Hierarchical Reinforcement Learning With Multi Discount Factors In A Differential Game

Reinforcement learning (RL) is an approach to solving differential games, especially pursuit-evasion games. Reinforcement learning, treats the model and the environment as black boxes. It is shown that using multi-agent reinforcement learning is a suitable tool to address pursuit-evasion games. This paper proposes a new approach based on hierarchical reinforcement learning to find an instantaneous reward function for the game of guarding a territory, which is a more complex version of pursuit-evasion games. To have greater control over the generated path, a novel modification to the fuzzy actor-critic learning algorithm is proposed, so the augmented algorithm is able to handle several reward functions with different time horizons. The proposed approach has shown to have a suitable control over the combination of a terminal reward function and an instantaneous reward.