hydromodpy.calibration.adapters.da_mh_gp_adapter#
Delayed-Acceptance Metropolis-Hastings with Gaussian-process surrogate.
Two-stage MCMC for expensive simulators: stage 1 filters proposals with a
sklearn GaussianProcessRegressor surrogate of the log-posterior; stage 2
corrects the acceptance ratio with one full-model call, so the chain targets
the exact posterior. The surrogate is retrained every retrain_interval
new evaluations. The objective is assumed to return RMSE; the
log-likelihood is built internally as -0.5 * (RMSE / sigma_noise)**2.
The sampler pulls evaluations from the engine via a background thread
(like scipy_adapter).
Classes
|
Delayed-Acceptance Metropolis-Hastings with GP surrogate. |