pystra.active_learning.pc_kriging.PCKrigingSurrogate#

Use it: User guide · Worked example · Theory

class PCKrigingSurrogate(*, degree=(1, 2, 3), q_norm=1.0, max_interaction=None, length_scale=1.0, correlation='matern52', optimize=True, noise=1e-08, n_restarts=0, seed=None)[source]#

Bases: Surrogate

Sequential sparse PC-Kriging in independent normal coordinates.

Parameters:

  • degree (int | Sequence[int]) – Sparse Hermite trend candidates; default (1, 2, 3).

  • q_norm (float) – Hyperbolic truncation, default 1.

  • max_interaction (int | None) – Maximum polynomial interaction order.

  • correlation (str) – ‘matern52’ (default, consistent with KrigingSurrogate) or ‘gaussian’.

  • length_scale (float | Sequence[float]) – Initial correlation scales, default 1; a scalar is broadcast across dimensions. Fixed scales are possible with optimize=False.

  • optimize (bool) – Default True: fit anisotropic scales by profiled maximum likelihood. Bounds are [0.01, 100] in standard-normal coordinate units.

  • noise (float) – Positive relative diagonal nugget, default 1e-8. Predictions describe the latent response, excluding new observation noise.

  • n_restarts (int) – Additional random optimizer starts, default 0. Failed optimization raises RuntimeError; it cannot silently produce a successful fit.

  • seed (int | None) – Local generator seed for additional optimization starts.

Notes

The sparse trend is selected before optimizing the correlation model; this is sequential PC-Kriging, not optimal PC-Kriging’s search along every LAR support. GLS re-estimates the trend for each correlation candidate. Universal predictive variance includes trend-estimation uncertainty. Exactly representable training data yield zero residual process variance; this does not certify correctness outside the observations. Requires only NumPy/SciPy, with row-wise input points and scalar responses.

Methods

fit

Replace fitted state; errors leave prediction unavailable.

predict

Return universal mean/std conditional on fitted trend and scales.
fit(points, values)[source]#

Replace fitted state; errors leave prediction unavailable.

Return type:

None

predict(points)[source]#

Return universal mean/std conditional on fitted trend and scales.

Return type:

tuple