Integration of Evolutionary Computing and Reinforcement Learning for Robotic Imitation Learning

This paper proposes an evolutionary reinforcement learning method by combining Estimation of Distribution Algorithm and Reinforcement Learning. The Reinforcement Learning method in our method is based on Policy Improvement with Path Integrals (PI2). Estimation of Distribution Algorithm is incorporated into this reinforcement learning method to improve the generation of roll outs with certain noises. This method can accelerate the converging of the learning results and improve the overall system performance. Additionally, this method provides a potential solution to integrate the exploratory evolutionary algorithms and the greedy policy learning method. The proposed method is applied in a robotic imitation learning experiment in this paper and the experimental results demonstrate the effectiveness and robustness of our proposed algorithm.