Learning Abstract Models for Strategic Exploration and Fast Reward Transfer