A Pursuit-Evasion Algorithm Based on Hierarchical Reinforcement Learning

Abstract

This paper proposed a pursuit-evasion algorithm based on the Option method from hierarchical reinforcement learning and applied it into multi-robot pursuit-evasion game in 2D-Dynamic environment. The algorithm efficiency is studied by comparing it with Q-learning. We decompose the complex task with option method, and divide the learning process into two parts: High-level learning and Low-level learning, then design a new mechanism in order to make the learning process perform parallel. The simulation result shows the Option algorithm can efficiently reduce the complexity of pursuit-evasion task, avoid traditional reinforcement learning curse of dimensionality, and improve the learning result.