PCA-assisted reproduction for continuous multi-objective optimization with complicated Pareto optimal set
Rui Wang,Nan-Jiang Dong,Dun-Wei Gong,Zhong-Bao Zhou,Shi Cheng,Guo-Hua Wu,Ling Wang
DOI: https://doi.org/10.1016/j.swevo.2020.100795
IF: 10.267
2021-02-01
Swarm and Evolutionary Computation
Abstract:<p>In terms of generating offspring solutions, simulated binary crossover (SBX) and differential evolution (DE) are two of the most representative reproduction operators in evolutionary multi-objective algorithms (EMOAs). However, they are found as less effective on multi-objective problems (MOPs) with complicated Pareto optimal set (PS). Under mild conditions, the PS of an MOP is an <span class="math"><math>(m−1)</math></span>-dimensional piecewise continuous manifolds where <span class="math"><math>m</math></span> is the number of objectives. Inspired from this regularity property, this study proposes a simple yet effective reproduction operator, namely, PCA-assisted reproduction (PCA-ar). Specifically, the PCA-ar first applies principal component analysis (PCA) method to construct a new decision space with reduced number of dimensions based on the information extracted from several well converged solutions. The PS is then estimated by a hyperplane in the new decision space. To this end, new offspring are sampled from the estimated PS, and then re-converted to the original decision space for fitness calculation. In order to systematically examine the effectiveness of the PCA-ar operator, we integrate it into NSGA-II and MOEA/D, and compare the derived algorithms, nNSGA-II and nMOEA/D, with their original versions (NSGA-II with SBX operator and MOEA/D with DE operator) as well as the regularity model based multi-objective estimated distribution algorithm (RM-MEDA) on the modified DTLZ benchmarks with up to 8 objectives. Experimental results show that nNSGA-II and nMOEA/D outperform the competitor EMOAs for most of problems, which indicate that the PCA-ar is effective. Lastly, the PCA-ar is also demonstrated to have good scalability to the number of decision variables.</p>
computer science, artificial intelligence, theory & methods