Abstract
The usage of broad sets of training data is paramount to evolve adaptive agents. In this respect, competitive co-evolution is a widespread technique in which the coexistence of different learning agents fosters adaptation, which in turn makes agents experience continuously varying environmental conditions. However, a major pitfall is related to the emergence of endless limit cycles where agents discover, forget and rediscover similar strategies during evolution. In this work, we investigate the use of competitive co-evolution for synthesizing progressively better solutions. Specifically, we introduce a set of methods to measure historical and global progress. We discuss the factors that facilitate genuine progress. Finally, we compare the efficacy of four qualitatively different algorithms, including two newly introduced methods. The selected algorithms promote genuine progress by creating an archive of opponents used to evaluate evolving individuals, generating archives that include high-performing and well-differentiated opponents, identifying and discarding variations that lead to local progress only (i.e., progress against the opponents experienced and retrogressing against others). The results obtained in a predator-prey scenario, commonly used to study competitive evolution, demonstrate that all the considered methods lead to global progress in the long term. However, the rate of progress and the ratio of progress versus retrogressions vary significantly among algorithms. In particular, our outcomes indicate that the Generalist method introduced in this work outperforms the other three considered methods and represents the only algorithm capable of producing global progress during evolution.