A method based on multi-agent deep reinforcement learning algorithm is proposed for the multiple satellites and non-cooperative targets in the orbital pursuit-evasion game， which has complex dynamics model， unknown maneuver information of non-cooperative targets， and challenge to coordinate effectively among satellites. Firstly， the game scenario is modeled， the reward function is reshaped and improved under the scenarios of minimum time， optimal fuel and collision avoidance. The multi-agent deep deterministic policy gradient algorithm is used for centralized training to obtain the optimal pursuit policy parameters of each pursuing satellite and evading satellite. Then the distributed execution enables multiple pursuing satellites and evading satellites to complete the pursuit-evasion game. The simulation results show that the method can complete the pursuit-evasion game of multiple satellites against non-cooperative targets and use the numerical advantage to effectively increase the success rate of pursuit and reduce the energy consumption in the pursuit process. Moreover， a series of intelligent game behaviors such as “interception”， “siege”， “infiltration”， and “capture” emerge， which are conducive to effectively achieve the game purpose.