Extend and reuse designs

The mergen.sequential module works on existing designs: it adds points, reuses previous campaigns, orders runs, and partitions designs for cross-validation. Every function looks only at parameter-space geometry, so nothing here depends on your measured responses.

Extend an existing design

Add space-filling points to a design while keeping every original run unchanged:

from mergen import sequential

extended = sequential.extend(sampler, base.best_design, n_new=10)

The first rows of extended are exactly your original points; the n_new additions are optimised to fill the space around them. This is the right tool when a pilot design is to be grown into a larger one without wasting completed runs.

Reuse a previous campaign

To rebuild a result object around a design you already have, for example to re-report or extend a design loaded from a file, use load_design:

sampler.load_design(previous_campaign)

load_design deliberately refuses to combine with n_samples or add_prescribed: the design is taken as given rather than re-optimised, and mixing the two intents silently would be a trap. A labelled variant records a name and colour for reporting:

sampler.load_design(prev_df, name='campaign_2024', color='#ff8800')

Order runs for sequential execution

Assign an execution order in which every leading subset of the design is itself space-filling, so an early stop still leaves a balanced design:

ordered = sequential.run_order(sampler, extended.best_design)

This matters whenever runs are executed serially and the campaign might be halted early, on a budget cut or an interim analysis.

Sub-sample and split

Pick a small representative subset of a larger design with Kennard-Stone selection:

pilot = sequential.subsample(sampler, extended.best_design, n_select=6)

Or partition a design into space-filling folds for cross-validation, so that each fold and each training set covers the whole space:

folds = sequential.k_fold_split(sampler, extended.best_design, k=3)

A nested constructor builds an inner design that is a subset of a larger outer design, for multi-fidelity studies where two budget tiers share one space.