Apollonius partitions based pursuit-evasion strategies via multi-agent reinforcement learning

Abstract

This paper considers a multi-agent pursuit-evasion game in which no less than two pursuers cooperate to capture a sensible evader. In contrast, the slower evader takes its strategy to extend life as long as possible. Since the classical Apollonius circle is effective in generating a capture strategy for multiple pursuers around a single evader scenario, we use Apollonius circle method to divide the entire state space into their respective dominant regions. The overlapping area of the Apollonius circles is assigned as the dominant region of the slower evader. In addition, we design a potential game model for this game, which guarantees the existence of the Nash equilibrium solution. In this paper, the evader’s strategy is derived using the dominant region determined by current global states while pursuers’ strategies are generated by the designed algorithm, where pursuers cooperate to shrink the evader’s dominant area to achieve capture during the motion. Finally, three simulations of different scenarios based on the QMIX algorithm demonstrate the effectiveness of the proposed method.