Multi-agent reinforcement learning research and implementation project for UAIS 26/27. Focus: Train AI agents to go from simple predator-prey chases to playing a live 2v2 soccer match using mult-agent reinforcement learning.
Description
This project introduces team members to Multi-Agent Reinforcement Learning (MARL) by training autonomous AI agents to master competitive and cooperative environments. Starting with a simple predator-prey setup, the project culminates in a live 2v2 soccer tournament game using Unity ML-Agents. The system has to learn good behavior purely through trial and error, guided by a reward signal. The environment is non-stationary, which means that through the self-play process the agents will need to adapt their behaviour to a changing environment as the teammate and opponents learn as well.
Skills Gained
Over 6 months, project members will learn practical skills in reward shaping, self-play algorithms, and experiment tracking using industry standard RL tools.
Technical Details
Python, PyTorch, PettingZoo, Stable-Baselines3, Weights & Biases, and Unity ML Agents.