README.MD

Introduction

This is our project for Reinforcement Learning with PyBoy, where we trained agents to play GameBoy games, namely Super Mario Land and Kirby's Dream Land.

Report: report

Training time:

• Super Mario Land: 26 hours on a 4-core CPU
• Kirby's Dream Land: 20 hours on CUDA GPU
• Kirby's Dream Land Bossfight: 12 hours on CUDA GPU

Requirements

• Python3.7 on Windows
• Python3+ on Linux
• PyBoy (https://github.com/Baekalfen/PyBoy)
• SDL2
  • Ubuntu: sudo apt install libsdl2-dev
  • Fedora: sudo dnf install SDL2-devel
  • macOS: brew install sdl2
  • Windows: PyBoy guide https://github.com/Baekalfen/PyBoy/wiki/Installation#windows-10-64-bit Download link https://www.libsdl.org/download-2.0.php
• For package requirements see requirements.txt
• GameBoy ROM files for Super Mario Land or Kirby's Dream Land (place these in /games)
  • Filename must be "Kirby_Dream_Land" and "Super_Mario_Land" respectively

Run with Python

To run from source, first install dependencies:

• pip3 install -r requirements.txt

Then, run:

• python3 main.py

Docker

Build command: docker build --tag pyboy-rl .

Once inside the image run python3 main.py to start the program. The docker container only supports headless mode, and the game emulator ui will not show up.

Based on

DDQN - https://arxiv.org/abs/1509.06461

PyTorch RL Super Mario Bros Example - https://pytorch.org/tutorials/intermediate/mario_rl_tutorial.html