Dynamic Gain Military Game Algorithm Based on Episodic Memory
Jun He, Jun Jie Yang · 2021 International Conference on Computer Engineering and Application (ICCEA) · 2021
In recent years, the development of deep reinforcement learning has become a new research frontier in the field of artificial intelligence. In the complex and changeable battlefield environment of military confrontation, how to use reinforcement learning methods to assist in formulating a set of efficient military game strategies has become a new research direction. The military game confrontation environment usually uses static reward settings ignoring the importance of captured target in the tactical arrangement. This paper starts from the use of episodic memory Q-network model to train drone swarm confrontation and uses expert knowledge to design a local state target threat assessment method. By analyzing the threat indicator of killed target, it provides drones with dynamic rewards to help drones comb local situation. Aiming at the problem that the static weight of episodic memory is very inefficient by manual adjustment in complex military confrontation environment, we use dynamic weight adjustment strategy inspired by multitask learning. Through the improvement of above method, we establish a dynamic gain drones confrontation model based on episodic memory with methods of dynamic reward mechanism and dynamic weight adjustment. The effectiveness of model is verified by machine-to-machine confrontation which provides thoughts for the analysis of military game decision-making in the complex battlefield environment.