Prediction of Reward Functions for Deep Reinforcement Learning via Gaussian Process Regression

Inverse reinforcement learning (IRL) is a technique for automatic reward acquisition, however, it is difficult to apply to high-dimensional problems with unknown dynamics. This article proposes an efficient way to solve the IRL problem based on the sparse Gaussian process (GP) prediction with l-1-regularization only using a highly limited number of expert demonstrations. A GP model is proposed to be trained to predict a reward function using trajectory-reward pair data generated by deep reinforcement learning with different reward functions. The trained GP successfully predicts the reward functions of human experts from their collected demonstration trajectory datasets. To demonstrate our approach, the proposed approach is applied to the obstacle avoidance navigation of the mobile robot. The experimental results clearly show that the robots can clone the experts' optimality in navigation trajectories avoiding obstacles using only with a very small number of expert demonstration datasets (e.g., \leq 6). Therefore, the proposed approach shows great potential to be applied to complex real-world applications in an expert data-efficient manner.