Age-Layered Strategies for Many-Objective Optimization

Abstract

Many-objective optimization problems (MaOPs) are multi-objective problems that have four or more objectives. MaOPs face significant challenges because of search inefficiency, computational cost, decision making, and visualization. Most MaOP systems use variants of non-dominated sorting (Pareto ranking). However, Pareto dominance is ineffective when the number of objectives exceeds four. In this research, we explore different strategies for solving MaOPs. We use Hornby’s Age-Layered Population Structure (ALPS) evolutionary algorithm in order to mitigate premature convergence and improve results. Instead of Pareto ranking, we use the many-objective evaluation strategy called sum of ranks (SR). SR is more appropriate than Pareto dominance for problems that require a majority of objectives to be optimized. We introduce and compare different objective reduction methods for ALPS, including random and correlated objective reduction. Because hypervolume and IGD performance measurements are not necessarily suitable to SR strategies, we introduce a new minimum distance measurement. Results show that different strategies are suitable for different problems, and depend strongly on the performance measure being used. Random objective reduction was the least effective strategy, while correlated reduction was more successful. The research shows that the ALPS framework with objective reduction is a promising framework for MaOPs.