Cooperative AI Using Multi-Agent Reinforcement Learning for Team-Based Game

Abstract

This thesis explores the use of Multi-Agent Reinforcement Learning (MARL) to develop cooperative behavior in a team-based game environment. A capture-the-flag scenario is implemented in Unity ML-Agents, where agents must coordinate to attack, defend, and capture objectives. The system uses a decentralized training and execution framework with parameter sharing, allowing agents to learn a common policy while acting independently based on local observations. The results demonstrate that agents can develop coordinated strategies and role differentiation without centralized control or explicit communication. These findings highlight the effectiveness of decentralized MARL with shared policies in complex, partially observable environments.