From Robotics to Sepsis Treatment: Offline RL via Geometric Pessimism
Sarthak Wanjari · 2026
Read From Robotics to Sepsis Treatment: Offline RL via Geometric Pessimism by Sarthak Wanjari (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2602-08655.txt.
Description
Offline reinforcement learning promises policies learned from static data, but over-rating unseen actions can wreck it, especially on fractured, sparse data. Sarthak Wanjari proposes Geometric Pessimism, a frame that adds density-based penalties from nearest-neighbor distances in state-action embedding space to standard expectile learning. His arXiv paper targets the efficiency-versus-rigor trade that dogs existing methods.
Penalties are precomputed, so caution enters through reward shaping at negligible training cost. On a standard robot benchmark, the method beats its base algorithm by over eighteen points on unstable replay tasks and cuts seed variance fourfold. On a sepsis critical-care dataset, it improves over clinician agreement where the base method collapses to imitation. Readers get a safe, cheap guard against out-of-distribution optimism.
Catalog record
| CREATOR | Sarthak Wanjari |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 31 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2602-08655