Co-evolution improves the efficiency of preference learning methods when the Decision Maker's aspirations develop over time
Michał K. Tomczyk, Miłosz Kadziński · Proceedings of the Genetic and Evolutionary Computation Conference · 2023
This paper's research scope is interactive evolutionary multiple objective optimization founded on the preference learning paradigm. It concerns a scenario in which the Decision Maker's (DM's) aspirations develop over time. In this view, the interactive method may be forced to occasionally re-learn the DM's value system and, thus, re-orient the search during optimization. In preliminary studies, we observed that although a satisfactory recommendation can ultimately be discovered, it is often attainable with more significant computational power. To resolve this issue, we propose a co-evolutionary method, evolving sub-populations that approximate the Pareto front or align with the DM's preferences. This diversifies the maintained solution set, which is useful when re-understanding the DM's aspirations during interactions. Also, it helps reallocate the preference-driven sub-population quickly, avoiding an extensive computational burden. We demonstrate the usefulness of such hybridization in a series of extensive experiments that involve different test problems, numbers of objectives, and behavioral models of the Decision Maker.