Efficient Reinforcement Learning Through Evolving Neural Network Topologies

摘要: Neuroevolution is currently the strongest method on pole-balancing benchmark reinforcement learning tasks. Although earlier studies suggested that there was an advantage in evolving network topology as well connection weights, leading neuroevolution systems evolve fixed networks. Whether structure can improve performance open question. In this article, we introduce such a system, NeuroEvolution of Augmenting Topologies (NEAT). We show when evolved (1) with principled crossover, (2) by protecting structural innovation, and (3) through incremental growth from minimal structure, significantly faster stronger than best fixed-topology methods. NEAT also shows it possible to populations increasingly large genomes, achieving highly complex solutions would otherwise be difficult optimize.