An Empirical Analysis of Fitness Assignment and Diversity-Preserving in Evolutionary Multi-Objective Optimization

,


Introduction
Many real-world optimization problems often involve multiple conflicting objectives, namely multi-objective optimization problems (David A. et al., 2000).Without loss of generality, a single-objective optimization problem has a particular global optimum.For a multi-objective optimization problem, there exists a set of solutions.These solutions are also called Pareto-optimal solutions or trade-off solutions (David A. et al., 2000).Without further knowledge, they are identically important.Usually, it is more difficult to solve a multi-objective optimization problem than to solve a single-objective optimization problem.This is because the search space of a multi-objective optimization problem is often more complex and larger than that of a single-objective optimization problem.Compared with the traditional methods, evolutionary algorithms (EAs) can exhibit more advantages such as gaining a solution set in a single simulation run and often being less susceptible to the characteristics of problems when applied to solve multi-objective optimization problems (Carlos M. et al., 1995;E. Mezura-Montes, et al., 2008).Currently, EAs have been widely applied to solve multi-objective optimization problems from scientific community and engineering fields (C. A. Coello Coello, 2006).
The aim of designing EAs is to gain a well-converged and well-distributed set approximation to the true Pareto-optimal front of problem in a single simulation run (C. A. Coello Coello, 2006).Hence, it requires to improve the convergence speed and perverse the diversity of solutions during the search of EAs.Accordingly, EAs employ corresponding fitness assignment approaches (E.Zitzler, et al., 1999;2002) and diversity-preserving mechanisms (J.Horn, et al., 1994;N. Srinivas, et al., 1995;J. Horn, et al., 1993;C. Igel, et al., 2007;D. E. Goldberg, et al., 1987) to achieve their two intentions, respectively.In EAs, the fitness assignment approaches are utilized to guide the search of EAs towards optimal regions, and the representatives of these approaches involve Goldberg's Pareto ranking (D.E. Goldberg, 1989) and Fonseca and Fleming's Pareto ranking (C.M. Fonseca, et al., 1993); while diversity-preserving mechanisms are utilized to enrich the population and avoid the premature convergence, and in the diversity-preserving mechanisms, the effective fitness sharing techniques can improve the diversity of solutions.This paper first introduces two alternative fitness assignment approaches, i.e., Goldberg's Pareto ranking and Fonseca and Fleming's Pareto ranking, and develops three alternative pruning techniques.In order to implement these pruning techniques and prune excess individuals more effectively, a fast and effective density-estimating technique without using parameters is employed (K.Deb, 2002).Then, through combining these approaches and techniques, several differential evolution (R. Storn, et al., 1997) algorithms for multi-objective optimization are given.Numerical experiments are conducted on a test set involving seven scalable multi-objective optimization problems to be minimized.The experimental results indicate the relative performances between two alternative Pareto ranking approaches and among three alternative pruning techniques, respectively.Finally, in order to further enrich the diversity of solutions, a dynamic mutation operator (Z.Michalewicz, 1996) is introduced.This paper is organized as follows: First, we introduce two alternative fitness assignment approaches based on Pareto ranking in Section 2.Then, we develop three pruning techniques in Section 3 and introduce a fast and effective technique for estimating the individual crowded density in Section 4. Next, we present the experimental results, compare different fitness assignment approaches, and contrast different pruning techniques in Section 5. Finally, some conclusions and future work are given in Section 6.

Goldberg's Pareto Ranking
Based on the concept of Pareto optimality (David A., 2000), Goldberg (D. E. Goldberg, 1989) first proposed the basic idea of Pareto ranking to determine the individual rank for each individual in the current population.According to these ranks and a certain selection strategy, some individuals with better ranks have the relatively bigger probabilities to be selected into the next generation population.The procedure of Goldberg's Pareto ranking is described as follows (D.E. Goldberg, 1989): First, the current population is sorted by the concept of Pareto optimality, some non-dominated individuals are obtained and assigned rank 1, then remove these individuals from the current population and the remaining individuals constitute the current population; Repeatedly, first, the current population is sorted by the concept of Pareto optimality, some non-dominated individuals are obtained and assigned rank 2, then remove these individuals from the current population and the remaining individuals constitute the current population; Continue to the above steps till the current population becomes empty when all individuals have certain ranks.The basic idea of Pareto ranking can be illustrated in Figure 1.

Fonseca and Fleming's Pareto Ranking
Based on the concept of Pareto optimality, Fonseca and Fleming (Fonseca & Fleming, 1993)   , then remove ( The basic idea of the second pruning technique is illustrated in Figure 4; 3) We calculate the individual density for each individual from the union of the th i subpopulation ) (t P i n and the intermediate population . The basic idea of the third pruning technique is illustrated in Figure 5.

Estimating the Individual Crowded Density
In order to maintain the diversity of solutions more effectively, some individuals which are located in the more crowded regions will be removed when the number of individuals in the population is larger than a predefined constant value.The crowding-distance estimating approach provides a fast and effective measuring technique for estimating the individual crowded density for each individual in the population.The basic idea of estimating the individual crowding-distance is described as follows (K.Deb, et al., 2002): First, along the th K dimensional direction in the objective space, the th K distance between the individual i x and the individual j x from the population P is calculated as x near to two individuals located at each side of the individual i x is calculated, where


. In order to further enrich the diversity of solutions and retain the individuals located at the extreme positions, if ( max ) (

Benchmark Test Functions and Performance Metrics
In order to gain a specific and important insight into existing approaches, through combining these approaches, we give several differential evolution algorithms for multi-objective optimization.Numerical experiments are conducted on a test set of multi-objective optimization problems (K.Deb, et al., 2001;S. Huband, et al., 2006).These multi-objective optimization problems are scalable and to be minimized.They have been widely used to valid the performances of evolutionary algorithms for multi-objective optimization.The details of these multi-objective optimization problems are described in Appendix A.
In order to measure the performances including convergence, diversity as well as convergence and diversity of obtained solution set, let Z and * PF be the obtained solution set and the true Pareto-optimal front (typically, * PF is produced by uniformly sampling), then convergence metric  , diversity metric  as well as convergence and diversity metric IGD can be correspondingly defined in the following.
Obviously, the smaller the value of  . Storn, and K. Price, 1997;T. Okabe, et al., 2003) Let the diversity metric  be used to measure the distribution of diversity of Z and the spread along the true Pareto-optimal front * PF .This metric can be calculated by  Definition 3 (convergence and diversity metric IGD ) (T. Okabe, et al., 2003;Q. Zhang, et al., 2008) Let the convergence and diversity metric IGD be used to measure the average distance from * PF to Z .This

Investigating Two Alternative Pareto Ranking Approaches
In this study, we first investigate the relative contributions of two alternative Pareto ranking approaches to the quality of obtained solutions.According to Goldberg's Pareto ranking and Fonseca and Fleming's Pareto ranking, we give two corresponding differential evolution algorithms called PR and SP.The parameter settings for all problems (Set A) are listed in Table 1.Experimental Parameters utilized by differential evolution algorithms are given in Table 2. Table 3 gives the convergence results over 20 runs; Table 4 gives the diversity results over 20 runs; Table 5 gives the convergence and diversity results over 20 runs.
According to Table 1, we give the corresponding Pareto-optimal set (PS) and Pareto-optimal front (PF) for each problem in Figure 7 to Figure 13.For DTLZ4, it is not easy to sample the true Pareto-optimal set and the true Pareto-optimal front.Figure 10 might only stand for the part of the true Pareto-optimal set and the true Pareto-optimal front of DTLZ4.
In order to more intuitively observe the experimental results, we present the obtained Pareto-optimal front for each problem in terms of two alternative differential evolution algorithms PR and SP in Figure 14 to Figure 20.According to Figure 14 to Figure 20, we can observe that SP and PR can find an approximation set to the true Pareto-optimal front for each problem above in a single simulation run.
According to Table 3, we can find that for all problems, the best results obtained by SP are slightly better than those obtained by PR in terms of convergence.For DTLZ1, DTLZ5, and DTLZ6, the mean results obtained by SP are slightly better than those obtained by PR, while for DTLZ2, and DTLZ7, the mean results obtained by SP are slightly worse than those obtained by PR.It is interesting to note that for DTLZ3, SP can find an approximation set for each run out of 20 runs.Unfortunately, PR failed to find an approximation set for each run out of 20 runs.At the same time, for most of problems DTLZ1, DTLZ2, DTLZ5, and DTLZ6, the worst results obtained by SP are slightly worse than those obtained by PR, while for DTLZ3 and DTLZ7, the worst results obtained by SP are slightly superior to those obtained by PR.It signifies that SP outperforms PR when applied to solve more complex problems.
According to Table 4, we can observe that for all problems, SP and PR can obtain the approximate mean, best, worst, std (standard deviation) results.It is worthy of noting that for DTLZ3, although PR failed to find an approximation set for each run out of 20 runs, the obtained approximation set keeps the better diversity of solutions.
According to Table 5, for DTLZ1, DTLZ3, DTLZ5, and DTLZ7, the mean results produced by SP are slightly superior to those produced by PR, while for other problems, the mean results produced by SP are worse than those produced by PR.For DTLZ1-DTLZ5, we can find that the best results produced by SP are slightly worse than those produced by PR, while for other problems, the best results produced by SP are slightly better than those produced by PR.For DTLZ1~ DTLZ2, and DTLZ4~ DTLZ7, the worst results obtained by SP are approximate to those obtained by PR.In general, the mean performance of SP is superior to that of PR.

Investigating Three Alternative Pruning Techniques
In this section, we continue to further investigate the relative contributions of three alternative pruning techniques to the quality of obtained solutions.According to the first pruning technique, the second pruning technique, and the third pruning technique, we create three corresponding differential evolution algorithms called DPA, DPB, and DPC.They all employ Fonseca and Fleming's Pareto ranking approach.The parameter settings for all problems (Set B) are listed in Table 6.Experimental Parameters utilized by DPA, DPB, and DPC are given in Table 7. Table 8 gives the convergence results over 10 runs; Table 9 gives the diversity results over 10 runs; Table 10 gives the convergence and diversity results over 10 runs.
In the process of experimental simulation, we find that for two problems DTLZ1 and DTLZ3, each algorithm DPA, DPB, or DPC has a certain possibility of getting struck in local Pareto-optimal fronts in a single simulation run.In order to more intuitively observe the shapes of local Pareto-optimal fronts, Figure 21 illustrates two different local Pareto-optimal fronts obtained in two independent simulation runs with respect to DTLZ1; Figure 22 illustrates two different local Pareto-optimal fronts obtained in two independent simulation runs with respect to DTLZ3.Seen from Figure 21 and Figure 22, these local Pareto-optimal fronts still exhibit better diversity and profile.It indicates that diversity-preserving mechanisms employed in these algorithms are effective.
According to Table 8, for four problems DTLZ2, DTLZ5, DTLZ6, and DTLZ7, the mean and best results obtained by three algorithms DPA, DPB, and DPC are approximate.For DTLZ1, The mean performance of DPA is superior to those of DPB and DPC, the best and std (standard deviation) results obtained by these algorithms are slightly different, and the worst result of DPB is the worst among three algorithms.For DTLZ3, three algorithms DPA, DPB, and DPC failed to find an approximation set for each run out of 10 runs, and they perform approximately.
According to Table 9, for each problem, each algorithm can obtain the approximate performances with respect to the mean, best, worst, and std experimental results.
According to Table 10, for most of problems DTLZ1, DTLZ2, DTL3, DTLZ5, and DTLZ7, the mean performance of DPC is superior to that of DPA or DPB.The initial results indicate that the third pruning technique is more effective than the first pruning technique, and the second pruning technique.

Improving the Search Abilities by Incorporating Other Search Techniques
In this section, in order to improve the search capability and avoid the premature convergence, we introduce a dynamic mutation operator, which can be described as follows (Storn and Price, 1997): , where N is the population size.We call this modified version H-DPC of DPC.
The parameter settings for two problems DTLZ1 and DTLZ3 are referred to Table 6; Experimental parameters for two differential evolution algorithms DPC and H-DPC are referred to Table 7. Finally, Table 11 gives the convergence results over 10 runs with respect to DTLZ1 and DTLZ3; Table 12 gives the diversity results over 10 runs with respect to DTLZ1 and DTLZ3; Table 13 gives the convergence and diversity results over 10 runs with respect to DTLZ1 and DTLZ3.
For DTLZ1 and DTLZ3, Table 11 shows that the convergence performances of H-DPC are obviously superior to those of DPC; Table 12 shows that the diversity performances of H-DPC are approximate to those of DPC; Table 13 shows that the convergence and diversity performances of H-DPC are obviously better than those of DPC.It indicates that an effective search operator can improve the performance of an evolutionary algorithm.

Conclusions and Future Work
This paper introduces two alternative fitness assignment approaches based on Pareto ranking to guide the search of EAs towards optimal regions, three alternative diversity-preserving pruning techniques (i.e., three alternative fitness sharing techniques), and incorporates a dynamic mutation operator into differential evolution algorithms in order to enrich the diversity of solutions to premature convergence.Through combining these approaches and/or techniques, we present several specific differential evolution algorithms for multi-objective optimization, and compare and contrast these algorithms.
As to future work, some research directions will be considered as follows: On the one hand, during the search, how to identify better individuals more effectively is still our research work.As shown in Figure 23, the individual (white circle) is easier to be rejected than any other individuals (back circles) in the population.In fact, the individual is a potential solution.On the other hand, we will also further introduce more effective fitness assignment approaches to improve the convergence speed, and more effective fitness sharing techniques and diversity-preserving mechanisms to enrich the diversity of solutions.   .The total number of variables is of the individual i x in the population P is obtained, where M is the number of objectives, max objective value and the minimum objective value along the th K dimensional direction in the objective space, respectively.The basic idea of estimating the individual crowding-distance is illustrated in Figure6.
measure the closeness of the obtained solution set Z to the true Pareto-optimal front * PF .This metric can be calculated by closer the distance from Z to * PF is, and the value of ) , ( * PF Z  is equal to zero, if and only if distance between two points x and y .A smaller value of ) -distributed diversity of Z along the true Pareto-optimal front * PF .If the obtained solution set Z is well distributed and involves the extreme points in * the minimum Euclidean distance from the point z in * PF to all points in Z .If | | * PF is large enough to represent the true Pareto-optimal front, measure the convergence and diversity of the obtained solution set Z .A smaller value of

,
and the objective function values lie on the linear hyper-plane    proposed another alternative Pareto ranking.According to the basic idea of Fonseca and Fleming's Pareto ranking, the individual

Table 1 .
Parameter settings for all problems (Set A)

Table 3 .
Experimental results (over 20 runs) with respect to convergence  (Set A)

Table 8 .
Experimental results (over 10 runs) with respect to convergence  (Set B)

Table 9 .
Experimental results (over 10 runs) with respect to diversity  (Set B)

Table 11 .
Experimental results (over 10 runs) with respect to convergence  (for DTLZ1 and DTLZ3 in Set B)

Table 13 .
Experimental results (over 10 runs) with respect to convergence and diversity IGD (for DTLZ1 and