A Multi-agent Reinforcement Learning Framework for Simulating Asymmetric UAV Swarm Combat

Abstract

This paper introduces a sophisticated 2D simulation framework designed for evaluating Unmanned Aerial Vehicle (UAV) swarm tactics in a competitive multi-agent environment. The simulation features two opposing teams, Blue and Red, each composed of heterogeneous drones with distinct attributes (speed, endurance, firepower, defense, etc.). The core objective revolves around neutralizing the opponent's "General"(G-type) drone or achieving a predefined score target. The Blue team employs a hybrid AI strategy combining Ant Colony Optimization (ACO) for scout patrol pathing and a Brown-Robinson game-theoretic approach for a specialized Hunter drone to track and neutralize Evader drones. The Red team utilizes reactive behaviors, including fight-or-flee logic and strategic G-drone positioning, with a focus on assaulting the Blue team's designated operational zone.