Self-play training for general impulsive orbital game

Abstract In this paper, the decision and training architecture based on intelligent methods is provided for a more general impulsive orbital game problem. Then, the impulsive maneuver intervals are incorporated into the game action set and the reinforcement learning method is utilized to make decisions. Based on prioritized fictitious self-play, a method to rapidly generate an opponent population is proposed to train the agent to defeat opponents with various strategies. Finally, the evaluation matrix is decomposed by dimensionality reduction and the transitivity and cyclicity between strategies are analyzed, demonstrating that the agent can continuously improve strategy even in the face of an opponent population with increasingly rich strategies.