Evolutionary reinforcement learning via cooperative coevolutionary negatively correlated search

Evolutionary algorithms (EAs) have been successfully applied to optimize the policies for Reinforcement Learning (RL) tasks due their exploration ability. The recently proposed Negatively Correlated Search (NCS) provides a distinct parallel search behavior and is expected facilitate RL more effectively. Considering that commonly adopted neural usually involves millions of parameters be optimized, direct application NCS may face great challenge large-scale space. To address this issue, paper presents an NCS-friendly Cooperative Coevolution (CC) framework scale-up while largely preserving its behavior. issue traditional CC can deteriorate also discussed. Empirical studies on 10 popular Atari games show method significantly outperform three state-of-the-art deep methods with 50% less computational time by effectively exploring 1.7 million-dimensional