<ul><li>Successor representation (SR) in reinforcement learning (RL) has gained attention for addressing key challenges like exploration, credit assignment, and generalization by encoding transition dynamics.</li><li>A new representation called default representation (DR) is introduced, which considers both reward and transition dynamics while learning and making decisions.</li><li>The paper provides theoretical foundation for DR, methods for learning it, and its extension to function approximation.</li><li>Empirical analysis shows that DR outperforms SR in various RL tasks like reward shaping, option discovery, exploration, and transfer learning.</li></ul>

Reward-Aware Proto-Representations in Reinforcement Learning

Discover more