Seol mar théacs é seo: Adaptive representations for reinforcement learning