Escape with Self-adaptive Decision Radius based on Deep Deterministic Policy Gradient in Pursuit Games

The Pursuit-Evasion (PE) game of Unmanned Surface Vehicles (USVs) is a classic antagonistic problem for the intelligent agent system. To enhance the escaping success rate of evaders with better effort, this paper proposes an escape strategy based on the geometrical characteristics of Apollonius circles. An improved self-adaptive escaping strategy for the evader utilizing the deep deterministic policy gradient algorithm is proposed. Then, the criteria for successful encirclement by pursuer are given. A DDPG algorithm-based framework is proposed on the basis of markov decision process formulation. Specifically, an action space based on adaptive decision radius of evader is designed. Our simulation shows the proposed method has more advantages in terms of escape distance and escape time.