Deep Reinforcement Learning Task Assignment Based on Domain Knowledge
Jiayi Liu, Gang Wang, Xiangke Guo, Siyuan Wang, Qiang Fu
Abstract
Deep Reinforcement Learning (DRL) methods are inefficient in the initial strategy exploration process in large-scale complex scenarios. This is becoming one of the bottlenecks in their application to large-scale game adversarial scenarios. This paper proposes a Safe reinforcement learning combined with Imitation learning for Task Assignment (SITA) method for a representative red-blue game confrontation scenario. Aiming at the problem of difficult sampling of Imitation Learning (IL), this paper combines human knowledge with adversarial rules to build a knowledge rule base; We propose the Imitation Learning with the Decoupled Network (ILDN) pre-training method to solve the problem of excessive initial invalid exploration; In order to reduce invalid exploration and improve the stability in the later stages of training, we incorporate Safe Reinforcement Learning (Safe RL) method after pre-training. Finally, we verified in the digital battlefield that the SITA method has higher training efficiency and strong generalization ability in large-scale complex scenarios.