Proposal: Cloud AI Gym Retro Environment

[giongto35]

OpenAI Gym retro is a framework to set up AI environment for Retro games. Some pain points

1. AI Agent is for a single-player game. Competition is just about getting as high score as possible (https://openai.com/blog/retro-contest/)
2. Setting environment is troublesome: still need to download rom + setup gym environment (getting memory access for scoring, input ...)
   Demo
   https://github.com/uvipen/Contra-PPO-pytorch
   https://github.com/uvipen/Super-mario-bros-PPO-pytorch

With Cloud Gaming enhancement for Retro Game

1. AI Agent can be applied to fighting games. Create a real competition between users. Playoff series:

• Picking "Street Fighter" game. Two AI agents are uploaded for each player. Every people can watch the AI competition.
• Watching AI training agent:
  • It's a fun process that I also like doing and I think many people also do. Mario Trainig 10M views.

2. Ease environment setup

• This is a very high-level goal and hard to get it right but extremely useful.
• Program on local/remote, but compile and run in server and watch it run.
• Server is a VM server that we provide by integration with cloud provider.

Welcome for brainstorming, technical discussion: