International Journal of Automation and Computing  2018, Vol. 15 Issue (4): 462-473 PDF
Enhancing the Performance of JADE Using Two-phase Parameter Control Scheme and Its Application
Qin-Qin Fan1,2, Yi-Lian Zhang3, Xue-Feng Yan2, Zhi-Huan Wang1
1 Logistics Research Center, Shanghai Maritime University, Shanghai 201306, China;
2 Key Laboratory of Advanced Control and Optimization for Chemical Processes of Ministry of Education, East China University of Science and Technology, Shanghai 200237, China;
3 Key Laboratory of Marine Technology and Control Engineering Ministry of Communications, Shanghai Maritime University, Shanghai 201306, China
Key words: Differential evolution (DE) algorithm     evolutionary computation     dynamic optimization     control parameter adaptation     chemical processes
1 Introduction

Differential evolution (DE) algorithm, which was introduced by Storn and Price[1], is a simple yet competitive meta-heuristic algorithm. Although DE has been widely utilized to deal with a large number of benchmark test functions and industrial application problems, the search behavior of DE is mainly determined by three main parameters (i.e., mutation control parameter F, crossover control parameter CR, and population size NP) and two strategies (i.e., mutation and crossover)[2, 3]. To improve DE performance, researchers have proposed different empirical guidelines to set control parameters in the previous studies[46]. However, the above guidelines lack sufficient experimental justifications[7] since they are achieved through some particular experiments.

Besides the above studies, many methods have also been introduced to improve the search performance of the DE. For instance, in [17], an opposition-based learning is utilized to enhance the global search capability of the DE, the experimental results demonstrate that the approach is effective. Gong et al.[18] introduced a crossover rate repair method to enhance the performance of other self-adaptive DE algorithms. Recently, Gong et al.[19] introduced an adaptive ranking mutation operator (ARMOR) for the DE algorithm, which is used to deal with the constrained optimization problems introduced in CEC2006 and CEC2010. Yang et al.[20] introduced an auto-enhanced population diversity (AEPD) that is employed to improve the population diversity when the population is stagnant and premature. Fan et al.[21] introduced an auto-selection mechanism (ASM) to automatically choose a suitable algorithm from an algorithm pool. The results indicate that the ASM can improve the algorithm robustness. In [22], an eigenvector-based crossover strategy is used to improve the search ability of the DE. For more related studies, interested readers can refer to [2331].

Although various adaptation techniques can improve the performance of the DE effectively, most of them are based on the population information. In most cases, this information may be deceptive or misleading in complex optimization environment[32], thus a well-known empirical guideline is incorporated into the JADE (called TPC-JADE) to improve the search efficiency of DE in the current study. In the TPC-JADE, the parameter adaptation technique used in the JADE is employed to search the promising region in the early evolution stages, while the empirical guideline is utilized to enhance the search efficiency of the JADE in the later evolution stages. The main reason is that the fitness landscape of the current population may be not very complex in the later phases of the search. The performance of the proposed algorithm is compared with that of four DE variants on two widely used test suites, i.e., IEEE CEC2005[33] and IEEE CEC2015[34]. The experimental results reveal that the TPC-JADE is competitive among all competitors.

Several problems in chemical/biochemical processes can be described as a set of nonlinear differential equations, which are named as dynamic optimization problems. To solve these dynamic optimization problems, three main approaches[35, 36], which include dynamic programming (DP) methods, indirect optimization methods, and direct optimization methods, have been introduced.

For the DP, it is based on Bellman′s optimality conditions[37] and is a promising approach to solve dynamic optimization problems. However, DP is not always available for solving high-dimensional dynamic optimization problems. Subsequently, Luus[38] proposed an iterative dynamic programming (IDP) to overcome the defects of the DP, in which the coarse grid points and search region reduction strategy are utilized. The results indicate that the IDP is able to improve the solution accuracy and computation efficiency when compared with the DP.

In indirect methods, a dynamic optimization problem is converted into a boundary value problem (BVP), which can be addressed by some methods[36, 39, 40]. In fact, these approaches are very effective, but they are difficult to be used in solving dynamic optimization problems due to the active inequality constraints[41].

For direct methods, a dynamic optimization problem can be usually transcribed into a nonlinear programming (NLP) problem and is addressed by parameterization, which can be classified into two different techniques[42], namely, control vector parameterization (CVP) and complete parameterization. In the CVP, only control variables are discretized, whereas complete parameterization discretizes both control variable and state variable. In the current study, the CVP and the TPC-JADE are used to solve actual dynamic optimization problems. The results show that the TPC-JADE is competitive when compared with the other approaches.

2 Differential evolution

A minimization problem can be expressed as follows:

 $f({{{x}}^ * }) = \mathop {\min }\limits_{{x_i} \in \Omega } f({{{x}}_i}) ,\;{{{x}}_i} \in {{{P}}_0} = \prod\limits_{j = 1}^D {\left[ {{L_j},\;{U_j}} \right]}$ (1)

where f denotes the objective function; ${{{x}}_i} = [{x_{i,1}} \cdots\!\;{x_{i,D}}]$ and ${{{x}}^ * }$ are the D-dimensional vector and the global optimum solution of problem, respectively; ${L_j}$ and ${U_j}$ ( $j = 1, \cdots\!,\;D$ ) are the lower and upper bounds of the j-th variable of ${{{x}}_i}$ respectively; and ${{{P}}_0}$ is the feasible search space.

DE is a population-based stochastic optimization approach. ${{x}}_i^G = [x_{i,1}^G \cdots\!\;x_{i,D}^G]$ and ${{{X}}^G} = \{ x_1^G, \cdots\!,\;x_{NP}^G\}$ denote the i-th target vector and the population at the G-th generation, respectively.

The procedure of executing DE is described as follows[43]:

1) Initialization: Determine F, CR, NP, and maximum number of generations ${G_{\max }}$ . The initial individuals ${{x}}_i^{\rm{0}}( i = 1, \cdots\!,\;NP)$ are generated randomly in ${{{P}}_0}$ . The current generation G is set as 0.

2) Mutation: After initialization operation, for each ${{x}}_i^G$ , the mutant vector ${{v}}_i^G$ is produced by mutation operation. Some of the most frequently used mutation strategies are given as follows:

“DE/rand/1”

 ${{v}}_i^G = {{x}}_{{r_1}}^G + F \times ({{x}}_{{r_2}}^G - {{x}}_{{r_3}}^G)$ (2)

“DE/rand/2”

 ${{v}}_i^G = {{x}}_{{r_1}}^G + F \times ({{x}}_{{r_2}}^G - {{x}}_{{r_3}}^G) + F \times ({{x}}_{{r_4}}^G - {{x}}_{{r_5}}^G)$ (3)

“DE/current-to-best/1”

 ${{v}}_i^G = {{x}}_i^G + F \times ({{x}}_{{\rm{best}}}^G - {{x}}_i^G) + F \times ({{x}}_{{r_1}}^G - {{x}}_{{r_2}}^G)$ (4)

“DE/current-to-best/2”

 ${{v}}_i^G = {{x}}_i^G + F \times ({{x}}_{{\rm{best}}}^G - {{x}}_i^G) + F \times ({{x}}_{{r_1}}^G - {{x}}_{{r_2}}^G + {{x}}_{{r_3}}^G - {{x}}_{{r_4}}^G)$ (5)

“DE/rand-to-best/1”

 ${{v}}_i^G = {{x}}_{{r_1}}^G + F \times ({{x}}_{{\rm{best}}}^G - {{x}}_i^G) + F \times ({{x}}_{{r_2}}^G - {{x}}_{{r_3}}^G)$ (6)

where ${r_1}$ , ${r_2}$ , ${r_3}$ , ${r_4}$ and ${r_5}$ are mutually exclusive integers randomly chosen within the range [1, NP], and are also different from the index i (i.e., ${r_1} \ne {r_2} \ne {r_3} \ne {r_4} \ne {r_5} \ne i$ ); ${{x}}_{{\rm{best}}}^G$ is the target vector with the best fitness value at the G-th generation.

3) Crossover: For each ${{x}}_i^G$ , trial vector ${{u}}_i^G$ is generated as follows:

 u_{ij}^G = \left\{ \begin{aligned}& v_{ij}^G,\;\;\;\;{\rm {if}}\;\;{R_j} \le CR \;\; {\rm or} \;\; j = {j_{rand}}\\& x_{ij}^G ,\;\;\;\;{\rm{otherwise}}\end{aligned} \right.\;\;\;\;j = 1,2,\cdots\!,\;D (7)

where ${R_j}$ is a uniform random number within the range [0, 1] and ${j_{rand}}$ is a randomly chosen integer within the range [1, D], respectively.

4) Selection: The trial vector ${{u}}_i^G$ competes with its target vector ${{x}}_i^G$ and the better one of them is selected for the next generation:

 {{x}}_i^{G + 1} = \left\{ \begin{aligned}& {{u}}_i^G,\;\;\;\;\;\;{\rm {if}}\;\;f({{u}}_i^G) \le f({{x}}_i^G)\\& {{x}}_i^G,\;\;\;\;\;\;{\rm{otherwise}}.\end{aligned} \right. (8)

5) G = G + 1.

6) Implement Steps 2 to 5 repeatedly until the number of generations is equal to Gmax.

3 JADE using two-phase parameter control scheme

For most evolutionary algorithms, it is generally believed to be a good idea to encourage the global search in the early stages of the evolution and ensure the local search in the later evolution phases[44], thus the above view can be considered as the empirical guideline. For the DE algorithm, a large value of F can provide good exploration capability, whereas a small value of F can accelerate the convergence speed[3]. Meanwhile, a large value of CR can provide good exploitation capability[45]. In the current study, this empirical guideline can be used to improve the search efficiency of the DE in the later evolution stages. Namely, the value of F gradually decreases and the value of CR gradually increases during the entire evolutionary process.

The adaptation technique can enable control parameters to make timely adjustment when compared with the empirical guideline. In other words, the adaptability of the control parameters produced by the adaptation approach is better than that of the control parameters generated by the empirical guideline. Additionally, the fitness landscape of the population is usually very complex in the early stages of the search since individuals in the current population may be distributed in the entire search space, but individuals may gather in a small search region in the later phases of the evolution. Based on the above introduction, the parameters of JADE are adjusted through two different approaches in the current study. Namely, the parameter adaptation method proposed in JADE[10] is used in the early evolution stages and the empirical guideline is utilized to guide the evolution of control parameters in the later phases of the evolution. The main target is to take advantages of the two different parameter control approaches.

3.1 Evolution of control parameters

If G < $gs \times {G_{\max }}$ , the F value of each individual is produced in [10]; otherwise, the F value of each individual is generated as follows:

 $\begin{split} F_i^{G + 1} = \;& Cauchy \left(\left(0.6 - 0.5 \times \frac{{G - gs \times {G_{\max }}}}{{(1 - gs) \times {G_{\max }}}}\right),\;\sigma_1\right)\\& i = 1, \cdots\!,\;NP\end{split}$ (9)

where Cauchy is a Cauchy distribution function, ${\sigma _1} = 0.6$ , gs = 0.6. It can be seen from (9) that TPC-JADE has good local search ability in the later search stages since Cauchy distribution function with a low location parameter tends to generate a small value of F (i.e., promoting exploitation). Moreover, the location parameter of Cauchy distribution function is within the range of [0.1, 0.6].

If G < $gs \times {G_{\max }}$ , the CR value of each individual is generated in [10]; otherwise, the CR value of each individual is generated as follows:

 $\begin{split}CR_i^{G + 1} =\; & N \left(\left(1 - 0.5 \times \left(1 - \frac{{G - gs \times {G_{\max }}}}{{(1 - gs) \times {G_{\max }}}}\right)\right),\;\sigma_2\right)\\ & i = 1, \cdots\!,\;NP\end{split}$ (10)

where N is a normal distribution function, ${\sigma _2} = 0.6$ . We can observe from (10) that a large CR value generated by a normal distribution function with large mean value can speed up convergence in the later phases of the search. Moreover, the mean value of normal distribution function is within the range of [0.5, 1].

Based on the above descriptions, it can be observed that the empirical guideline can provide a good exploitation capability for the DE algorithm. Therefore, it is useful to improve the search efficiency of DE in the later search process. Additionally, the parameter adaptation approach used in the early evolution stages can help to locate the promising search region. This is because the adaptation technique based on the population information can automatically adjust the control parameters when the fitness landscape of the population is very complex.

1) Initialization

Determine the population size NP and maximal number of generation Gmax. Initialize a population ${{P}}_1^0$ that is generated randomly within the feasible search space ${{{P}}_0}$ . Set the current generation G = 0, gs = 0.6.

2) Population evolution

Mutation operation: For each ${{x}}_i^G$ , a mutant vector ${{v}}_i^G$ is generated as follows[10]:

 ${{v}}_i^G = {{x}}_i^G + {F_i} \times ({{x}}_{pbest}^G - {{x}}_i^G) + {F_i} \times ({{x}}_{{r_1}}^G - {{x}}_{{r_2}}^G)$ (11)

where ${{x}}_{pbest}^G$ is randomly selected from the best 100p% individuals in the current population, and p is within the range (0, 1].

Crossover operation: Equation (7) is used to generate a trial vector ${{u}}_i^G.$

The more detailed descriptions can be seen in Section 3.1.

4) G = G + 1.

5) Steps 2 to 4 are repeated until the maximum number of generations is equal to Gmax.

The framework of the proposed algorithm is shown in Fig. 1.

4 Experimental results

In this section, two famous test suites (i.e., IEEE CEC2005 and IEEE CEC2015) are utilized to assess the performance of the TPC-JADE. Moreover, the performance of the TPC-JADE is compared with that of four state-of-the-art DE variants, namely, jDE[9], SaDE[7], JADE[10], and CoBiDE[13]. All the compared algorithms are coded in Matlab (Matlab R2012a) and run on Windows 7 operating system (64 bit). For each test function, the maximum numbers of fitness evaluations are set to 300 000 for 30-dimensional functions and 500 000 for 50-dimensional functions. Since each DE variant has its own suitable NP, recommended NP from their literatures are utilized, namely, 100 for jDE, JADE, and TPC-JADE, 50 for SaDE, and 60 for CoBiDE. Furthermore, to compare the performance of all the algorithms, the Wilcoxon′s rank sum test[46] at the 0.05 significance level and the Friedman′s test[47] are utilized. For the Wilcoxon′s rank sum test, the “+”, “−” and “≈” signs denote that the search performance of the proposed algorithm is significantly better than, worse than, and almost similar to that of the other compared algorithms in a statistically significant way, respectively.

4.1 Comparison with four DE variants on 30-dimensional CEC2005 functions

Based on the above comparisons, it can be observed from Table 1 that the average performance of TPC-JADE is the best among five compared DE variants, the adaptation technique can assist DE to find optimal/near search region in the early evolution stages, while the empirical guideline can enhance the local search capability of DE at later phases of the search.

Table 1
Results of five DE variants on 30-dimensional CEC2005 test functions

4.2 Comparison with four DE variants on 50-dimensional CEC2005 functions

Based on the above analyses, the statistical analysis results shown in Table 2 indicate that the overall performance of TPC-JADE is better than that of other competitors on twenty-five 50-dimensional CEC2005 functions.

Table 2
Results of five DE variants on 50-dimensional CEC2005 test functions

4.3 Comparison with four DE variants on 30-dimensional CEC2015 functions

In this experiment, fifteen 30-dimensional functions introduced in IEEE CEC2015 are employed to demonstrate the performance of TPC-JADE. For each function, the run times are set to be 30, and the mean and standard deviation values are presented in Table 3. Additionally, the statistical analysis results obtained by the Wilcoxon′s rank sum test are also shown in Table 3.

According to the results summarized in Table 3, it can be observed that the performance of TPC-JADE is significantly better than that of jDE, SaDE, JADE, and CoBiDE on ten, twelve, nine, and seven functions, respectively. Note that jDE, SaDE, and JADE cannot provide significantly better results than TPC-JADE on any functions. We can find that the control parameters guided by the empirical guideline can improve the search efficiency of JADE. It can be also seen from Table 3 that CoBiDE significantly outperforms TPC-JADE on four functions. This is because the covariance matrix learning is used in the DE. Overall, TPC-JADE provides the best average performance among all compared algorithms on 15 30-dimensional IEEE CEC2015 functions.

Table 3
Results of all compared algorithms on 15 30-dimensional CEC2015 test functions

5 Parameter analysis

In this part, 25 30-dimensional CEC2005 test functions are utilized to investigate the effect of σ, which is selected from the set (0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8). The maximum number of function evaluations and run times are set to be 300 000 and 30, respectively.

The statistical analysis results achieved by the Friedman′s test are plotted in Fig. 2. As shown in Fig. 2, we can find that the overall performance of TPC-JADE is susceptible to the value of sigma and is the best when σ = 0.6. Therefore, σ = 0.6 is chosen in our proposed algorithm.

 Download: larger image Fig. 2. Rankings obtained by Friedman′s test under different values of sigma

6 Case study

A typical dynamic optimization problem can be formulated as follows[39, 48]:

 $\begin{split}& \min \;\;\;J({{u}}(t)) = G(x({t_{{f}}})) + \int_{{t_0}}^{{t_{\rm{f}}}} {F({{x}},{{u}}){\rm d}t} \\& {\rm s.t.}\;\;\;\frac{{{\rm d}{{x}}(t)}}{{{\rm d}t}} = f({{x}}(t),{{u}}(t)),\;\;{{x}}(0) = {{{x}}_0}\\&\;\;\;\;\;\;\;\; 0 \le t \le {t_{{f}}},\;{{{u}}_{{\rm{min}}}} < {{u}} < {{{u}}_{{\rm{max}}}}\end{split}$ (12)

where J denotes the performance index; x and u are the state and control vectors, respectively; G and F are the terminal time performance index and integrated performance, respectively; t0 is the initial time and tf is the final time; umin and umax are the lower and upper bounds of u, respectively.

6.1 Feeding-rate optimization for Lee-Ramirez bioreactor

In this section, TPC-JADE is utilized to solve a dynamic optimization problem proposed by Lee and Ramirez[49]. Additionally, the direct method adopted in [50] is employed. This dynamic optimization problem can be described as follows:

 $\begin{array}{l} \frac{{{\rm{d}}{x_1}}}{{{\rm{d}}t}} = {u_1} + {u_2},\;\;\;\;\;\;\;\;\;\;\;\;{x_1}(0) = 1\\ \frac{{{\rm{d}}{x_2}}}{{{\rm{d}}t}} = \mu {x_2} - \frac{{{u_1} + {u_2}}}{{{x_1}}}{x_2},\;\;\;\;{x_2}(0) = 0.1\\ \frac{{{\rm{d}}{x_3}}}{{{\rm{d}}t}} = \frac{{{u_1}}}{{{x_1}}}{C_{nf}} - \frac{{{u_1} + {u_2}}}{{{x_1}}}{x_3} - {Y^{ - 1}}\mu {x_2},\;\;\;{x_3}(0) = 40\\ \frac{{{\rm{d}}{x_4}}}{{{\rm{d}}t}} = {R_{fp}}{x_2} - \frac{{{u_1} + {u_2}}}{{{x_1}}}{x_4},\;\;\;\;{x_4}(0) = 0\\ \frac{{{\rm{d}}{x_5}}}{{{\rm{d}}t}} = \frac{{{u_2}}}{{{x_1}}}{C_{if}} - \frac{{{u_1} + {u_2}}}{{{x_1}}}{x_5},\;\;\;\;{x_5}(0) = 0\\ \frac{{{\rm{d}}{x_6}}}{{{\rm{d}}t}} = - {k_1}{x_6},\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;{x_6}(0) = 1\\ \frac{{{\rm{d}}{x_7}}}{{{\rm{d}}t}} = {k_2}(1 - {x_7}),\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;\;{x_7}(0) = 0\\ 0 \le {u_1},\;{u_2} \le 0.01 \end{array}$ (13)

where ${x_1}$ (L), ${x_2}$ (g/L), ${x_3}$ (g/L), ${x_4}$ (g/L) and ${x_5}$ (g/L) are the reaction volume, cell density, nutrient concentration, foreign protein concentration, and inducer concentration, respectively; ${x_6}$ is inducer shock factor and ${x_7}$ is inducer recovery factor on the cell growth rate; ${u_1}$ (L/h) is the glucose feed rate, and ${u_2}$ (L/h) is the inducer feed rate; Y, ${C_{{{nf}}}}$ (g/h), ${C_{{{if}}}}$ (g/h), $\mu$ (1/h), and ${R_{{{fp}}}}$ (1/h) are the growth yield coefficient, nutrient concentration in the nutrient feed, inducer concentration in the inducer feed, specific growth rate, and foreign protein production rate, respectively; ${k_1}$ is the shock parameter and ${k_2}$ is recovery parameter.

In (13), some parameters can be defined as follows:

 $\mu = \frac{{0.407{x_3}}}{{\displaystyle\frac{{0.108 + {x_3} + x_3^2}}{{14\;814.8}}}}\left( {{x_6} + \frac{{0.22}}{{0.22 + {x_3}}}{x_7}} \right)$
 ${k_1} = {k_2} = \frac{{0.09{x_5}}}{{0.034 + {x_5}}}$
 ${R_{fp}} = \frac{{0.095{x_3}}}{{\displaystyle\frac{{0.108 + {x_3} + x_3^2}}{{14\;814.8}}}}\left( {\frac{{0.005 + {x_5}}}{{0.022 + {x_5}}}} \right)$
 ${C_{{{if}}}} = 4, \; {C_{{{nf}}}} = 100, \; {{Y = 0}}{\rm{.51}}.$

For the above fed-batch production of induced foreign protein, the objective is to maximize the economic benefit of this fed-batch system. Therefore, the objective function can be defined as

 ${\rm{max }}J{\rm{(}}{u_1}{\rm{,}}{u_2}{\rm{) = }}{x_1}({t_{{f}}}){x_4}({t_{{f}}}) - Q\int_0^{{t_{\rm{f}}}} {{u_2}(t)} {\rm{d}}t$ (14)

where Q = 5 and tf = 10 h.

For this dynamic optimization problem, researchers have introduced various methods to obtain satisfactory results, which are 0.814 9[51], 0.816 7[52], 0.815 8[53] and 0.816 4[54].

In this experiment, NP and ${G_{\max }}$ of TPC-JADE are set to be 100 and D×10 000, respectively. Moreover, three different discrete time degrees (i.e., D) are discussed as follows.

1) The value 0.816 43 is achieved by TPC-JADE when D is set to be 10. It can be seen that our result is better than the results achieved by Roubos et al.[51], Zhang et al.[53], and Fan et al.[54] respectively. However, the result of TPC-JADE is slightly worse than the result obtained by Sarkar and Modak[52]. The main reason may be that Sarkar and Modak[52] consume much more computational resources and the discrete time length is smaller than our proposal. In addition, the curves of two control variables (i.e., u1 and u2) are plotted in Figs. 3 and 4, respectively.

 Download: larger image Fig. 3. Optimal control profile of glucose feed rate (D = 10)

 Download: larger image Fig. 4. Optimal control profile of inducer feed rate (D = 10)

2) Let D = 20, the value 0.816 47 is achieved by using TPC-JADE. Clearly, the result achieved by TPC-JADE is better than the other competitors except for the result obtained by Sarkar and Modak[52]. This is because Sarkar and Modak consume more computational resources when compared with the proposed algorithm. Moreover, the obtained result is better than the result obtained by TPC-JADE when D = 10. It means that the solution precision is directly influenced by the discrete time degree. Additionally, the curves of the optimal glucose feed rate and inducer feed rate are presented in Figs. 5 and 6, respectively.

 Download: larger image Fig. 5. Optimal control profile of glucose feed rate (D = 20)

 Download: larger image Fig. 6. Optimal control profile of inducer feed rate (D = 20)

3) In this experiment, D is set to be 30. The value 0.816 48 is achieved by the proposed algorithm. Although the result is slightly worse than Sarkar and Modak[52], the obtained result is better than the results reported by Roubos et al.[51], Zhang et al.[53] and Fan et al.[54], respectively. The obtained result is also better than the results achieved by TPC-JADE when D = 20. It can be observed from the obtained result that TPC-JADE is a very competitive optimization tool for solving a complex dynamic optimization problem. The curves of optimal u1 and u2 are shown in Figs. 7 and 8, respectively.

 Download: larger image Fig. 7. Optimal control profile of glucose feed rate (D = 30)

 Download: larger image Fig. 8. Optimal control profile of inducer feed rate (D = 30)

Based on the above simulation results, it can be concluded that the solution precision is highly influenced by the dimensionality of the problem and the TPC-JADE can perform well when the dimensionality of the problem is increasing. It should be noted that the value 0.816 7 is achieved by Sarkar and Modak[52] when the number of function evaluations is more than 2 000 000. For our proposal, the maximum number of function evaluations is 300 000; therefore, the proposed algorithm is very competitive for solving this dynamic optimization problem. Additionally, from Figs. 38, it can be observed that the nutrient can be neglected when Q = 5. This is because glucose feed rate is maintained at zero during the entire period of operation.

7 Conclusions