Improved Bayesian inverse reinforcement learning based on demonstration and feedback

A major obstacle to traditional reinforcement learning is that rewards need to be artificially set to have a strong subjectivity. The inverse reinforcement learning algorithm solves this problem. Traditional inverse reinforcement learning requires an optimised demonstration, which is often not met in reality. Therefore, an interactive learning method was proposed to enhance the learned reward function by combining the feedback with the demonstration and using the improved Bayesian rule iteration of the imagery to improve the Agent strategy. The proposed method was tested in experimental and simulation tasks. The results showed that the efficiency of the method was significantly improved under different degrees of non-optimal proof.