Research On The Pursuit-Evasion Problem With Multi-robot Based On Reinforcement Learning

The pursuit-evasion problem is a kind of problem widely focused, a typical problem for Multi-Agent collaboration and cooperation strategy research in dynamic environment. It includes real-time processing knowledge, wireless communication, multi-robot cooperation and coordination and dynamic path planning and so on. This paper focused on the multi-robot cooperation and coordination based on pursuit-evasion problem. It comprises four predator agents whose goal is to capture a prey agent by surrounding it on four sides.Reinforcement learning is an unsupervised and online learning method. It improves its behaviours by trial and error with environment. The empirical knowledge not required in reinforcement learning. Therefore, reinforcement learning is a real-time and online learning method. The typical reinforcement learning methods include Q-learning proposed by Watkins and Temporal Difference method proposed by Sutton.The complexity degree of reinforcement learning increases exponently with the increase of agent number. To avoid the so-called'curses of dimensionality', some methods have been put forward, HRL (hierarchical reinforcement learning) was among of them. The famous HRL methods include HAM, MAXQ and Option. Because of its flexibility and its simplicity, Option is widely applied into multi-robot system. This paper adopted Option method to deal with pursuit-evasion problem. Simulation results show Option method is better than Q-learning in training time and pursuit quality.