Is it possible to do curriculum learning with a heterogenous multi agent population trained by POCA? I struggle to understand how I could express in the config that certain lessons only apply to certain types of agents.
Unfortunately, POCA does not currently support training heterogeneous multi agent groups with different behavior names. However, you can approximate this using a goal sensor for 'roles' as long as the agents have the same observation and action spaces. In this setting, curriculum would be used as it is in the single agent case.
I don't think the ML-Agents academy supports POCA + Curriculum Learning. I had to implement an explicit Training Level Manager in my code when I want to use POCA. Increase or decrease the Training Objectives, based upon Training performance within my code. (JulesVerny/DuffBallML: The use of Unity ML Agents to learn to play a simple form of Football (github.com)