【摘要】 深度强化学习在处理序列决策与策略探索问题上取得了很大的成功，大多从游戏中展开研究获得启发，其应用领域从单智能体场景扩展到多智能体场景中。基于纸牌的多人对战策略游戏是一种多智能体系统，但现有研究较少，且大多都来自于斗地主、德州扑克。为拓展基于纸牌的多智能体策略游戏的研究，提出了一种基于强化学习的三国杀多智能体博弈方法(SGS-MAPG),自建了以三国杀游戏为背景的2v2对战游戏场景作为实验环境，基于策略梯度的思想对合作的多个智能体建模，在其决策过程中包含了多智能体系统的团队合作与对抗，解决了多个智能体环境下的不稳定性问题。经计算机模拟对战过程，上述方法使智能体经过训练具有良好的学习决策能力，并且能够尝试获得多于基础算法的最终团队奖励，并得到高出至少12%胜率。更多还原

【Abstract】 Deep reinforcement learning has achieved great success in dealing with sequential decision-making and strategy exploration, and most of them are inspired by in-game research, and its application field has expanded from single-agent scenarios to multi-agent scenarios. Solitaire-based multiplayer strategy games are a multi-agent system, but there are few existing studies, and most of them come from Doudi Landlord and Texas Hold’em. In order to expand the research of multi-agent strategy games based on cards, this paper proposes a 2v2 three-country killing multi-agent game method(SGS-MAPG)based on reinforcement learning, which builds a 2v2 battle game scene with the background of three-kingdom killing game as the experimental environment, models cooperative multiple agents based on the idea of strategy gradient, and includes teamwork and confrontation of multi-agent systems in its decision-making process, which solves the problem of instability in multiple agent environments. Through computer simulation of the battle process, this method enables the agent to be trained to have good learning and decision-making ability, and can try to obtain more final team rewards than the basic algorithm, and get at least 12% higher win rate.更多还原