Safe RL With Stability Guarantees: Learning Without Ever Falling Down
tomssilver · x · 2026-09-07
Tom Silver's paper pick of the week is Berkenkamp et al.'s NeurIPS 2017 work on safe model-based RL with stability guarantees.
- Key distinction: most safe RL guarantees the learned policy is safe; this paper guarantees safety during learning — crucial for RL on real systems.
- Method: extends control-theoretic Lyapunov stability verification, using Gaussian process dynamics models to obtain high-performance policies with provable stability certificates; under regularity assumptions it safely collects data to improve control and expand the safe region.
- Experiments: safely optimizes a neural network policy on a simulated inverted pendulum without it ever falling.
More from Research
- TAMER, the First General-Purpose RLHF Algorithm From 2008, Gets a New Open-Source Release — dhadfieldmenell · 2026-09-07
- AI Math Podcast Sits Down With CMU's Jeremy Avigad: Can Mathematics Be Automated? — EchoShao8899 · 2026-09-07
- 'The honest claim' emerges as telltale AI-writing phrase in bioRxiv preprints — lpachter · 2026-09-07
- Is Reproducibility a Lost Cause in ML Research? A Debate — NeighborhoodFatCat · 2026-09-07
- GPT Astra Solves 1962 Erdős–Sós Conjecture in 1 of 3 Tries for $363 — burny_tech · 2026-09-07
- Microsoft Paper: Distilling 50 Trajectories Into Skill Cards Matches Costly Reasoning — rohanpaul_ai · 2026-09-07