Leveraging self-play techniques from poker and go, this reinforcement learning model combines human data with self-play algorithms by penalizing the model for making unlikely actions from the human data set.