Data Driven Evasion Policy for Repeated Pursuit Evasion Problem with Unknown Imperfect Pursuer

A novel repeated pursuit-evasion (PE) problem with unknown imperfect pursuer is proposed and addressed using model based reinforcement learning. Grounded in aerospace applications, the problem adopts a repeated game formulation, where the evader engages repeatedly with the pursuer with unknown imperfect policy for multiple rounds, and the objective for the evader is to optimize the evasion policy based on data of preceding rounds so that evasion is achieved with minimum number of rounds. Unlike previous works that rely on prior knowledge of the pursuer's policy, our approach focus on efficient data exploitation in each round. Specifically, we adopt the probabilistic inference for learning control (PILCO) method, where a Gaussian process is adopted to model the unknown imperfect pursuer. This allows for efficient data exploitation and transforms the problem into the deterministic pay-off optimization with analytical gradients, enabling successive policy optimization in each round. Numerical results on a typical aerospace application demonstrate that the proposed method achieves successful evasion in remarkably few rounds, with the optimized policy aligning closely with the analytical optimal solution assuming complete knowledge of the pursuer.