Min–Max Q-learning for multi-player pursuit-evasion games

Abstract

In this paper, we address a pursuit-evasion game involving multiple players by utilizing tools and techniques from reinforcement learning and matrix game theory. In particular, we consider the problem of steering an evader to a goal destination while avoiding capture by multiple pursuers, which is a high-dimensional and computationally intractable problem in general. In our proposed approach, we first formulate the multi-agent pursuit-evasion game as a sequence of discrete matrix games. Next, in order to simplify the solution process, we transform the high-dimensional state space into a low-dimensional manifold and the continuous action space into a feature-based space, which is a discrete abstraction of the original space. Based on these transformed state and action spaces, we subsequently employ min–max Q-learning, to generate the entries of the payoff matrix of the game, and subsequently obtain the optimal action for the evader at each stage. Finally, we present extensive numerical simulations to evaluate the performance of the proposed learning-based evading strategy in terms of the evader’s ability to reach the desired target location without being captured, as well as computational efficiency.