Terminal Representation: A New RL Abstraction Method
Researchers introduce the terminal representation (TR), a novel spatio-temporal abstraction for reinforcement learning. Unlike the successor representation (SR) and default representation (DR), TR encodes reward-weighted trajectories as a lower-dimensionality object and can be used directly for downstream tasks like option discovery, reward shaping, transfer learning, and exploration without eigenvector computations. The work is described in arXiv:2605.31289.
Key facts
- arXiv:2605.31289 introduces the terminal representation (TR).
- TR is a spatio-temporal abstraction for reinforcement learning.
- TR encodes reward-weighted trajectories similarly to the default representation (DR).
- TR can be learned as a lower-dimensionality object.
- TR can be used directly for option discovery, reward shaping, transfer learning, and exploration.
- TR does not require eigenvector computations.
- The successor representation (SR) encodes states by future trajectories.
- The default representation (DR) weights trajectories with reward.
Entities
—