This project details a solution OpenAI's CarRacing-v0 game. This model uses a three layer DQN regularized with dropout and using a curriculum learning approach. The trained model exceeds 900 in average score, sufficient to meet the criterion for solution. More information at: https://github.com/openai/gym/wiki/Leaderboard#carracing-v0
Reinforcement Learning at UCLA IPAM RIPS 2018. Industry Sponsor: Advanced Micro Devices.
Contributers:
- Patrik Gerber*, University of Oxford
- Jiajing Guan*, George Mason University
- Elvis Nunez*, Johns Hopkins University
- Kaman Phamdo*, University of Maryland
- Tonmoy Monsoor, UCLA
- Nicholas Malaya, AMD Research
* Denotes Equal Contribution.
Initial goals:
- Add papers and background
- Get OpenAI gym Sandbox working
- Get game working/imported
References:
https://blog.insightdatascience.com/reinforcement-learning-from-scratch-819b65f074d8
https://blog.openai.com/requests-for-research-2/
http://www.andreykurenkov.com/writing/ai/a-brief-history-of-game-ai/
https://blog.openai.com/retro-contest/
https://medium.com/@dhruvp/how-to-write-a-neural-network-to-play-pong-from-scratch-956b57d4f6e0