부산대학교 도서관

In this paper, a problem of multi-agent pursuit-evasion in which multiple pursuers try to round up a single evader as soon as possible in a 2D limited space is considered. A cooperative approach of pursuit strategy through sensing among pursuits is developed under the multi-agent deep deterministic policy gradient (MADDPG) reinforcement learning (RL) method, which adopts the centralized training and distributed execution to deal with the pursuit problem by a fully distributed approach. Instead of using the communication information, each pursuer is supposed that can only obtain the sensing information of its neighbors and the evader. By introducing the sensing range segment and relative-position average strategy, we allow that the number of the pursuer can be changeable, which is different from some existing results that assume that the number of pursuers is fixed. To demonstrate the feasibility of the proposed method above, we implement it in a simulation environment. Simulation results show that the pursuer can learn a highly cooperative control strategy and capture the evader with a high success rate.