Sonic-the-Hedgehog-A3C-LSTM-tensorflow2

Introduction

Trained an AI playing Sonic the Hedgehog by Deep Reinforcement Learning model:
Asynchronous Advantage Actor-Critic(A3C) with LSTM cell using tensorflow2-keras API

Setup

1. Clone the repository:https://github.com/Chang-Chia-Chi/Sonic-the-Hedgehog-A3C-LSTM-tensorflow2.git
2. Run pip3 install -r requirements.txt to install packages required.
3. Sonic 1, 2 and 3 & Knuckles ROMs are available on Steam:
   • Sonic The Hedgehog
   • Sonic The Hedgehog 2
   • Sonic 3 & Knuckles
4. Once you buy any of games, use script python -m retro.import.sega_classics. You'll be asked to type
   Steam username, password and Guard code:
   • Open a private session and log into Steam, there will be a guide code shown as below:
     
5. After installation, use script python -m retro.import <path to steam folder>
6. For more detail please follow this link

Train

1. Open parser.py, change argements of --game and --state to which Sonic game, Zone and Act you want to train.
   You could find all levels in these two links: -sonic-train.csv, -sonic-validation.csv
2. run python train.py to start training.
3. By default, it'll save model weights, model performance in distance, and video (.bk2) every 10 episodes.
4. model and training setting parameters are in parser.py.
5. If you want to load pre-trained weights, open train.py and set pretrain=True.`

Test

1. Open test.py, change game and state to level you want.
2. If you don't want to record gameplay, set record=False
3. Run python test.py
4. P.S. Uploaded pre-trained weights are trained with Impala CNN model

Enviroment

• python == 3.7
• tensorflow == 2.3.0

Render .bk2 record to Mp4

1. Run script python3 -m retro.scripts.playback_movie <Record file name>.bk2 <output file name>.mp4
2. For more detail please follow this link

Reference

1. https://contest.openai.com/2018-1/details/
2. https://retro.readthedocs.io/en/latest/python.html#replay-files