부산대학교 도서관

Multiobjective evolution and competitive coevolution are sub-fields of evolutionary computation which each significantly complicate the concept of fitness. Together, these make multiobjective competitive coevolution very difficult to work with by combining multiple objective values with non-absolute fitness, preventing the use of many established techniques for improving performance in multiobjective or competitive coevolutionary algorithms. Nonetheless, multiobjective scenarios arise frequently in competitive coevolution, such as whenever coevolving agents must consider costs for their actions. This paper proposes a new evaluation method of pairing opponents with similar skill levels in each objective, so that evaluations more efficiently distinguish the performance of similar individuals. This is enabled through the use of per-objective Elo ratings as a surrogate fitness function that prevents bias against individuals assigned stronger opponents. Ratings can further be assigned for asymmetric, non-zero-sum objectives such as cost, allowing individuals to be paired with opponents that incidentally challenge those asymmetric objectives. Mixed results are presented, showing significant benefits from pairing similar opponents, but finding that the use of Elo rating instead of raw fitness harms evolution. A novel statistical test for comparing multiobjective coevolutionary algorithms is also introduced.