Skip to content

SYSTEM Cited by 1 source

Scikit-Optimize

Definition

Scikit-Optimize (aka skopt) is a Python library for sequential model-based optimization — most commonly Bayesian optimization — over real-valued, integer, or categorical parameter spaces. It is part of the scikit-learn ecosystem but is a separate package.

Core primitive

Given a parameter search space and an objective function, the optimizer proposes candidate parameter sets, observes the returned objective value, and uses prior observations to propose subsequent candidates more likely to minimize (or maximize) the objective. The underlying surrogate model is typically a Gaussian process, but random forest / gradient boosting surrogates are also supported.

The sampling primitives (skopt.space.Real, skopt.space.Integer, skopt.space.Categorical) are what users declare in configuration to describe the search space.

How Yelp uses it

Yelp's Back-Testing Engine (2026-02-02) uses Scikit-Optimize as its default optimizer. The YAML search_type: 'scikit-opt' with minimize_metric: 'average-cpl' drives iterative candidate selection against the ad-budget-allocation simulation's output metrics.

Verbatim from the post: "The optimizer (Bayesian in this case) is designed to extract the most promising candidates that aim at minimizing a given metric."

Yelp also supports grid search and listed search (each candidate specified directly in YAML) as alternatives, but the post notes: "for all kinds of search except Scikit-Opt, the optimizer doesn't really act as an optimizer but just a wrapper that yields the next candidate to try."

Why it matters for system design

The relevance for systems work is not the statistical machinery but the iterative loop pattern: expensive evaluation (Yelp's back-testing simulation takes non-trivial time per candidate), small evaluation budget (max_evals = 25 in the example), and learning from prior results to guide future proposals. This shape recurs in any simulation-driven parameter-search workload: performance autotuning, schema-advisor hypothesis-index evaluation, feature-flag-space exploration.

Canonical instance of concepts/bayesian-optimization-over-parameter-space.

Seen in

Last updated · 476 distilled / 1,218 read