MLAGENTS: Is it possible to omit a failed agent's data from learning model?
I realize this isn't the most effective way to train one's agents in most situations, but I'm actually trying to simulate something more organic, so am interested in preventing "omniscience" as it were. My idea is that the agent will essentially just die after certain conditions are reached, and under those conditions, all of its learning is basically omitted from the model.
Essentially, I want the model/brain to only learn under certain specific conditions and even "delete" certain bits of what it would have otherwise learned. Is this possible?
Comment