Self-play and multi-agent reinforcement learning

Is self-play or multi-agent reinforcement learning possible with the current state of ML Agents? I have't tinker with it much yet but I have a project in mind which would require having multiple agents in one scene interacting with each other and cooperating.
Does ML Agents allow to set up many controllable brains learning at the same time?

 

We recently added a self-play option for symmetric, adversarial games. This is on the master branch now and will be in the next release (0.14.0); the documentation is available here https://github.com/Unity-Technologies/ml-agents/blob/0.14.0/docs/Training-Self-Play.md

However, it sounds like what you described doesn't fit that description (not symmetric or adversarial). I think we're planning out some more features in this area, but don't have the details right now.

Not specifically for cooperation, but all of the examples have the same brain shared by multiple agents.

(Edited July 14 2002: updated link to the 0.14.0-specific docs)

 

Nice! Is there an updated example project, leveraging this new option?

 

@mbaske - the Soccer and Tennis environments now use the self-play option

 

I cannot find the example code. The link https://github.com/Unity-Technologies/ml-agents/blob/master/docs/Training-Self-Play.md is expired. Can anyone help?

 

There is a bunch of links here https://github.com/Unity-Technologi...ng-in-multi-agent-environments-with-self-play
Also some more info here https://github.com/Unity-Technologi...docs/Training-Configuration-File.md#self-play

and the file you were looking for is in an old version of the documentation here: https://github.com/Unity-Technologies/ml-agents/blob/0.15.1/docs/Training-Self-Play.md

 

Hi I have done this but it say " WARNING [trainer.py:240] Your environment contains multiple
teams, but PPOTrainer doesn't support adversarial games. Enable self-play to train adversarial games.
" how can i solve this , thanks in advance

 

Hi @unity_6gCk04bKBvTBXw,
You already have a thread open where we are trying to help you with this issue. Please keep questions to that thread.