Investigating Relaxed Selection in Test-Based Pareto Coevolution

In previous studies, we proposed four relaxed selections schemes for test-based Pareto coevolution as implemented by a variant of the Population based Pareto Hill Climber (P-PH C- P) Three of them outperformed the default selection used in P-PHC-P in which a parent is only replaced in the next generation by its child if the latter Pareto-dominates the former. While the results were particularly encouraging, more work is needed to fully understand the reasons behind this improved performance. In this work, we therefore extend previous results by revisiting the relaxed selection methods from the perspective of both the distribution of candidate solutions in different Pareto layers, and the concept of hyper volume commonly used in the evolutionary multi-objectives optimization literature. Extensive experimental analysis shows that relaxed selection (Upward-Horizontal Selection) improves convergence while maintaining diversity in the converging population, better than base selection.