search

dgriff777 / rl_a3c_pytorch

A3C LSTM Atari with Pytorch plus A3G design
Apache License 2.0
563 stars 119 forks source link

Pretrained models #27

Closed lweitkamp closed 6 years ago

lweitkamp commented 6 years ago

Hello, is it possible to get access to some of the pre-trained models? (specifically looking for sea quest, pong and space invaders but any or all would be brilliant)

dgriff777 commented 6 years ago

which versions? i.e. -v0, v4, Deterministic-v4, NoFrameskip-v4...?

lweitkamp commented 6 years ago

I would love the NoFrameskip-v4, but Deterministic-v4 is not bad either. Also: what are the best parameters to tweak when training a more complexer game than pong?

dgriff777 commented 6 years ago

Well looks like I deleted most but did locate one, the SeaquestNoFrameskip-v4. I added to repo. It performs flawlessly with a perfect score of 999,999 beating the game every time. Enjoy!! :)

oh and make sure to set time limit for environment to the max with --max-episode-length 100000

dgriff777 commented 6 years ago

Sorry missed the parameters part lol. It will perform just fine how is. That said if you reduced the gamma parameter the agent will try to score quicker and if the environment is more deterministic reducing the tau parameter will introduce more variance

lweitkamp commented 6 years ago

Ah that's alright, thanks for the model! I'll try and see if i can fix some params to train easier. I've also heard that the latest version of pytorch kind of threw a lot of models off for some reason so maybe ill downgrade to 0.3 and try again.

dgriff777 commented 6 years ago

Yeah the code is only tested up to 0.3. Have seen too many bugs in more current versions so not ready to upgrade till its more stable as no performance improvement in doing so as well.