Stagewise Reinforcement Learning and the Geometry of the Regret Landscape
Fuente:
arXiv
Salvato in:
| Autori principali: | , , , , |
|---|---|
| Natura: | Preprint |
| Pubblicazione: |
2026
|
| Soggetti: | |
| Accesso online: | |
| Tags: |
Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
|
| _version_ | 1866915816543354880 |
|---|---|
| author | Elliott, Chris Urdshals, Einar Quarel, David Farrugia-Roberts, Matthew Murfet, Daniel |
| author_facet | Elliott, Chris Urdshals, Einar Quarel, David Farrugia-Roberts, Matthew Murfet, Daniel |
| contents | Singular learning theory characterizes Bayesian learning as an evolving tradeoff between accuracy and complexity, with transitions between qualitatively different solutions as sample size increases. We extend this theory to reinforcement learning, proving that the concentration of a generalized posterior over policies is governed by the local learning coefficient (LLC), an invariant of the geometry of the regret function. This theory predicts that deep reinforcement learning with SGD should proceed from simple policies with high regret to complex policies with low regret. We verify this prediction empirically in a gridworld environment exhibiting stagewise policy development: phase transitions over training manifest as "opposing staircases" where regret decreases sharply while the LLC increases. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2601_07524 |
| institution | arXiv |
| publishDate | 2026 |
| record_format | arxiv |
| spellingShingle | Stagewise Reinforcement Learning and the Geometry of the Regret Landscape Elliott, Chris Urdshals, Einar Quarel, David Farrugia-Roberts, Matthew Murfet, Daniel Machine Learning Singular learning theory characterizes Bayesian learning as an evolving tradeoff between accuracy and complexity, with transitions between qualitatively different solutions as sample size increases. We extend this theory to reinforcement learning, proving that the concentration of a generalized posterior over policies is governed by the local learning coefficient (LLC), an invariant of the geometry of the regret function. This theory predicts that deep reinforcement learning with SGD should proceed from simple policies with high regret to complex policies with low regret. We verify this prediction empirically in a gridworld environment exhibiting stagewise policy development: phase transitions over training manifest as "opposing staircases" where regret decreases sharply while the LLC increases. |
| title | Stagewise Reinforcement Learning and the Geometry of the Regret Landscape |
| topic | Machine Learning |
| url | https://arxiv.org/abs/2601.07524 |