Solving complex sequential decision-making problems by deep reinforcement learning with heuristic rules

Deep reinforcement learning (RL) has demonstrated great capabilities in dealing with sequential decision-making problems, but its performance is often bounded by suboptimal solutions in many complex applications. This paper proposes the use of human expertise to increase the performance of deep RL methods. Human domain knowledge is characterized by heuristic rules and they are utilized adaptively to alter either the reward signals or environment states during the learning process of deep RL. This prevents deep RL methods from being trapped in local optimal solutions and computationally expensive training process and thus allowing them to maximize their performance when carrying out designated tasks. The proposed approach is experimented with a video game developed using the Arcade Learning Environment. With the extra information provided at the right time by human experts via heuristic rules, deep RL methods show greater performance compared with circumstances where human knowledge is not used. This implies that our approach of utilizing human expertise for deep RL has helped to increase the performance of deep RL and it has a great potential to be generalized and applied to solve complex real-world decision-making problems efficiently.