Exploring compact reinforcement-learning representations with linear regression