Researcher probes new tuning-free offline RL model selection method's LSTD ties
nanjiang_cs · x · 2026-10-29
Nan Jiang engaged with Lars van der Laan's paper on selecting the best Q-function or occupancy-ratio estimator in offline RL and OPE, noting the first variant resembles an LSTD-tournament used in their own experiments (possibly without Q itself in the feature span). He adds he tried a pairwise-comparison version but couldn't express it as a non-negative loss without strong assumptions. The proposed method claims to be fast, tuning-free, requiring no extra model fitting, with oracle guarantees even when every candidate is misspecified.
Related event: PBV Method for Offline RL Model Selection Draws Peer Scrutiny(3 posts)→
More from Research
- Build-A-Chip-Bench: end-to-end chip design benchmark, top agent scores just 16.6% — teortaxesTex · 2026-10-29
- Niantic Spatial's GroundControl uses satellite images to guide multi-view diffusion for street-level novel view synthesis — kwangmoo_yi · 2026-10-29
- GroundControl paper conditions multi-view diffusion on satellite images for novel view synthesis — kwangmoo_yi · 2026-10-29
- Stanford's FACT: Force-Aware Robot Foundation Model Trained on 2,000h Human Data — yining_hong · 2026-10-29
- 86M-param decoder-only transformer plays chess on Lichess, full notebook public — andreisavu · 2026-10-29
- Harvard experiment finds AI reviews dominate human peer reviews — RexDouglass · 2026-10-29