On/off-policy Hybrid Deep Reinforcement Learning and Simulation in Control Tasks

Abstract

Recently, deep reinforcement learning with neural network shows great performance in tasks such as game AI and robotics control tasks. However, on-policy and off-policy reinforcement learning methods proposed in related works have problems such as slow exploration speed. To solve these problems, we propose a hybrid deep reinforcement learning method which combines on-policy and off-policy reinforcement learning in this paper. The comparison experiment shows that the proposed method outperforms classic DDPG and DPPO method with an obvious advantage.