Adaptive Interception in Dynamic Domains: Exploration of Hybrid Reinforcement Learning in Pursuit-Evasion Tasks

Abstract

This paper investigates a hybrid approach that integrates clas sical interception heuristics with reinforcement learning (RL) for solving real-time pursuit-evasion problems in both single agent and multi-agent scenarios. Drawing inspiration from Weintraub et al.’s range-limited pursuit-evasion formulation, we developed a simulation and learning architecture com bining Proximal Policy Optimization (PPO) with classical prediction-based interception. Through adaptive curriculum learning and an enhanced reward function, the hybrid agent learns to dynamically balance between geometric reason ing and learned behaviors. We first validate our approach in single-agent scenarios, then extend to a multi-agent set ting featuring coordinated pursuer teams defending stations against RL-trained evader flocks using adversarial self-play. Our evaluations reveal that the hybrid method significantly outperforms classical approaches in both settings, achiev ing a 94% capture rate versus 24% in single-agent scenarios and demonstrating robust coordination in multi-agent defense tasks against adaptive adversaries.