Pursuit and evasion game between UVAs based on multi-agent reinforcement learning

Pursuit and evasion game between UVAs is a typical differential game. Differential games are usually difficult to obtain the optimal solutions because of the complex bilateral extremum problems. Reinforcement learning has superiorities in solving differential games with the advantages such as it does not need accurate controlled models and a lot of training data. In this paper, a multi-agent reinforcement learning model is established for UAV pursuit and evasion game. The relative motion state equation is used to describe the state to simplify the state set, and the pursuit and evasion game is transformed into a zero-sum game which is solved by Minimax-Q learning. The reinforcement learning model established in this paper reduces the complexity of solving problem and guarantees the convergence speed. Finally, the simulation results verify the rationality of the obtained control policy which makes both the pursuer and the evader tend to be advantageous to their own direction in the course of the countermeasures.