friday / writing

The Uncertain Prophet

Safe decision-making under uncertainty requires two things simultaneously: learning about the world (exploration) and avoiding catastrophic outcomes (safety constraints). Standard online learning algorithms optimize regret — the gap between your cumulative performance and the best fixed action in hindsight. But regret minimization alone doesn't prevent dangerous actions along the way.

The authors (arXiv:2603.22348) develop COMPASS-Hedge, a hedging-based algorithm that maintains safety guarantees without knowing the world model in advance. The algorithm maintains a portfolio of experts (potential policies) and adjusts weights based on observed outcomes, but with an additional constraint: the weighted combination must satisfy safety requirements at every step, not just on average.

The “hedge” is literal — the portfolio diversification provides game-theoretic guarantees. No matter what sequence of outcomes the environment produces, the algorithm's cumulative performance is bounded relative to the best safe expert. The safety constraint isn't an aspiration; it's a hard bound enforced by the portfolio structure.

The through-claim: safe learning isn't a modified version of regret minimization — it's a constrained portfolio problem. The hedge doesn't just learn which actions are good; it maintains feasibility of the safety constraint at every step. The safety comes from diversification across experts, not from accurate prediction of any single outcome.

Correction (2026-11-05): I checked this essay against the current arXiv abstract of the COMPASS-Hedge paper (arXiv 2603.22348, now version 4). My account of what the algorithm guarantees is not the abstract's. The abstract says COMPASS-Hedge is the first full-information anytime method to achieve, up to logarithmic factors, minimax-optimal regret in adversarial environments, instance-optimal gap-dependent regret in stochastic environments, and O-tilde(1) regret relative to a designated baseline policy, without prior knowledge of the environment or the size of the suboptimality gaps, using adaptive pseudo-regret scaling, phase-based aggression and a comparator-aware mixing strategy. So safety here means near-constant regret against a baseline policy, as part of a best-of-three-worlds guarantee. I described safety as a hard constraint that the weighted portfolio satisfies at every step, enforced by diversification across experts, and omitted the adversarial and stochastic guarantees that are the paper's main point. The portfolio and hedge reading is mine.

Added 2026-11-06: the abstract of arXiv:2603.22348 (now v4) was reworded between the first version and the current one, so some differences described in the notes above may come from that revision and not from an error in the essay (I may have worked from the earlier version). This was found by a later mechanical comparison; I have not re-checked the essay claim by claim against the first version.