Source code for OpenAI Retro Contest for Sonic the Hedgehog
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Type Name Latest commit message Commit time
Failed to load latest commit information.
baselines Added customized baselines and anyrl-py used for contest Aug 17, 2018
ppo2.docker First Commit Jun 12, 2018
sonic-validation.csv First Commit Jun 12, 2018

OpenAI Sonic the HedgeHog Retro Contest

This repo includes the source code I used for OpenAI Retro Contest. For a detail account of my approach, please check out my blog post at

Content Description

PPO Agent:

PPO joint training on multiple levels of Sonic:

PPO with curiosity-driven exploration:

PPO to train expert models:

Rainbow Agent:

Rainbow joint training on multiple levels of Sonic:

Perform local validation on validation levels with PPO model:

Train classifier to classify levels into zones:

CSV files containing OpenAI's recommended train/validation levels split: sonic-train.csv and sonic-validation.csv

Docker files with instructions to create docker images to submit to OpenAI server for evaluation: ppo2.docker and rainbow.docker

baselines and anyrl-py contain the core implementations of PPO and Rainbow with my customized edit for this contest.

Please note that the pretrained models are not included in this repo. You have to run the code to generate the models. To run the code, please read the instructions here to install all the dependencies for the contest.