problems in chemical processes are dynamic In the past ten years, multi-objective evolutionary algorithms
optimization problems, which are described by a set of nonlinear differ- (MOEAs) have gained great popularity in handling MDOPs because of
ential equations. In these dynamic optimization problems, there may their ability to yield a number of Pareto optimal solutions in a single
exist several objectives conflicting to each other [1]. For instance, run. Sarkar and Modak employed the enhanced version of nondominated
some chemical processes are characterized by a conflict between yield sorting genetic algorithm (i.e. NSGA-II) to solve multi-objective optimal
and productivity, and the optimization problem is to find an operating control problems of fed-batch bioreactors [5]. In [6], He and Chen pro-
mode that achieves the best tradeoff between the yield and the produc- posed a method called immune mechanism based multi-objective ant
tivity [2]. In multi-objective optimization problems (MOPs), there does colony algorithm to solve a MDOP with both path and terminal con-
not exist a solution which simultaneously optimizes all objectives. straints. For the optimal design of the bioreactors, Mandli and Modak de-
Instead, it gives rise to a set of tradeoff solutions, which are called as veloped a computational methodology based on genetic algorithm to
Pareto optimal sets. The designer can select one solution from the set solve the challenging multi-objective optimization problem with multiple
of solutions according to their preference. decision variables [7]. In order to maximize the final product while reduc-
The research on multi-objective dynamic optimization problems ing the by-product in batch process, Jia et al. proposed a novel diversity
(MDOPs) has attracted much attention over the years. Traditional preservation strategy in multi-objective particle swarm optimization
method transforms the multi-objective functions into single objective (MOPSO) to enhance the convergence and diversity of the attained Pareto
function by weighted combination, and then the methods used in single optimal set [8]. Recently, Wang et al. presented a modified Bare-bones
objective optimization can be employed [3,4]. Wang and Shieh extend- MOPSO algorithm for dynamic multi-objective optimization of chemical
ed iterative dynamic programming to solve MDOPs through converting processes [9]. The research in this direction is quite promising, which en-
the original problems into single-objective problem including con- courages us to develop new MOEA to solve MDOPs.
straints [3]. To mitigate the drawbacks of the classic weighted sum ap- Differential evolution (DE) is one of the most powerful stochastic
proach, Logist et al. investigates two alternative approaches, normal search algorithms, originally proposed by Storn and Price [10]. After
boundary intersection and normalized normal constraint, for multiple its great success in single-objective optimization problems (SOPs),
objective optimal control of (bio)chemical processes [4]. These methods some researchers have extended it to solve MOPs, such as Pareto-
frontier differential evolution (PDE) [11], Pareto differential evolution
⁎ Corresponding author. Tel.: +86 21 64252060. approach (PDEA) [12], differential evolution for multi-objective optimi-
E-mail address: wldu@ecust.edu.cn (W. Du). zation (DEMO) [13], the third version of generalized differential

http://dx.doi.org/10.1016/j.chemolab.2014.05.007
0169-7439/© 2014 Elsevier B.V. All rights reserved.
86 X. Chen et al. / Chemometrics and Intelligent Laboratory Systems 136 (2014) 85–96

evolution (GDE3) [14], multi-objective self-adaptive DE (MOSaDE) Reciprocally, u2 is said to be dominated by u1, denoted as u2 ≺ u1.
[15], and so on. A detailed survey of multi-objective DE has been pub-
lished in [16], in which the advantages and disadvantages of the most Definition 2. A solution u* is said to be a Pareto optimal solution if there
known multi-objective differential evolution (MODE) methods are does not exist another solution u such that u ≻ u *. All Pareto optimal so-
discussed. lutions constitute a Pareto optimal set, denoted as U *.
Recently, inspired by the nature phenomenon that good individuals Definition 3. Pareto front is defined as PF = {J(u*), u* ∈ U*}, i.e., the
which contain good information often have more chance to be utilized mapping of Pareto optimal set U* in the objective space.
to guide other individuals, Gong and Cai proposed a ranking-based mu-
tation operator for the single-objective DE algorithm [17]. In this study,
by incorporating fast nondominated sorting and crowding distance, we 2.2. Differential evolution
extend this ranking-based mutation operator for MODE, and hence, a
new MODE variant, namely multi-objective differential evolution with The DE algorithm is a stochastic, population-based algorithm devel-
ranking-based mutation operator (MODE-RMO, for short) is proposed. oped by Storn and Price, designed for optimization problems in con-
The ranking-based mutation operator can enhance the performance of tinuous search space [10]. It has won the third place at the first
MODE by accelerating the convergence rate. To evaluate the perfor- international contest on evolutionary computation in a real-valued
mance of MODE-RMO, it is firstly used to solve ten classic test functions function test-suite. DE has its own evolutionary strategies: mutation,
and compared with other efficient MOEAs. Finally, by combining with crossover and selection. Different from genetic algorithm, mutation is
control vector parameterization (CVP), MODE-RMO is applied to solve the key operator of DE. In the mutation operation, three distinct individ-
three MDOPs taken from literature. The obtained results indicate that uals xgr1, xgr2, and xgr3 are randomly selected from the current population,
MODE-RMO is an effective and efficient approach for MDOPs. and the mutation operator is mathematically described as:
The rest of this paper is organized as follows. Section 2 briefly de- gþ1 g  g g 
scribed the formulation of MDOP and the DE algorithm. Section 3 pre- mi ¼ xr1 þ F  xr2 −xr3 ð5Þ
sents our proposed algorithm MODE-RMO in detail. Section 4 presents g
comparative results between MODE-RMO and other MOEAs, and the where mi + 1 is the mutant vector; r1, r2 and r3 are three mutually exclu-
effectiveness of ranking-based mutation operator is also demonstrated. sive integers different from index i; F is the mutation scale factor usually
In Section 5, MODE-RMO is applied to solve three MDOPs of chemical in [0,1].
g
processes using CVP. Finally, we draw some conclusions. The crossover operation modifies the mutant vector mi + 1 and
g+1
obtains the trial vector vi as follows:
2. Preliminary ( gþ1
gþ1 mij ; if ðrand≤CRÞor j ¼ sn
vij ¼ g ; j ¼ 1; 2; …; D ð6Þ
This section firstly introduces the formulation of MDOP and some xij ; otherwise
basic concepts for multi-objective optimization; and then, the proce-
dure of the DE algorithm is briefly described. where rand generates a random value in [0,1]; CR is the crossover con-
stant in [0,1]; sn is an arbitrary number in {1,2,…,D} that ensure vgi + 1
g
2.1. Multi-objective dynamic optimization problem gets at least one parameter from mi + 1.
The selection used in DE is a greedy operator. The better individual
g g
A typical MDOP can be formulated as follows [4]: between vi + 1 and xi is survived into the next generation:
(
min f J 1 ; J 2 ; …; J m g ð1Þ gþ1
gþ1
vi ; if vi
gþ1 g
is better than xi
xi ¼ g : ð7Þ
xi ; otherwise
subject to:

dx h i DE repeats the above three operators until a termination criterion is


¼ f ðxðt Þ; uðt Þ; t Þ; t∈ t 0 ; t f ð2Þ satisfied.
dt

3. Our proposed method: MODE-RMO


L U
u ≤u ≤u ð3Þ
In this section, the proposed MODE-RMO algorithm is presented in
where x(t)∈Rn are the state variables, with initial conditions x(t0) = x0; detail. In MODE-RMO, the ranking-based mutation operator is intro-
u∈Rm are the control variables. The function f represents the dynamic duced into MODE to enhance its performance by accelerating the
system equations describing the model of the chemical process (on convergence speed. Therefore, in the following part, we will firstly de-
the time interval [t0,tf]). Each of the objective functions Ji (i = 1,…,m) scribe the ranking-based mutation operator for multi-objective optimi-
can be of the following type: zation; and then our MODE-RMO is minutely elucidated.
    Z tf
J i ¼ ψi x t f ; t f þ ϕi ðxðt Þ; uðt Þ; t Þdt ð4Þ 3.1. Ranking-based mutation operator for multi-objective optimization
t0

t
In nature, good individuals always contain good information, and
with ψi(x(tf), tf) the Mayer item and ∫t f ϕi ðxðt Þ; uðt Þ; t Þdt the Lagrangian they are more likely to be utilized to guide other individuals. Based on
0
term. The Mayer item usually represents the terminal cost and the these considerations, Gong and Cai proposed a ranking-based mutation
Lagrangian term represents the integral cost. operator for DE in single-objective optimization [17], in which the par-
Contrary to SOPs, no single global solution exists in MOPs. Therefore, ents are proportionally selected according to their rankings in the
it is necessary to introduce the concept of Pareto optimal solutions in current population. The higher ranking a parent obtains, the more op-
multi-objective optimization. portunity it will be selected. The authors integrated the proposed
ranking-based mutation operator into some DE algorithms, and experi-
Definition 1. A solution u1 is said to dominate another solution u2, mental results indicated that the ranking-based mutation operator is
denoted as u1 ≻ u2,if and only if ∀ i ∈ {1, 2, …, m}, Ji(u1) ≤ Ji(u2) and able to enhance the performance of the DE algorithms in single-
∃ j ∈ {1, 2, …, m}, Jj(u1) b Jj(u2). objective optimization.
X. Chen et al. / Chemometrics and Intelligent Laboratory Systems 136 (2014) 85–96 87

To our knowledge, until now, the ranking-based mutation operator the population to the offspring. In our ranking-based mutation operator
has not been introduced into MODE for MOPs. Therefore, we are inter- for multi-objective optimization, only the base vector and the terminal
ested in extending this operator for MODE. However, the challenge for vector are selected based on their selection probabilities, while the
this extension is that, the population can be directly sorted from best starting vector is randomly selected. This is because if the two vectors
to worst in single-objective optimization, while there exist many solu- in the difference vector are both chosen from better vectors, then the
tions which are nondominated with each other in multi-objective opti- search step-size of the difference vector may decrease quickly and
mization. Considering that the MODE algorithm needs to generate may lead to premature convergence [17,19].
approximate Pareto solutions with both good convergence and spread,
hence, fast nondominated sorting and crowding distance are incorpo- 3.2. Multi-objective differential evolution with ranking-based mutation
rated into ranking-based mutation operator to deal with MOPs. operator

3.1.1. Fast nondominated sorting and crowding distance In MODE, the crossover is conducted in the same way as that in
The fast nondominated sorting and crowding distance are proposed single-objective optimization. However, the selection needs to be
by Deb et al. in the NSGA-II algorithm [18]. In the fast nondominated redesigned, because the trial vector and the target vector are often
sorting procedure, for each solution, we calculate two entities: 1) dom- nondominated with each other. In MODE-RMO, we use the following
ination count np, i.e. the number of solutions which dominate the solu- selection operator, which has three steps:
tion p; and 2) Sp: a set of solutions that the solution dominates. All
solutions in the first nondominated front will have their domination (i) If the trial vector dominates the target vector, use the trial vector
count as zero. Now, for each solution p with np = 0, we visit each mem- to replace the target vector.
ber (q) of its set Sp and reduce its domination count by one. In doing so, (ii) If the target vector dominates the trial vector, the trial vector is
if for any member the domination count becomes zero, we put it in a discarded.
separate list Q. These members belong to the second nondominated (iii) Otherwise, the trial vector is added to the population.
front. Now, the above procedure is continued with each member of
and the third front is identified. This process continues until all fronts Thus, at the end of a generation, total size of the population is be-
are identified. tween NP and 2NP. This population is truncated for the next step of
Crowding distance is used to get an estimate of the density of solu- the algorithm. The truncation process consists of nondominated sorting
tions surrounding a particular individual i in the population, it calcu- and evaluating the individuals of the same front with the crowding dis-
lates the average distance of two solutions on either side of solution i tance. The truncation procedure keeps only the best NP vectors in the
(i.e. i + 1 and i − 1) along each of the objectives. The crowding- population.
distance computation requires sorting the population according to Now, by integration with the ranking-based mutation operator, the
each objective function value in ascending order of magnitude. Thereaf- MODE-RMO algorithm can be described as follows:
ter, for each objective function, the boundary solutions (solutions with
smallest and largest function values) are assigned an infinite distance Step1: Set the generation number g = 0, randomly initialize the pop-
value. All other intermediate solutions are assigned a distance value ulation P g = {xg1, xg2, …, xgNp} with xgi = {xgi1, xgi2, …, xgiD}(i = 1, …, Np)
equal to the absolute normalized difference in the function values of uniformly distributed in search space. Set mutation scale factor F,
two adjacent solutions. This calculation is continued with other objec- crossover constant CR, and the maximum number of generations
tive functions. The overall crowding-distance value is calculated as the Maxgen.
sum of individual distance values corresponding to each objective.
Step2: Evaluate the fitness value of each target vector xgi .
Each objective function is normalized before calculating the crowding
distance. Step3: For each individual vector, do the following steps from step4
to step7.
3.1.2. Ranking assignment and selection probability Step4: Determine the selected vector indexes r1, r2, and r3 through
After obtaining the nondomination front number ifront and crowding the method described in Fig. 1.
distance icd for each solution i, we can define a partial order for the Step5: Use ranking-based mutation operator to generate a mutant
whole population. The solution i is better than solution j, if any of the vector mgi + 1 corresponding to the target vector xgi according to
two conditions is satisfied: (i) ifront b jfront; (ii) ifront = jfront and icd N jcd. Eq. (5).
Now, we can sort the population in ascending order based on the Step6: Use crossover operation to generate a trial vector vgi + 1 for
partial order defined above. The ranking of an individual is assigned as
each target vector xgi according to Eq. (6).
follows:
Step7: Evaluate the trial vector vgi + 1, and use the following selection
Ri ¼ Np−i; i ¼ 1; 2; …; Np ð8Þ operation:
g g g g
(i) If vi + 1 dominates xi , xi + 1 = vi + 1.
g g+1 g+1
where Np is the population size. According to Eq. (8), the best individual (ii) If xi dominates vi , vi is rejected.
g g g
in the current population will obtain the highest ranking. (iii) If xi and vi + 1 are nondominated with each other, add vi + 1
After assigning the ranking for each individual, the selection proba- to the population.
bility pi of the ith individual is calculated as: Step8: After Step7, the size of population ranges from Np to 2Np.
Sorting the population based on fast nondominated sorting and
Ri
pi ¼ ; i ¼ 1; 2; …; Np: ð9Þ crowding distance, and the best Np individuals are survived into
Np
the next generation.
Step9: Set g = g + 1, return back to Step 3 until the maximum
After calculating the selection probability of each individual by
generations Maxgen has been reached.
Eq. (9), the ranking-based mutation operator of “DE/rand/1” for multi-
objective optimization can be described in Fig. 1. From the approach It is worth noting that, the current population has been sorted based
in Fig. 1, the individual with a higher ranking will have a larger probabil- on fast nondominated sorting and crowding distance in Step8 of the last
ity to be selected as base vector or terminal vector in the mutation oper- generation; and hence, from the second generation, it does not need to
ator; therefore, it is beneficial for the propagation of good information in repeatedly sort the current population in Step4.
88 X. Chen et al. / Chemometrics and Intelligent Laboratory Systems 136 (2014) 85–96

Fig. 1. Ranking-based mutation operator of “DE/rand/1” for multi-objective optimization.

The advantages of our proposed MODE-RMO can be listed as the set of uniformly distributed Pareto optimal solutions P*. It is de-
follows: fined as:
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
uX   2
(i) Since the good parents are more likely to be chosen in the   u
 t v∈P d v; P
ranking-based mutate operator, it is helpful to enhance the per- GD P; P ¼ ð10Þ
jP j
formance of the MODE algorithm.
(ii) Compared with the original MODE algorithm, MODE-RMO only
where d(v,P*) is the minimum Euclidean distance between v and
needs a small extra computational cost in sorting the population
points in P*, |P| is the number of points in P.
and selecting the individuals. Our approach does not increase the
The second metric used is inverted generational distance (IGD) [20],
overall complexity of the original MODE algorithm.
which can be defined as
(iii) Different from most of the state-of-the-art MODE variants which
have very complex structure, MODE-RMO has a very simple vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
uX
   u
2
structure; therefore, it can be easily implemented. t v∈P dðv; P Þ
IGD P ; P ¼  ð11Þ
jP j
4. Numerical simulations on test functions
where d(v,P) is the minimum Euclidean distance between v and points
This section presents the numerical experiments on test functions in P, |P*| is the number of points in P*. If |P*| is large enough to represent
performed in this work. First, we state ten test functions and two perfor- the Pareto front very well, IGD could measure both the diversity and
mance metrics for measuring the performance of the MOEA algorithms. convergence of P in a sense. As for both performance metrics, a lower
Next, to evaluate our proposed MODE-RMO, we compare it with two value means a better approximation set.
representative MOEAs: GDE3 [14] and NSGAII-DE [20]. Finally, the effec-
tiveness of ranking-based mutation operator in MODE is demonstrated 4.2. Comparison of MODE-RMO with other MOEAs
by comparing MODE-RMO with the algorithm without ranking-based
mutation operator. In order to know how competitive the proposed approach is, MODE-
RMO is compared with two MOEAs: the GDE3 [14] and NSGAII-DE [20],
4.1. Test problems and performance metric which are representative of the state-of-the art. MODE-RMO uses the
following parameter values: F = 0.5, CR = 0.3 [13], and the parameters
Ten test problems are used to evaluate the performance of our of GDE3 and NSGAII-DE keep the same as in the literature [14] and [20].
proposed MODE-RMO. The first five are bi-objective problems: The population size Np is set to 100 for all the three algorithms. The
ZDT1–ZDT4 and ZDT6, which are defined in [18]. The other five maximum generations Maxgen is set to be 250 for ZDT1-4 and ZDT6,
problems are composed of DTLZ1–DTLZ5 [21], which are more than 300 for DTLZ1 and DTLZ2, 500 for DTLZ3, 200 for DTLZ4 and DTLZ5,
two objectives. In our experiments, tri-objective problems are con- which are the same as those in their original literature [18] and [21]. It
sidered. The formulations of these ten problems are presented in is important to point out that the boundary-handling method has signif-
Appendix A. icant influence to the performance of the evolution algorithms. Since
Two performance metrics are employed to evaluate the attained the Pareto optimal solutions of some test problems (like ZDT1–ZDT4,
approximation sets of nondominated solutions in the objective ZDT6) lies on the boundaries, we use the re-initialization method for
space from the MOEAs. The first is generational distance (GD) [18]. all MOEAs in this paper, i.e., when one of the decision variables is be-
GD measures, in the objective function space, the average distance yond its boundary constraint, it is generated with the uniform distribu-
of the obtained approximation set of nondominated solutions P to tion within the boundary.
X. Chen et al. / Chemometrics and Intelligent Laboratory Systems 136 (2014) 85–96 89

Table 1
Comparison results of MODE-RMO and other MOEAs based on GD.

Algorithm MODE-RMO GDE3 NSGAII-DE

Mean SD Mean SD Mean SD

ZDT1 3.85E−03 3.22E−04 2.40E−03 5.62E−04 = 5.83E−03 1.22E−03 +


ZDT2 6.97E−03 7.25E−04 8.20E−03 2.33E−03 = 7.75E−03 5.99E−04 +
ZDT3 4.76E−03 5.45E−04 2.76E−03 9.22E−04 − 5.31E−03 9.23E−04 =
ZDT4 7.29E−02 1.63E−01 3.87E−02 3.77E−02 − 1.16E−02 2.15E−02 =
ZDT6 2.58E−02 1.12E−02 1.71E−01 5.23E−02 + 5.90E−02 4.59E−03 +
DTLZ1 2.58E−04 8.75E−06 7.08E−02 7.21E−02 + 4.80E−03 1.71E−02 +
DZTL2 7.28E−04 1.79E−05 7.25E−04 2.69E−05 = 1.96E−03 3.02E−04 +
DZTL3 7.88E−03 2.32E−02 3.27E + 00 4.87E + 00 + 1.82E−01 4.29E−01 +
DZTL4 7.26E−04 2.12E−05 7.12E−04 2.35E−05 = 1.46E−03 1.94E−04 +
DZTL5 9.12E−06 4.41E−07 1.09E−05 5.04E−06 = 1.87E−04 5.02E−05 +
w/t/l 3/5/2 8/2/0

To avoid randomness, each method is run 20 independently, and the From the above comparisons and analyses, we can say that, compared
statistics values including mean value (Mean) and standard deviations with other MOEAs, MODE-RMO can generate approximation Pareto sets
(SD) are presented in Tables 1–3. In addition, the paired Wilcoxon with comparable convergence and diversity in most of test problems;
signed-rank test is used to compare the significance between two algo- therefore, MODE-RMO can be used to deal with complex MOPs.
rithms. In Tables 1–3, “+”, “–” and “=” indicate MODE-RMO is respec-
tively better than, worse than, or similar to its competitor according to
the Wilcoxon signed-ranked test at significance level α = 0.05. The re- 4.3. Effectiveness of ranking-based mutation operator
sults are summarized as “w/t/l”, which denotes that MODE-RMO wins
on w functions, ties on t functions, and loses on l functions, compared In this section, we evaluate the effectiveness of our proposed
with its corresponding competitor. ranking-based mutation operator for MODE. For this purpose, MODE-
Table 1 shows the results of convergence metric GD using the algo- RMO is compared with MODE. The only difference between MODE-
rithms MODE-RMO, GDE3 and NSGAII-DE. From Table 1, MODE-RMO RMO and MODE is that in MODE three vector indices r1, r2, and r3 in
outperforms GDE3 in 3 problems, loses in 2 problems, and ties in 5 prob- the mutation are randomly selected. The GD and IGD values are com-
lems. ZDT1–ZDT4 are relative easy to solve, GDE3 performs better than pared as shown in Table 3.
MODE-RMO in these 4 problems in terms of GD. However, MODE-RMO For all the bi-objective problems (ZDT1–ZDT4, ZDT6), MODE-RMO
can obtain better GD values in other 6 more difficult problems: ZDT6, consistently obtains better results than MODE in terms of GD and IGD.
DTLZ1–DTLZ5, which means that the obtained approximate Pareto For the tri-objective problems, MODE-RMO achieves better GD value
fronts from MODE-RMO are closer to the optimal Pareto fronts than in DTLZ1 and better IGD value in DTLZ2. The results of MODE-RMO
those computed by GDE3 in these problems. Moreover, MODE-RMO ob- and MODE are similar in DTLZ3, DTLZ4, and DTLZ5. This is probably be-
tains smaller standard deviation values than GDE3 in all test problems cause the maximum generations are large enough for both algorithms
except ZDT4, which means that the stability of MODE-RMO is better to reach the approximation Pareto fronts in these 3 problems, which is
than GDE3. When compared with NSGAII-DE, MODE-RMO has an obvi- shown in the convergence graphs Figs. 2 and 3.
ous superiority, as it wins in 8 problems, and ties in 2 problems. In addi- The average convergence graphs of GD and IGD metrics in problems
tion, MODE-RMO obtains smaller standard deviation values than ZDT1–ZDT4, ZDT6, DTLZ1, DTLZ4, and DTLZ5 are plotted in Figs. 2 and 3.
NSGAII-DE in 7 problems. For ZDT1–ZDT4, ZDT6, MODE-RMO has converged to the satisfied ap-
The results of IGD obtained by three algorithms are presented in proximation Pareto fronts in the maximum generations, while MODE
Table 2. MODE-RMO achieves better IGD values than GDE3 in all prob- has not reached. For DTLZ1, DTLZ4, and DTLZ5, although both algo-
lems except ZDT3. This means that MODE-RMO provides better approx- rithms have reached the satisfied fronts, the convergence speed of
imation Pareto fronts when considering both convergence and diversity MODE-RMO is faster than MODE in both GD and IGD metrics.
in these 9 problems. In addition, the standard deviation values of Based on the results and analysis, we can see that the ranking-based
MODE-RMO are smaller than those of GDE3 in 8 problems, demonstrat- mutation operator is of benefit to the performance enhancement of
ing the stability of our MODE-RMO. MODE-RMO outperforms NSGAII- MODE. It not only accelerates the algorithms, but also makes them
DE in 7 problems, and ties in 3 problems. more efficient.

Table 2
Comparison results of MODE-RMO and other MOEAs based on IGD.

Algorithm MODE-RMO GDE3 NSGAII-DE

Mean SD Mean SD Mean SD

ZDT1 1.62E−03 1.11E−04 2.22E−03 6.61E−04 + 2.29E−03 1.16E−04 +


ZDT2 2.95E−03 2.78E−04 3.87E−03 1.12E−03 + 3.46E−03 2.80E−04 +
ZDT3 2.70E−03 1.65E−04 1.74E−03 5.27E−04 − 3.03E−03 2.68E−04 +
ZDT4 1.93E−03 3.47E−03 1.54E−02 1.42E−02 + 2.14E−03 2.35E−03 =
ZDT6 9.52E−03 7.25E−04 8.74E−02 2.48E−02 + 2.68E−02 1.40E−03 +
DTLZ1 4.00E−04 4.24E−04 8.06E−03 7.74E−03 + 4.14E−04 2.13E−05 +
DZTL2 1.01E−03 3.25E−05 1.06E−03 4.79E−05 + 1.06E−03 5.78E−05 +
DZTL3 1.78E−03 3.04E−03 1.59E−01 1.05E−01 + 5.93E−03 6.95E−03 +
DZTL4 1.03E−03 4.09E−05 1.07E−03 4.04E−05 + 1.04E−03 3.95E−05 =
DZTL5 9.72E−05 7.58E−06 1.02E−04 6.46E−06 + 9.73E−05 4.38E−06 =
w/t/l 9/0/1 7/3/0
90 X. Chen et al. / Chemometrics and Intelligent Laboratory Systems 136 (2014) 85–96

Table 3
Comparison results between MODE-RMO and MODE.

Algorithm GD IGD

MODE-RMO MODE MODE-RMO MODE

Mean SD Mean SD Mean SD Mean SD

ZDT1 3.85E−03 3.22E−04 5.32E−03 1.09E−03 + 1.62E−03 1.11E−04 2.14E−03 1.69E−04 +


ZDT2 6.97E−03 7.25E−04 9.81E−03 1.28E−03 + 2.95E−03 2.78E−04 4.13E−03 4.66E−04 +
ZDT3 4.76E−03 5.45E−04 6.37E−03 8.56E−04 + 2.70E−03 1.65E−04 3.49E−03 2.12E−04 +
ZDT4 7.29E−02 1.63E−01 1.98E−01 2.18E−01 + 1.93E−03 3.47E−03 9.86E−03 1.01E−02 +
ZDT6 2.58E−02 1.12E−02 3.02E−02 2.49E−03 + 9.52E−03 7.25E−04 1.36E−02 1.11E−03 +
DTLZ1 2.58E−04 8.75E−06 2.66E−04 1.43E−05 + 4.00E−04 4.24E−04 3.99E−04 1.42E−05 =
DZTL2 7.28E−04 1.79E−05 7.36E−04 2.01E−05 = 1.01E−03 3.25E−05 1.03E−03 4.56E−05 +
DZTL3 7.88E−03 2.32E−02 8.62E−03 2.37E−02 = 1.78E−03 3.04E−03 1.77E−03 2.96E−03 =
DZTL4 7.26E−04 2.12E−05 7.51E−04 4.21E−05 = 1.03E−03 4.09E−05 1.01E−03 2.86E−05 =
DZTL5 9.12E−06 4.41E−07 9.65E−06 6.62E−07 = 9.72E−05 7.58E−06 1.02E−04 7.05E−06 =
w/t/l 6/4/0 6/4/0

5. Applications to solve MDOPs interval is partitioned into several stage of equal length; and piece-
wise constant functions is employed to approximate the control
In this section, MODE-RMO is applied to solve three MDOPs trajectory.
of chemical processes taken from literature. In MDOP, each solution The major steps to optimize a MDOP using MODE-RMO are outlined
is a control trajectory in the time interval [t0 , tf ]; therefore, the as follows.
CVP approach is employed to deal with the dynamic problem. The
detail description of CVP can be found in [22]. In this study, we em- Step1: Divide the entire time interval [t0, tf] into N time intervals
ploy uniform discretization-based CVP approach, that is, the time [ti − 1, ti] (i = 1,…,N) with equal length.

MODE MODE -RMO

-0.5 0 -0.5

-0.5
-1 -1
log(GD)

log(GD)

log(GD)

-1
-1.5 -1.5
-1.5
-2 -2
-2

-2.5 -2.5 -2.5


0 50 100 150 200 250 0 50 100 150 200 250 0 50 100 150 200 250
(a) ZDT1: GD (b) ZDT2: GD (c) ZDT3: GD
1.5 0 2
1 1
-0.5
0.5 0
log(GD)

log(GD)

log(GD)

0 -1 -1
-0.5 -2
-1.5
-1 -3
-1.5 -2 -4
0 50 100 150 200 250 0 50 100 150 200 250 0 50 100 150 200 250 300
(d) ZDT4: GD (e) ZDT6: GD (f) DTLZ1: GD
-1 -1

-1.5 -2
log(GD)

log(GD)

-2 -3

-2.5 -4

-3 -5

-3.5 -6
0 50 100 150 200 0 50 100 150 200

(g) DTLZ4: GD (h) DTLZ5: GD


Fig. 2. Convergence graphs of GD metric.
X. Chen et al. / Chemometrics and Intelligent Laboratory Systems 136 (2014) 85–96 91

MODE MODE -RMO

-1 -0.5 -1

-1
log(IGD) -1.5 -1.5

log(IGD)

log(IGD)
-1.5
-2 -2
-2
-2.5 -2.5
-2.5

-3 -3 -3
0 50 100 150 200 250 0 50 100 150 200 250 0 50 100 150 200 250
(a) ZDT1: IGD (b) ZDT2: IGD (c) ZDT3: IGD
1 -0.5 0

0 -1 -1
log(IGD)

log(IGD)

log(IGD)
-1 -1.5 -2

-2 -2 -3

-3 -2.5 -4
0 50 100 150 200 250 0 50 100 150 200 250 0 50 100 150 200 250 300
(d) ZDT4: IGD (e) ZDT6: IGD (f) DTLZ1: IGD
-2
-2
-2.5
-2.2
log(IGD)

log(IGD)

-2.4 -3

-2.6 -3.5
-2.8
-4
-3
-3.2 -4.5
0 50 100 150 200 0 50 100 150 200

(g) DTLZ4: IGD (h) DTLZ5: IGD


Fig. 3. Convergence graphs of IGD metric.

Step2: Employ piece-wise constant functions to approximate the Step3: Suppose that the value of U is determined; then, employ an
control variables; and the parameters to be optimized are U = ordinary differential equation integrator to obtain the value of the
[u11,…,u1N, u21,…,u2N,…, um1,…,umN], where uij is the constant objective functions.
value of the ith variable in the jth time interval. Step4: Apply the MODE-RMO algorithm to optimize U; and obtain
the approximate Pareto optimal solutions.

5.1. Temperature control of a batch reactor

0.3 In the first problem, we consider the continuous stirred tank


T1
reactor (CSTR) [8,23], in which a first-order irreversible exothermic
J2: Concentration of A

0.25
T2 Table 4
The values of five Pareto optimal reaction temperatures.
T3
0.2 T1 T2 T3 T4 T5
T4
Stage 1 364.83188 377.50293 396.52125 396.85303 393.86480
T5
Stage 2 340.05753 371.92351 380.48428 396.09577 389.53583
Stage 3 340.48360 360.64328 372.04548 387.13968 393.57131
0.15 Stage 4 331.40026 357.63833 372.13378 381.34104 393.79417
Stage 5 334.67518 354.74714 352.38131 374.68806 392.54186
Stage 6 332.08104 342.68563 366.69167 370.36242 390.19363
Stage 7 334.29419 346.81884 365.01573 383.69559 379.30976
0.1 Stage 8 326.60700 347.10562 362.39021 375.83676 383.58584
-0.65 -0.6 -0.55 -0.5 -0.45 -0.4 -0.35 -0.3 -0.25 Stage 9 327.46827 347.00606 371.23060 374.44771 397.94964
J1: Concentration of B Stage 10 323.12137 349.49003 364.55216 378.92342 384.67188
J1 −0.60963 −0.55180 −0.44050 −0.33081 −0.24777
J2 0.29615 0.21757 0.17387 0.14728 0.13183
Fig. 4. Pareto optimal front for problem 1.
92 X. Chen et al. / Chemometrics and Intelligent Laboratory Systems 136 (2014) 85–96

400 400 400

Reaction Temperature

Reaction Temperature

Reaction Temperature
T1 T2 T3
380 380 380

360 360 360

340 340 340

320 320 320

300 300 300


0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
Time Time Time
(a)T1 (b) T2 (c) T3
400 400
Reaction Temperature

Reaction Temperature
T4
380 380

360 360 T5

340 340

320 320

300 300
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
Time Time
(d) T4 (e) T5

Fig. 5. Trajectories of Pareto optimal reaction temperature for problem 1.

k k
A→1
B→
2
C takes place. The process can be described by the following 5.2. Catalyst mixing problem in a tubular reactor
dynamic equations:
This problem considers a steady-state plug flow reactor, which
8 is packed with two catalysts [24,25]. These catalysts are required
>
> dC A 2 to stimulate a series of reactions (one reversible and one irreversible
>
> ¼ −k1 C A
>
> dt S1 ↔ S2 → S3). These assumptions give rise to the following model:
< dC
B 2
¼ k1 C A −k2 C B ð12Þ
>
> dt 8
>
> 298 ≤T ≤398; C A ð0Þ ¼ 1; C B ð0Þ ¼ 0; t f ¼ 1 > dx1
>
> >
> ¼ uðzÞ½10x2 ðzÞ−x1 ðzÞ
: 3 −2500=T 5 −5000=T >
< dt
k1 ¼ 4  10  e ; k2 ¼ 6:2  10  e
dx2 ð14Þ
>
> ¼ −uðzÞ½10x2 ðzÞ−x1 ðzÞ−½1−uðzÞx2 ðzÞ
>
> dt
: T
where C A and C B represent the dimensionless concentrations of A xð0Þ ¼ ½10 ; 0≤uðzÞ≤1; z f ¼ 12
and B, T is the temperature of the reactor. The objective is to obtain
the optimal reactor temperature, which simultaneously maximizes where x1 and x2 are the concentrations of S1 and S2, u is the fraction of
the concentration of B and minimizes the concentration of A: catalyst A, and z is the spatial coordinate. The objective is to determine
the optimal mixing policy of the two catalysts in order to maximize
8  
< min J 1 ¼ −C B t f
  ð13Þ
: min J ¼ C t :
2 A f
0.35

For this problem, the time interval is divided into 10 stages, the 0.3
population size is set as Np = 100, and maximum generation u1
Maxgen = 100. The attained Pareto optimal front is plotted in Fig. 4. 0.25
To assist decision making, we provide the values of five Pareto optimal
J2: -Catalyst use

reaction temperatures with equal distribution in Table 4, and the corre- 0.2
sponding Pareto optimal control trajectories are presented in Fig. 5.
u2
Babu and Angira only considered the single objective problem which 0.15
maximizes the concentration of B, and their best result is 0.610079 u3
[23]. From Table 4, our MODE-RMO obtained the value 0.60963 in 0.1 u4
boundary point T1, which is very close to 0.610079. The Pareto optimal u5
control trajectory of T1 is depicted in Fig. 5(a). When the complete focus
0.05
is put on minimizing the concentration of A, our result in boundary
point T5 is 0.13183, and the corresponding Pareto optimal control tra-
0
jectory of T5 is depicted in Fig. 5(e). The other three solutions T2, T3, -0.05 -0.04 -0.03 -0.02 -0.01 0
and T4 are intermediate points. From Fig. 5, it's obvious that increasing J1: Conversion
the reaction temperature in the entire time interval can reduce the con-
centration of A in the final time. Fig. 6. Pareto optimal front for problem 2.
X. Chen et al. / Chemometrics and Intelligent Laboratory Systems 136 (2014) 85–96 93

Table 5 5.3. Optimal operation of a fed-batch reactor


The values of five Pareto optimal mixing policy of the catalysts.

u1 u2 u3 u4 u5 In the third problem, we consider a batch reactor based on the fol-


lowing reaction system:
Stage 1 0.99963 0.64760 0.68575 0.26613 0.02095
Stage 2 0.47590 0.18442 0.00246 0.07997 0.00466
k1
Stage 3 0.22703 0.33236 0.01382 0.02580 0.00316 A þ B→C
Stage 4 0.24161 0.08481 0.01596 0.01041 0.02534 k2
Stage 5 0.18872 0.03159 0.01786 0.00133 0.00292 B þ B → D:
Stage 6 0.19141 0.00784 0.01008 0.00242 0.01166
Stage 7 0.30256 0.03179 0.00131 0.00261 0.00867 The process can be described by the following dynamic equations [8,
Stage 8 0.02394 0.00022 0.00220 0.00371 0.00075
26]
Stage 9 0.00085 0.01749 0.00166 0.00564 0.00307
Stage 10 0.00273 0.00052 0.00113 0.00180 0.00174 8
−0.04798 −0.04049 −0.03053 −0.01893 −0.00367 >
> d½A
J1 >
> dt ¼ −k1 ½A½B−ð½A=V Þu
J2 0.26544 0.13386 0.07522 0.03998 0.00829 >
>  
>
> bfeed −½B
> d½B
> 2
>
> ¼ −k1 ½A½B−2k2 ½B þ u
>
> dt V
>
>
the production S3 at the reactor outlet and minimize the amount of the < d ½C 
¼ k1 ½A½B−ð½C =V Þu ð16Þ
most expensive catalyst A: > dt
>
> d ½ D
>
> 2
¼ 2k2 ½B −ð½D=V Þu
8      >
>
>
> dt
>
< min J 1 ¼ − 1−x1 z f −x2 z f >
> d ½V 
Z z >
> ¼u
ð15Þ >
>
>
f >
: 0dt
: min J 2 ¼ uðzÞdz: ≤u≤0:01; t f ¼ 120 min
0

where [A], [B], [C], and [D] denote, respectively, the concentrations of A,
For this problem, the time interval is divided into 10 stages, the pop- B, C, and D, V is the current reaction volume, u is the reactant feed rate,
ulation size is set as Np = 100, and the obtained Pareto optimal front and k1 = 0.5 and k2 = 0.5 are the reaction rate constants. At the start of
after maximum generation of Maxgen = 100 is illustrated in Fig. 6. this reaction, the initial values is taken: [A](0) = 0.2 mol/l, [B](0) =
The trade-off between the use of catalyst A and the conversion to species 0 mol/l, and V(0) = 0.5. The objective is to maximize the amount of
S3 is clearly visible. The values of five Pareto optimal mixing policy of the product C and minimize the amount of by-product D
catalysts with equal distribution are presented in Table 5, and the corre-
8     
sponding Pareto optimal control trajectories are presented in Fig. 7. In < min J 1 ¼ − ½C  t f V t f
[25], Vassiliadis et al. considered the single objective problem to maxi-     ð17Þ
: min J ¼ ½D t V t :
mize the production S3, their best value is 0.0480238. The result of our 2 f f

MODE-RMO is 0.04798 in boundary point u1, very close to 0.0480238.


When we only consider minimizing the amount of the catalyst A, our re- For this problem, the time interval is divided into 10 stages, the pop-
sult in boundary point u5 is 0.00829, and the Pareto optimal trajectory of ulation size is set as Np = 100, and the obtained Pareto optimal front
u5 is illustrated in Fig. 7(e). Other three intermediate Pareto optimal after maximum generation of Maxgen = 100 is displayed in Fig. 8. The
control trajectories u2, u3, and u4 are illustrated in Fig. 7(b)–(d). From values of five Pareto optimal feed rate are presented in Table 6, and
Fig. 7, the trajectories of mixing policy of the catalysts in the entire inter- the corresponding Pareto optimal control trajectories are presented in
val gradually drop to 0 from u1 to u5, leading to the reduction of the use Fig. 9. When the complete focus is put on maximizing the amount of
of catalyst A. product C, the results of MODE-RMO in boundary point u1 is 0.07936,

1 u1 1 u2 1 u3
Fraction of catalyst

Fraction of catalyst
Fraction of catalyst

0.8 0.8 0.8

0.6 0.6 0.6

0.4 0.4 0.4

0.2 0.2 0.2

0 0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
Spatial coordinate Spatial coordinate Spatial coordinate
(a) u1 (b) u2 (c) u3
1 u4 1 u5
Fraction of catalyst

Fraction of catalyst

0.8 0.8
0.6 0.6

0.4 0.4

0.2 0.2

0 0
0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1
Spatial coordinate Spatial coordinate
(d) u4 (e) u5
Fig. 7. Trajectories of Pareto optimal mixing policy of the catalysts for problem 2.
94 X. Chen et al. / Chemometrics and Intelligent Laboratory Systems 136 (2014) 85–96

with complex constraints. The application of MODE-RMO for other


0.1
u1 complex real-world problems is also an interesting future work.
J2: -the amount of by-product D

0.08 Acknowledgments

Authors would like to express their sincere thanks to the reviewers


0.06 u2
for their valuable suggestions and comments. This work is supported
by the Major State Basic Research Development Program of China
0.04 u3 (2012CB720500), National Natural Science Foundation of China
(61333010, 21276078), National Science Fund for Outstanding Young
u4
u5 Scholars (61222303), Fundamental Research Funds for the Central Uni-
0.02
versities, Shanghai Rising-Star Program (13QH1401200), Program for
New Century Excellent Talents in University (NCET-10-0885) and
0 Shanghai R&D Platform Construction Program (13DZ2295300).

-0.09 -0.08 -0.07 -0.06 -0.05 -0.04 -0.03 -0.02 Appendix A. Test problems for multi-objective optimization
J1: the amount of product C
A.1. ZDT1
Fig. 8. Pareto optimal front for problem 3.

and the Pareto optimal feed rate is displayed in Fig. 9(a). When we only
consider minimizing the amount of by-product D, our result in bound-
 pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 9 X n
ary point u5 is 0.00111, and the Pareto optimal feed rate is displayed f 1 ðxÞ ¼ x1 ; f 2 ðxÞ ¼ gðxÞ 1− x1 =gðxÞ; g ðxÞ ¼ 1 þ x ; n ¼ 30;
in Fig. 9(e). Other three Pareto optimal feed rate u2, u3, and u4 are illus- n−1 i¼2 i
trated in Fig. 9(b)–(d). From Fig. 9, the feed rate in the entire time inter- 0≤ xi ≤1; i ¼ 1; 2; …; n:
val gradually reduces to the low bound from u1 to u5, which can reduce
the amount of by-product D.
The Pareto optimal solutions are: x1 ∈ [0, 1], xi = 0, i = 2, 3, …, n.
6. Conclusion
A.2. ZDT2
To deal with the challenge MOPs, in this paper, we propose a new
algorithm named MODE-RMO, in which ranking-based mutation oper-
ator is integrated into the original MODE algorithm to accelerate the
convergence rate. In the ranking-based mutation operator for multi-
  9 X n
objective optimization, the individual with better nondomination 2
f 1 ðxÞ ¼ x1 ; f 2 ðxÞ ¼ g ðxÞ 1−ðx1 =g ðxÞÞ ; g ðxÞ ¼ 1 þ x ; n ¼ 30;
front number and crowding distance has a larger probability to be se- n−1 i¼2 i
lected to take part in the mutation operator; therefore, it is beneficial 0 ≤xi ≤1; i ¼ 1; 2; …; n:
for the propagation of good information in the population to the off-
spring. The performance of MODE-RMO is evaluated by comparing
with other MOEAs in ten test problems, and the results demonstrate The Pareto optimal solutions are: x1 ∈ [0, 1], xi = 0, i = 2, 3, …, n.
that MODE-RMO can generate Pareto optimal fronts with good conver-
gence and diversity. Moreover, our experiments indicate that the A.3. ZDT3
ranking-based mutation operator can accelerate the convergence  
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
speed of MODE. Finally, MODE-RMO is applied to solve three MDOPs f 1 ðxÞ ¼ x1 ; f 2 ðxÞ ¼ gðxÞ 1− x1 =g ðxÞ− gxð1xÞ sinð10πx1 Þ ; g ðxÞ ¼
of chemical processes using the CVP approach. The obtained results in- n
dicate that MODE-RMO is an effective and efficient method for MDOPs. 1 þ n−1
9
∑ xi ; n ¼ 30; 0 ≤xi ≤1; i ¼ 1; 2; …; n:
i¼2
Parameter adaptive strategy has been proven to be efficient in im-
proving the performance of DE [27,28]; therefore, one possible future
The Pareto optimal solutions are: x1 ∈ [0, 1], xi = 0, i = 2, 3, …, n.
work is to employ this strategy to further improve the performance of
MODE-RMO. In addition, all the problems considered in this paper are
A.4. ZDT4
unconstrained, and we are planning to extend MODE-RMO for MOPs
 pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
¼ x1 ; f 2 ðxÞ ¼ g ðxÞ 1− x1 =g ðxÞ ; g ðxÞ ¼ 1 þ 10ðn−1Þþ
f 1 ðxÞ
Table 6
n  
The values of five Pareto optimal feed rate.
∑ x2i −10cosð4πxi Þ ; n ¼ 10; 0≤x1 ≤1; −5≤xi ≤5; i ¼ 2; 3; …; n:
i¼2
u1 u2 u3 u4 u5

Stage 1 0.00950 0.00991 0.00927 0.00431 0.00179


The Pareto optimal solutions are: x1 ∈ [0, 1], xi = 0, i = 2, 3, …, n.
Stage 2 0.00968 0.00573 0.00361 0.00193 0.00088
Stage 3 0.00938 0.00855 0.00579 0.00193 0.00059
Stage 4 0.00990 0.00801 0.00271 0.00308 0.00264 A.5. ZDT6
Stage 5 0.00892 0.00360 0.00253 0.00228 0.00080
Stage 6 0.00968 0.00755 0.00558 0.00344 0.00021
 
6
Stage 7 0.00953 0.00422 0.00279 0.00121 0.00062 f 1 ðxÞ ¼ 1−e−4x1 sin ð6πx1 Þ; f 2 ðxÞ ¼ g ðxÞ 1−ðx1 =g ðxÞÞ2 ; g ðxÞ ¼
Stage 8 0.00980 0.00309 0.00301 0.00079 0.00040 0:25
!
n
Stage 9 0.00721 0.00281 0.00145 0.00101 0.00014 1þ9 ∑ xi =ðn−1ÞÞ ; n ¼ 10; 0≤xi ≤1; i ¼ 1; 2; …; n:
Stage 10 0.00524 0.00616 0.00140 0.00268 0.00241 i¼2
J1 −0.07936 −0.07067 −0.05936 −0.04073 −0.01984
J2 0.09584 0.04714 0.02164 0.00628 0.00111
The Pareto optimal solutions are: x1 ∈ [0, 1], xi = 0, i = 2, 3, …, n.
X. Chen et al. / Chemometrics and Intelligent Laboratory Systems 136 (2014) 85–96 95

0.01 0.01 0.01


u2 u3
0.008 0.008 0.008

Feed Rate

Feed Rate

Feed Rate
0.006 0.006 0.006

0.004 0.004 0.004

0.002 0.002 0.002


u1
0 0 0
0 50 100 0 50 100 0 50 100
Time Time Time
(a) u1 (b) u2 (c) u3
0.01 0.01
u4 u5
0.008 0.008
Feed Rate

Feed Rate
0.006 0.006

0.004 0.004

0.002 0.002

0 0
0 50 100 0 50 100
Time Time

(d) u4 (e) u5
Fig. 9. Trajectories of Pareto optimal feed rate for problem 3.

A.6. DTLZ1 The Pareto optimal solutions are: x1 ∈ [0, 1], x2 ∈ [0, 1], xi = 0.5, i =
3, 4, …, n.
1 1 1
f 1 ðxÞ ¼ x x ð1 þ g ðxÞÞ; f 2 ðxÞ ¼ x1 ð1−x2 Þð1 þ gðxÞÞ; f 3 ðxÞ ¼ ð1−x1 Þð1 þ gðxÞÞ;
2 1 2 2 2 A.10. DTLZ5
h X n  i
2
gðxÞ ¼ 100 ðn−3 þ 1Þ þ ðxi −0:5Þ − cosð20π ðxi −0:5ÞÞ ;
i¼3
n ¼ 7; 0≤ xi ≤1; i ¼ 1; 2; …; n:

f 1 ðxÞ ¼ ð1 þ gðxÞÞ cosð0:5πθ1 Þ cosð0:5πθ2 Þ; f 2 ðxÞ ¼ ð1 þ gðxÞÞ cosð0:5πθ1 Þ sinð0:5πθ2 Þ;


The Pareto optimal solutions are: x1 ∈ [0, 1], x 2 ∈ [0, 1], xi =
Xn
ð1 þ 2x2 gðxÞÞ
0.5, i = 3, 4, …, n. f 3 ðxÞ ¼ ð1 þ gðxÞÞ sinð0:5πθ1 Þ; gðxÞ ¼
2
ðxi −0:5Þ ; θ1 ¼ x1 ; θ2 ¼
i¼3
2ð1 þ g ðxÞÞ
n ¼ 12; 0 ≤xi ≤ 1; i ¼ 1; 2; …; n:
A.7. DTLZ2

f 1 ðxÞ ¼ ð1 þ gðxÞÞ cosð0:5πx1 Þ cosð0:5πx2 Þ; f 2 ðxÞ ¼ ð1 þ gðxÞÞ cosð0:5πx1 Þ sinð0:5πx2 Þ;


The Pareto optimal solutions are: x1 ∈ [0, 1], x2 ∈ [0, 1], xi = 0.5, i =
Xn
2
3, 4, …, n.
f 3 ðxÞ ¼ ð1 þ gðxÞÞ sinð0:5πx1 Þ; gðxÞ ¼ ðxi −0:5Þ ; n ¼ 12; 0≤ xi ≤ 1; i ¼ 1; 2; …; n:
i¼3
References
The Pareto optimal solutions are: x1 ∈ [0, 1], x2 ∈ [0, 1], xi = 0.5, i = [1] F. Logista, M. Vallerioa, B. Houskab, M. Diehlb, J. Van Impea, Multi-objective optimal
3, 4, …, n. control of chemical processes using ACADO toolkit, Comput. Chem. Eng. 37 (10)
(2012) 191–199.
[2] K. Shimizu, A tutorial review on bioprocess systems engineering, Comput. Chem.
A.8. DTLZ3 Eng. 20 (6) (1996) 915–941.
[3] F.S. Wang, T.L. Shieh, Extension of iterative dynamic programming to multiobjective
optimal control problems, Ind. Eng. Chem. Res. 36 (6) (1997) 2279–2286.
[4] F. Logist, P.M.M. Van Erdeghem, J.F. Van Impe, Efficient deterministic multiple objec-
f 1 ðxÞ ¼ ð1 þ gðxÞÞ cosð0:5πx1 Þ cosð0:5πx2 Þ; f 2 ðxÞ ¼ ð1 þ gðxÞÞ cosð0:5πx1 Þ sinð0:5πx2 Þ;
# tive optimal control of (bio) chemical processes, Chem. Eng. Sci. 64 (11) (2009)
n  
f 3 ðxÞ ¼ ð1 þ gðxÞÞ sinð0:5πx1 Þ; g ðxÞ ¼ 100 ðn−3 þ 1Þ þ ∑ ðxi −0:5Þ2 − cosð20πðxi −0:5ÞÞ ; 2527–2538.
i¼3 [5] D. Sarkar, J.M. Modak, Pareto-optimal solutions for multi-objective optimization of
n ¼ 12; 0≤ xi ≤ 1; i ¼ 1; 2; …; n: fed-batch bioreactors using nondominated sorting genetic algorithm, Chem. Eng.
Sci. 60 (2) (2005) 481–492.
[6] Y. He, D. Chen, Immune mechanism based multi-objective ant colony algorithm ap-
The Pareto optimal solutions are: x1 ∈ [0, 1], x2 ∈ [0, 1], xi = 0.5, i = proach to batch reactor constrained dynamic multi-objective optimization prob-
3, 4, …, n. lems, J. Chem. Eng. Chin. Univ. 23 (2) (2009) 326–332.
[7] A.R. Mandli, J.M. Modak, Evolutionary algorithm for the determination of optimal
mode of bioreactor operation, Ind. Eng. Chem. Res. 51 (4) (2011) 1796–1808.
A.9. DTLZ4 [8] L. Jia, D. Cheng, M.S. Chiu, Pareto-optimal solutions based multi-objective particle
swarm optimization control for batch processes, Neural Comput. & Applic. 21 (6)
        (2012) 1107–1116.
f 1 ðxÞ ¼ ð1 þ gðxÞÞ cos 0:5πxα1 cos 0:5πxα2 ; f 2 ðxÞ ¼ ð1 þ g ðxÞÞ cos 0:5πxα1 sin 0:5πxα2 ;
[9] S. Wang, W. Du, X. Chen, F. Qian, Dynamic multi-objective optimization of chemical
  n
f 3 ðxÞ ¼ ð1 þ gðxÞÞ sin 0:5πxα1 ; gðxÞ ¼ ∑ ðxi −0:5Þ2 ; processes using modified Bare-bones MOPSO algorithm, The 3rd International
i¼3 Workshop on Advanced Computational Intelligence and Intelligent Informatics,
α ¼ 100; n ¼ 12; 0≤ xi ≤ 1; i ¼ 1; 2; …; n: 2013.
96 X. Chen et al. / Chemometrics and Intelligent Laboratory Systems 136 (2014) 85–96

[10] R. Storn, K. Price, Differential evolution—a simple and efficient heuristic for global [20] H. Li, Q. Zhang, Multiobjective optimization problems with complicated Pareto sets,
optimization over continuous spaces, J. Glob. Optim. 11 (4) (1997) 341–359. MOEA/D and NSGA-II, IEEE Trans. Evol. Comput. 13 (2) (2009) 284–302.
[11] H.A. Abbass, R. Sarker, C. Newton, PDE: a Pareto-frontier differential evolution ap- [21] K. Deb, L. Thiele, M. Laumanns, E. Zitzler, Scalable multi-objective optimization test
proach for multi-objective optimization problems, Proceedings of the 2001 Con- problems, Proceedings of the Congress on Evolutionary Computation (CEC-2002),
gress on, Evolutionary Computation, vol. 2, IEEE, 2001. (Honolulu, USA), 2002, pp. 825–830.
[12] N.K. Madavan, Multiobjective optimization using a Pareto differential evolution ap- [22] X. Chen, W. Du, H. Tianfield, R. Qi, W. He, F. Qian, Dynamic optimization of industrial
proach, Proceedings of the 2002 Congress on, Evolutionary Computation, vol. 2, processes with nonuniform discretization-based control vector parameterization,
IEEE, 2002, (CEC'02). IEEE Trans. Autom. Sci. Eng. (2014), http://dx.doi.org/10.1109/TASE.2013.2292582.
[13] T. Robic, B. Filipic, DEMO: differential evolution for multiobjective optimization, In Evolu- [23] B.V. Babu, R. Angira, Modified differential evolution (MDE) for optimization of non-
tionary Multi-Criterion Optimization (pp. 520–533), Springer, Berlin Heidelberg, 2005. linear chemical processes, Comput. Chem. Eng. 30 (6) (2006) 989–1002.
[14] S. Kukkonen, J. Lampinen, GDE3: the third evolution step of generalized differential [24] F. Logist, M. Vallerio, B. Houska, Multi-objective optimal control of chemical process-
evolution, The 2005 IEEE Congress on, Evolutionary Computation, vol. 1, IEEE, 2005. es using ACADO toolkit, Comput. Chem. Eng. 37 (10) (2012) 191–199.
[15] V.L. Huang, A.K. Qin, P.N. Suganthan, M.F. Tasgetiren, Multi-objective optimization [25] V.S. Vassiliadis, E.B. Canto, J.R. Banga, Second-order sensitivities of general dynamic
based on self-adaptive differential evolution algorithm, Congress on. IEEE, 2007, systems with application to optimal control problems, Chem. Eng. Sci. 54 (17)
Evolutionary ComputationIEEE, 2007, (CEC 2007). (1999) 3851–3860.
[16] L.V. Santana-Quintero, C.A.C. Coello, An algorithm based on differential evolution for [26] F. Herrera, J. Zhang, Optimal control of batch processes using particle swam optimi-
multi-objective problems, Int. J. Comput. Intell. Res. 1 (1) (2005) 151–169. sation with stacked neural network models, Comput. Chem. Eng. 33 (10) (2009)
[17] W. Gong, Z. Cai, Differential evolution with ranking-based mutation operators, IEEE 1593–1601.
Trans. Cybern. 43 (6) (2013) 2066–2081. [27] J. Brest, S. Greiner, B. Boskovic, M. Mernik, V. Zumer, Self-adapting control parame-
[18] K. Deb, A. Pratap, S. Agarwal, T. Meyarivan, A fast and elitist multiobjective genetic ters in differential evolution: a comparative study on numerical benchmark prob-
algorithm: NSGA-II, IEEE Trans. Evol. Comput. 6 (2) (2002) 182–197. lems, IEEE Trans. Evol. Comput. 10 (6) (2006) 646–657.
[19] W. Gong, Z. Cai, D. Liang, Engineering optimization by means of an improved [28] F. Qian, B. Xu, R. Qi, H. Tianfield, Self-adaptive differential evolution algorithm with
constrained differential evolution, Comput. Methods Appl. Mech. Eng. 268 (1) α-constrained-domination principle for constrained multi-objective optimization,
(2014) 884–904. Soft. Comput. 16 (8) (2012) 1353–1372.

Das könnte Ihnen auch gefallen