Reinforcement learning, efficient coding, and the statistics of natural tasks

•Reinforcement learning (RL) provides a rich conceptual framework for understanding human learning and decision making.•Most RL-inspired research in cognitive science and neuroscience has focused on identifying algorithms or procedures.•We argue for the importance of pursuing the complementary question of how RL problems are represented.•A foundation for this undertaking is provided by existing work employing the notion of efficient coding.•In an RL context, efficient coding theory points to exciting new targets for research, including the challenge of understanding the structure of naturalistic tasks.