Nash–Minmax Strategies for Multiagent Pursuit–Evasion Games With Reinforcement Learning
Yuhan Wang, Hao Zhang, Zhuping Wang, Huaicheng Yan
Abstract
This article investigates the pursuit-evasion games for target capture in multiagent systems. To address this challenge, a novel data-driven optimal control policy is proposed, leveraging off-policy reinforcement learning and Nash-minmax strategies. First, a comprehensive framework for multiagent pursuit-evasion games is developed, modeled as a two-layer game structure. In this framework, interactions among agents within the same team are characterized as nonzero-sum games, while interactions between opposing teams are adversarial and treated as zero-sum games. Second, Nash-minmax strategies are introduced to solve the formulated multiagent pursuit-evasion games. These strategies effectively derive distributed Nash solutions for agents within the same team and adversarial worst-case policies for agents in opposing teams. Furthermore, to eliminate the reliance on prior knowledge of agent dynamics and initial stabilizing control gains, a data-driven optimal control policy is designed, ensuring the achievement of target capture. Finally, a numerical example is provided to demonstrate the effectiveness and practical applicability of the proposed approach.