Proposal of a Realistic Many-Objective Test Suite

Weiyu Chen, Hisao Ishibuchi, Ke Shang

Abstract

Real-world many-objective optimization problems are not always available as easy-to-use test problems. For example, their true Pareto fronts are usually unknown, and they are not scalable with respect to the number of objectives or decision variables. Thus, the performance of existing multi-objective evolutionary algorithms is always evaluated using artificial test problems. However, there exist some differences between frequently-used many-objective test problems and real-world problems. In this paper, first we clearly point out one difference with respect to the effect of changing the value of each decision variable on the location of the objective vector in the objective space. Next, to make artificial test problems more realistic, we introduce a coefficient matrix to their problem structure. Then, we demonstrate that a different coefficient matrix leads to a different difficulty of a test problem. Finally, based on these observations, we propose a realistic many-objective test suite using various coefficient matrices.
BibTeX
@inproceedings{chen2020proposal,
  author = {Weiyu Chen and Hisao Ishibuchi and Ke Shang},
  title = {{Proposal of a Realistic Many-Objective Test Suite}},
  booktitle = {PPSN},
  year = {2020},
  doi = {10.1007/978-3-030-58112-1\_14},
}

Loading related content...

Figure