mrkwatkins.oakemu_rl.framework.agents.ActionMasker

Classes and functions

• ActionMasker

Constants

MASKED_ACTION_LOGIT

MASKED_ACTION_LOGIT = -1000000000.0

Value assigned to excluded actions' logits/Q-values, small enough to zero out their probability mass after softmax/argmax without producing NaN/inf in downstream loss computations.

Source