Elitism, Sharing, and Ranking Choices in Evolutionary Multi-criterion Optimisation

Elitism and sharing are two mechanisms that are believed to improve the performance of an evolutionary multi-criterion optimiser. The relative performance of the two most popular ranking strategies is largely unknown. Using a new empirical inquiry framework, this report studies the effect of elitism, sharing, and ranking design choices using a benchmark suite of two-criterion problems. Performance is assessed, via known metrics, in terms of both closeness to the true Pareto-optimal front and diversity across the front. Randomisation methods are employed to determine significant differences in performance. Informative visualisation of results is achieved using the attainment surface concept. Elitism is found to offer a consistent improvement in terms of both closeness and diversity, thus confirming results from other studies. Sharing can be beneficial, but can also prove surprisingly ineffective. Evidence presented herein suggests that parameter-less schemes are more robust than their parameter-based equivalents (including those with automatic tuning). Very little performance difference is evident between the two ranking strategies. A multi-objective genetic algorithm combining both elitism and parameter-less sharing is shown to offer very good performance across the test suite.