EI Stopping Rule for Cost-Aware Freeze-Thaw Bayesian Optimization
Original title: Cost-Aware Stopping for Freeze-Thaw Bayesian Optimization: When Does One-Step Lookahead Suffice?
One-step EI stopping balances expected improvement against evaluation costs in freeze-thaw Bayesian optimization.
The paper introduces a cost-aware stopping rule for freeze-thaw Bayesian optimization that advances candidates with the highest expected improvement under a surrogate and stops when no increment exceeds the cost. It constructs a lower bound certificate proving two-step lookahead corrections for premature stops and shows repeating the correction preserves expected utility. On hyperparameter optimization benchmarks, EI stopping is noninferior to corrections in most high-cost settings while gaining at low costs on two of three benchmarks.
- EI stopping rule advances candidates by largest expected improvement.
- Two-step lookahead certificate detects premature stops.
- EI stopping trades lower scores for fewer increments on two published surrogates.
- More information sharing in the surrogate raises utility.
From the original article
1 min read at arxiv.orgFreeze-thaw Bayesian optimization (BO) evaluates candidates in resumable increments but usually stops at a budget or convergence test. We maximize a budget-free utility, the best observed score minus a cost per increment, with a one-step lookahead: advance the candidate with the largest expected improvement (EI) under a fixed surrogate, and stop once no EI exceeds the cost.
By Long Vu · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.