osmose.calibration.surrogate_de

GP surrogate-assisted differential evolution.

Trains a Gaussian Process emulator on a small batch of real evaluations, runs DE on the GP-predicted objective (cheap), then real-evaluates the top-K candidates selected via Lower Confidence Bound acquisition, and retrains. Iterating yields convergence in 5–10× fewer real evaluations than vanilla DE on smooth problems.

Tier C1 of the speedup roadmap. Builds on the existing osmose/calibration/surrogate.py (sklearn GP with Matern kernel) for the emulator.

Functions

surrogate_assisted_de(objective, bounds[, ...])

Optimise objective over bounds using GP-assisted DE.