Gaussian-enhanced reinforcement learning for scalable evasion strategies in multi-agent pursuit-evasion games

Abstract

This paper introduces a Gaussian-enhanced multi-agent reinforcement learning framework for developing scalable evasion strategies in dynamic pursuit scenarios. The proposed methodology addresses two critical challenges in unknown environments: sparse reward structures and local optima convergence, while enhancing escape feasibility through probabilistic decision-making. By integrating Gaussian process regression with Q-function approximation, the framework enables efficient online parameter adaptation and demonstrates improved sample efficiency in high-dimensional state spaces. Comprehensive simulations and physical experiments across terrestrial and aerial robotic platforms validate the framework’s effectiveness and robustness in complex evasion tasks. The architecture’s modular design permits generalization to multi-agent pursuit-evasion scenarios with variable participant numbers, establishing a versatile foundation for strategic interactions in large-scale autonomous systems.