MAgent2 RL Final Project

Overview

This project trains a multi-agent RL model in the MAgent2 battle environment. The code sets up agents and evaluates them against:

• Random Agents
• A Pretrained Agent
• A Final (stronger) Agent

For each setting, the performance is measured by rewards and win rates.

The figure below depicts the results when the VDN mixer model competes with random opponents, pretrained agent, final pretrained agent.

Installation

pip install -r requirements.txt

Source Code Summary

• dqn_train/dqn.ipynb: Main training script for the DQN-based agent.

  

  you need to "copy&edit" on kaggle to run script, then upload the red.pt file to kaggle and correct the path in the source code

• dqn_train/dqn_noise_network.ipynb: Main training script for the DQN with Noise Network agent.

  

  you need to "copy&edit" on kaggle to run script, then upload the red.pt file to kaggle and correct the path in the source code

• vdn_mixer_train/vdn_mixer.ipynb: Main training script for the VDN agent.

  

  you need to "copy&edit" on kaggle to run script, then upload the red.pt file to kaggle and correct the path in the source code

• test_model/eval_test/test_dqn(others same).ipynb: Evaluation code against different opponents.
  you can find checkpoints at source model/dqn/checkpoints or model/vdn_mixer/

Report

Survey_on_MAgent2_Battle_Using_DQN_Variants_Report_Group22.pdf

• Experimental setup
• DQN variant details
• Performance comparison and discussion

Demos

For additional demos, see the video folder

References

1. MAgent2 GitHub Repository
2. MAgent2 API Documentation

For environment setup and agent interaction details, refer to the MAgent2 documentation.