Custom training script

[JulienHansen]

Hi everyone,

I was wondering if anyone here has tried implementing their own training algorithms, such as QMIX or others. If so, do you think it’s feasible to implement something like this within a custom train.py file? While I’m confident it can be done, I’m curious about the performance—would it be possible to achieve something efficient and robust?

Apologies if this question seems trivial; I understand that such implementations are possible, but I’m mainly wondering if they would provide any practical benefits. And if someone had already tried to do that

Looking forward to your thoughts!

[JulienHansen]

I m adding a question kinda related to a custom training script, is the MARL workflow also suitable for adversarial task ?
Thanks in advance