Import hotfix - Githubissues

HumanCompatibleAI / overcooked-demo

Web application where humans can play Overcooked with AI agents.

55 stars 25 forks source link

Import hotfix #24

Closed nathan-miller23 closed 3 years ago

nathan-miller23 commented 3 years ago

Fixed really subtle circular import error with human_aware_rl that caused race conditions with tensorflow imorts
- Moved harl imports to top of file, added some extra dependencies to Dockerfile
- Still don't quite understand what caused it but it appears to be working now
Added some dummy agents (ones that stay and perform random actions) as well as an rllib agent that was trained on overcooked for ~5 training iters that will hopefully server as examples for users
Added detailed instructions on the agent training and loading into demo pipeline I've been using in the README

nathan-miller23 commented 3 years ago

Probably want to leave this open until 1-2 real agents are trained and added into the pre-trained agents directory

micahcarroll commented 3 years ago

Thanks for doing this!

What do you mean by "real" agents by the way?

nathan-miller23 commented 3 years ago

What do you mean by "real" agents by the way?

I meant agents that were trained for more than a few timesteps. I tried to train an agent on counter_circuit but the potential function in master of harl is different than the one I have in my nathan_dev branch, and thus my hyperparameters didn't transfer over well. I believe @mesutyang97 trained an agent on cramped_room, in which case I think just adding that one successfully trained agent is sufficient. If you have any comments or think we should include additional trained agents please let me know

micahcarroll commented 3 years ago

Ok cool, yes this sounds like a good idea. Initially thought you had ended up training the agents yourself (till completion). But yes, let's wait for @mesutyang97 to add the actual agent(s) in this PR and then merge it?

mesutyang97 commented 3 years ago

Yes, I have trained some good PPO agents for cramped_room on my local machine. I am working to get overcooked_demo installed so that I can test them out before including them in this PR. Should be done soon.

mesutyang97 commented 3 years ago

The good PPO Rllib self-play agent for cramped_room has been uploaded, and I can verify locally that it is working as it should. @micahcarroll Please perform a final check before we merge this into master

micahcarroll commented 3 years ago

@mesutyang97 You should probably add to the agent README the exact command you used to train this agent in the human_aware_rl repo. Later on we should clean this up and store this information in a consistent manner, but for now this seems like the most obvious choice to give people something to work off of.