To solve the game problem of two agents in three-dimensional space, this paper proposes a deep reinforcement learning algorithm based on Experience Enhancement, by introducing the matrix game as the control decision of agents as the game opponent, in the interaction of agents to form the experience playback pool with the nature of the game, and then through the sampling and learning of neural networks, so that the agents that have been learnt will have the ability to play the game in three-dimensional space. Then through the sampling and learning of neural networks, the learned agents are equipped with the ability to play the game in three-dimensional space. In this paper, based on the basic DQN algorithm framework, we designed the Experience Enhancement mechanism with the nature of a matrix game, and after the agent training and confrontation simulation experiments, we verified that the proposed DQN-Experience Enhancement algorithm preliminarily solves the problem of convergence and generalisation ability of the DQN algorithm in solving the problem of the two agents game in the three-dimensional space. The problem of poor convergence and generalisation ability.