/retro-contest-sonic

Agent used on OpenAI meta-learning contest using Sonic The Hedgehog game

Primary LanguagePython

Retro Contest Open AI

https://contest.openai.com/

Reinforcement Learning with Policy Gradient

Uses 2 convolution layers and 1 fully connected layer to control Sonic

Video of agent playing Sonic! -> https://contest.openai.com/users/911

sonic

  1. To run this code first retro-contest and follow the instructions on the contest page: https://contest.openai.com/details

  2. pg_agent_train.py trains the agent over defined episodes. You can change the SonicTheHedgehog level and act in that file to train on different levels. pg.py contains the model used for training.

  3. pg_agent.py was used for submission.

  4. With only half day of training on CPU, this agent was not able to generalize well enough but still achieved a decent score

  5. Eventually ran a PPO baseline and was able to rank 140 out of 229 participants: https://contest.openai.com/leaderboard