search

nazaruka / gym-http-api

NSGA2-based Sonic agent + experimental code
MIT License
1 stars 1 forks source link

Apply an algorithm to an Atari game #7

Closed schrum2 closed 5 years ago

schrum2 commented 5 years ago

Not sure how much work is required for this. Open AI Gym supports interfacing to Atari games, but I think you still have to compile ALE (Arcade Learning Environment) with the Stella Atari emulator to take advantage. You may also need to have the appropriate ROM files.

Still, investigate the use of Open AI Gym for Atari games and see if you can apply a DQN to one.

schrum2 commented 5 years ago

From what I recall, the gym-based Atari options are Linux/Mac only. Can you update this thread with some of the details on this?

What about Docker and/or VirtualBox based work-arounds for this?

schrum2 commented 5 years ago

Atari might be possible. Look at this thread: https://github.com/openai/gym/issues/11 Specifically, skip down to the post by icoxfog417

schrum2 commented 5 years ago

This also looks promising: https://github.com/rybskej/atari-py

schrum2 commented 5 years ago

Thanks to gym-retro, we have also had success with Atari games. Therefore, we are close to closing this issue. This is what is left to do:

1) Get DQN to work on Atari games. Either modify our copy of DQN or get the DQN baseline implementation from Open AI's baseline repo: https://github.com/openai/baselines 2) Train a DQN on at least two of the games from this paper: https://www.cs.toronto.edu/~vmnih/docs/dqn.pdf 3) Log information in this GitHub thread about how long it takes to train, and what level of performance we attain on each game you test.

Then close the issue.

schrum2 commented 5 years ago

No more time for benchmarks, though we seem to be able to learn in Atari games