The reinforcement learning problem of complex action control in multiplayer online battlefield games has brought considerable interest in the deep learning field. This problem involves more complex states and action spaces than traditional confrontation games, making it difficult to search for any strategy with human-level performance. This paper presents a deep reinforcement learning model to solve this problem from the perspective of game simulations and algorithm implementation. A reverse reinforcement-learning model based on high-level player training data is established to support downstream algorithms. With less training data, the proposed model is converged quicker, and more consistent with the action strategies of high-level players’ decision-making. Then an intelligent deduction algorithm based on DDQN is developed to achieve a better generalization ability under the guidance of a given reward function. At the game simulation level, this paper constructs Monte Carlo Tree Search Intelligent Decision Model for turn-based antagonistic deduction games to generate next-step actions. Furthermore, a prototype game simulator that combines offline with online functions is implemented to verify the performance of proposed model and algorithm. The experiments show that our proposed approach not only has a better reference value to the antagonistic environment using incomplete information, but also accurate and effective in predicting the return value. Moreover, our work provides a theoretical validation platform and testbed for related research on game AI for deductive games.

多人在线战场游戏中复杂动作控制的强化学习问题引起了深度学习领域的极大兴趣。这个问题涉及到比传统对抗游戏更复杂的状态和动作空间，使得很难找到任何具有人类水平表现的策略。本文从游戏模拟和算法实现的角度提出了一种深度强化学习模型来解决这个问题。建立了基于高层球员训练数据的反向强化学习模型，以支持下游算法。在训练数据较少的情况下，所提出的模型收敛速度更快，更符合高层玩家决策的行动策略。然后开发了一种基于DDQN的智能推导算法，在给定奖励函数的指导下获得更好的泛化能力。在游戏模拟层面，本文构建了蒙特卡洛树搜索智能决策模型，用于回合制对抗推理游戏生成下一步动作。此外，还实现了结合离线和在线功能的原型游戏模拟器，以验证所提出模型和算法的性能。实验表明，我们提出的方法不仅对使用不完整信息的对抗环境具有更好的参考价值，而且在预测返回值方面也准确有效。此外，我们的工作为演绎游戏的游戏人工智能相关研究提供了理论验证平台和测试平台。