Mildly Conservative Q-Learning for Offline Reinforcement Learning

This paper explores mild but enough conservatism for offline learning while not harming generalization.