SKSurrogate
sksurrogate/concepts
Concepts
Three ideas do the work in SKSurrogate. Each is defined here in the terms the documentation uses, with the evidence the package publishes on its own behaviour.
optimization
Surrogate-based optimization
Fit an approximate model to expensive evaluations and use it to choose the next point — when that saves evaluations, and when it does not.
uncertainty
Split-conformal intervals
Finite-sample marginal coverage from a calibration split, and the assumptions that make the guarantee fail.
evaluation
Nested cross-validation
Keep model selection separate from the estimate of generalisation, so the score you report is not the score you optimised.