Preventing an RL agent from taking invalid actions by zeroing out their probabilities before selection.