通信与信息技术

在军事模拟战略中，作战任务的执行往往依赖于团队的高效协同，多智能体强化学习方法正是为应对这一需求而被广泛引入的。然而，在实际应用过程中，由于多智能体对所处环境状态的观测构建不够充分，导致其获取的信息较为有限，难以满足复杂军事任务中战场感知的要求。针对这一问题，提出一种基于RACE的多模态信息融合技术改进方法（RACE-MFT）。该方法通过整合属性、文本和图像三方面信息，构建了更为丰富和全面的智能体状态表示，从而增加状态维度，增强智能体的信息感知能力，使其能够做出更优的决策。实验在即时战略游戏《星际争霸Ⅱ》和自建的“争夺要地任务”环境中开展。结果显示，在《星际争霸Ⅱ》中，使用RACE-MFT的智能体对阵游戏自带AI时，胜率提升了3%。在改进算法与原算法的对抗中，胜率稳定在80%。在“争夺要地”环境里，相比其他单一模块改进，RACE-MFT的收敛奖励达到最大值。证实了RACE-MFT在处理多智能体团队协同任务时的有效性。

In military simulation strategies, the execution of combat tasks often relies on efficient teamwork, and multi-agent reinforcement learning methods have been widely introduced to meet this demand. However, in practical applications, due to insufficient observation and construction of the environmental state by multi-agent systems, the information they obtain is relatively limited, making it difficult to meet the requirements of battlefield perception in complex military tasks. A method for improving multi-modal information fusion technology based on RACE(RACE-MFT) is proposed to address this issue. This method integrates attribute, text, and image information to construct a richer and more comprehensive representation of the agent's state, thereby increasing the state dimension and enhancing the agent's information perception ability, enabling it to make better decisions. The experiment was conducted in the real-time strategy game StarCraft II and a self built "Battle for Key Tasks" environment. The results showed that in StarCraft II, when using RACE-MFT agents to compete against the game's built-in AI, the win rate increased by 3%. In the confrontation between the improved algorithm and the original algorithm, the winning rate remains stable at 80%. In the environment of "competing for important places", compared with other single module improvements, the convergence reward of RACE-MFT reaches the maximum. These all confirm the effectiveness of RACE-MFT in handling multi-agent team collaboration tasks.