Competitive coevolution through evolutionary complexification

摘要: Two major goals in machine learning are the discovery and improvement of solutions to complex problems. In this paper, we argue that complexification, i.e. incremental elaboration through adding new structure, achieves both these goals. We demonstrate power complexification NeuroEvolution Augmenting Topologies (NEAT) method, which evolves increasingly neural network architectures. NEAT is applied an open-ended coevolutionary robot duel domain where controllers compete head head. Because supports a wide range strategies, because coevolution benefits from escalating arms race, it serves as suitable testbed for studying complexification. When compared evolution networks with fixed complexifying discovers significantly more sophisticated strategies. The results suggest order discover improve solutions, evolution, search general, should be allowed complexify well optimize.