系统仿真学报 ›› 2026, Vol. 38 ›› Issue (3): 758-775.doi: 10.16182/j.issn1004731x.joss.25-0298

• 论文 • 上一篇    

LLM赋能的战术兵棋决策Agent构建方法

刘大勇1,2, 董志明1, 郭齐胜1, 高昂3, 邱雪欢1   

  1. 1.陆军兵种大学,北京 100072
    2.中国人民解放军32302部队
    3.陆军研究院,北京 100072
  • 收稿日期:2025-04-12 修回日期:2025-06-04 出版日期:2026-03-18 发布日期:2026-03-27
  • 通讯作者: 董志明
  • 第一作者简介:刘大勇(1984-),男,工程师,博士生,研究方向为装备作战仿真、智能决策。
  • 基金资助:
    军事类研究生资助课题重点项目(JG2024B2043)

Construction Approach of LLM-empowered Tactical Wargame Decision-making Agents

Liu Dayong1,2, Dong Zhiming1, Guo Qisheng1, Gao Ang3, Qiu Xuehuan1   

  1. 1.Army Arms University of PLA, Beijing 100072, China
    2.PLA 32302 Troops
    3.Academy of Army, Beijing 100072, China
  • Received:2025-04-12 Revised:2025-06-04 Online:2026-03-18 Published:2026-03-27
  • Contact: Dong Zhiming

摘要:

在战术兵棋推演中,决策Agent是人机、机机以及人机混合对抗的关键支撑,其智能化水平至关重要。针对传统决策Agent存在的适应性不足、策略单一、构建成本高等问题,提出一种大小模型融合驱动的决策框架,并重点研究了LLM与行为树、有限状态机、启发式搜索、深度强化学习等常规决策Agent构建方法的融合方式。本研究可为战术兵棋决策Agent构建提供新的思路和技术路径。

关键词: LLM, 战术兵棋, 决策Agent, 融合决策框架

Abstract:

Decision-making agents are critical enablers for implementing human-machine, machine-machine, and hybrid human-machine adversarial interaction in tactical wargaming, where the intelligence level of the agent is crucial. To address the limitations of traditional decision agents such as insufficient adaptability, simplistic strategies, and high construction costs, a fusion decision framework driven by the large and small models was proposed. It specifically investigated the fusion approach of large language models with conventional decision-making agent construction approaches, including behavior trees, finite state machines, heuristic search, and deep reinforcement learning. New ideas and technical pathways are provided for the construction of tactical wargame decision-making agents.

Key words: LLM, tactical wargame, decision-making agent, fusion decision framework

中图分类号: