Exploring Compact Reinforcement-learning Representations with Linear Regression