Guiding Deep Reinforcement Learning by Modelling Expert’s Planning Behavior

Deep Reinforcement Learning algorithm has shown its strength in solving complex robot control problems. Nevertheless, the existing challenge is that random and unnecessary exploration may lead to slow convergence, which causes difficulty in the training process. In this paper, we propose a novel framework of using a small amount of demonstration to guide the exploration. Unlike previous works, we model the demonstration as a function between direction preferences and goal-state distance using Neural Network. To evaluate our work, we designed several tasks in the robot simulation platform VREP. The results indicate that our algorithm can guide the exploration and hence improve the performance of Deep Deterministic Policy Gradient(DDPG).