Successor Representations
Overview
Successor representations predict which states or features an agent is likely to encounter in the future, allowing value functions to be recomputed when rewards change. Methods include successor features, forward-backward factorization of successor measures, and spectral or low-rank representations, with temporal abstraction and full-gradient training used to improve learning and stability. These representations are tested in control, navigation, visual reinforcement learning, and transfer settings, and are also being explored for hierarchy, safety constraints, and preference learning. A central challenge is balancing compact, reusable representations against errors at long horizons or when rewards and transition dynamics differ from those seen during training.
Latest papers
12 papers0 papers in the last four weeks, against 1 the four weeks before. 0.0% of all new papers.