HDMTK: Full Integration of Hierarchical Decision-Making and Tactical Knowledge in Multi-Agent Adversarial Games

In the field of adversarial games, existing decision-making algorithms primarily rely on reinforcement learning, which can theoretically adapt to diverse scenarios through trial and error. However, these algorithms often face the challenges of low effectiveness and slow convergence in complex wargame environments. Inspired by how human commanders make decisions, this paper proposes a novel method named Fully Integrating Hierarchical Decision-Making and Tactical Knowledge (HDMTK). This method comprises an upper reinforcement learning module and a lower multi-agent reinforcement learning module. To enable agents to efficiently learn the cooperative strategy, in HDMTK, we separate the whole task into explainable subtasks, and design their corresponding subgoals for shaping the online rewards upon tactical knowledge. Experimental results on the wargame simulation platform 'MiaoSuan' show that, compared to the advanced multi-agent reinforcement learning methods, HDMTK exhibits superior performance and faster convergence in the complex scenarios.

Duke Scholars

Author Kaizhu Huang DKU Faculty