Online surrogate repair reduces costly evaluations in ai design searches
Online Surrogate Repair: Decoupling High-Fidelity Feedback from Search Length in Closed-Loop Discovery
Machine LearningArtificial Intelligence
Summary
Generating new designs with AI can be cheap and fast, but checking which designs really work often requires expensive tests or experiments. The authors propose a method called online surrogate repair, where the AI mostly uses a fast, rough model to explore designs, but occasionally checks a few key designs with accurate tests to improve that model. This approach saves many costly tests while still guiding the AI toward good designs. They show it works better than older methods that either test every design or never update the model during the search.
What this means in practice
- •For industrial automation teams: Reduce expensive experimental testing during automated design by selectively verifying candidate designs to improve predictive models in real time.
- •For machine learning engineers: Improve hyperparameter or architecture search efficiency by updating surrogate models selectively with costly ground-truth evaluations during iterative optimization.
Authors
Xiaotang Feng, Philip Torr, Bruno Andreis
Abstract
Closed-loop AI scientists can generate candidate designs at low marginal computational cost, whereas reliable feedback may require wet-lab synthesis, characterization, or high-fidelity computation. Addressing this imbalance through custom laboratory automation remains infrastructure-intensive and costly, while replacing new experiments with a fixed surrogate leaves persistent model errors that can be amplified by optimization. We propose \emph{online surrogate repair} (OSR), a closed-loop algorithm that uses sparse high-fidelity evaluations to update the surrogate throughout a longer agent search conducted primarily with inexpensive surrogate feedback. An acquisition rule selects which designs from the agent's accumulated proposals receive high-fidelity evaluation, and the resulting labels update the surrogate used in subsequent episodes. Across controlled synthetic environments, we demonstrate that improving global surrogate fit does not necessarily reduce maximum regret, whereas Q90-UCB and expected improvement (EI) substantially reduce regret by directing evaluations toward regions that determine the optimizer's decisions. On MADE, controls receiving high-fidelity feedback after every episode require $6.36$--$7.23\times$ more oracle queries to match Online EI under two LLM orchestrators and $10.27\times$ more under the non-LLM Chemeleon+MLIP workflow. Online surrogate repair introduces a novel third feedback regime between fixed-surrogate operation and high-fidelity feedback after every episode, separating the frequency of high-fidelity evaluation from the duration of the agent's search.