Behavior cloning is very close or better

[engyasin]

Hi,

I'm just trying to test MAGAIL as per the paper on MPE env (https://pettingzoo.farama.org/environments/mpe/simple_spread/) However, when I just train with behavior cloning (supervised learning from states to actions) it trains super fast reaching good accuracy. When I try with MAGAIL, it gets forever to reach BC level, and sometimes it return to worse performance with further training.

I know this isn't about your code, put perhaps you have an idea of what's going on. I can share my parameters if that will help.

[RITCHIEHuang]

@engyasin Have you tried some other complicated environments or benchmarks? As BC can achieve good accuracy, I personally think the task is quite simple. As to the speed of convergence, MAGAIL is bound to be slower, as it's trained in a generative way.