Reasoning for Representations for Learning-Based Control