Objects and settings at a glance

Every object you can create and every setting you can change, on one page. Each entry links to the page that explains it in depth; the full signatures live in the API reference.

Defining the space

ParameterSpace holds the parameters of your study. Five parameter kinds are supported, freely mixed in one space: an explicit list of values (discrete), a ('continuous', low, high) range sampled on a fine grid (add 'log' for logarithmically spaced values), an ('integer', low, high) range, and ('nominal', ...) or ('ordinal', ...) categories. The parameter types guide walks through all of them.

Shaping the design (before running)

All of these are methods of Sampler, called before run():

set_design(n_samples, n_validation, extra_sets)

How many runs to place, how many validation points to reserve, and any additional named point sets. Guidance on choosing the size is in How large should a design be?.

add_prescribed(points)

Runs that already exist and must be kept in the design, for example from an earlier campaign.

load_design(points)

Resume from a complete previous design; run() then builds only what is missing around it.

add_focus(region, n)

A region of the space that deserves extra points.

add_exclusion(region) and feasibility constraints

Combinations that must never be proposed. See constraints and exclusions.

add_set(name, points)

Extra labelled point sets, such as a hand-picked test set.

set_optimizer(name, **options)

Tune the optimisation budget (iterations, restarts) of the chosen algorithm.

set_dimension_weights(weights)

Make some parameters count more than others in the distance calculations.

Running

Sampler.run(criteria, algorithm, seed, n_repeats, n_jobs)

Builds the design. Seven criteria and three algorithms are available; the criterion guide and the algorithm guide explain when to pick which.

Sampler.compare(criteria, algorithms, n_repeats)

When you are unsure, run several combinations under a shared baseline and get a ranked table; see comparing designs.

Working with the result

run() returns a result object with everything in one place:

summary() and quality_report()

The design counts, and the statistical quality evidence against a Monte Carlo baseline. How to read the report is covered in quality metrics.

plot(kind)

Eight plot kinds: 'pairplot', '1d', '2d', 'distances', 'correlation', 'quality', 'comparison', 'comparison_matrix'.

to_csv(), to_excel(), to_json(), to_markdown(), to_latex(), to_html() : Exports for running, sharing and publishing; see exporting results.

compare() returns a comparison object of its own, with summary(), plot() (the percentile heat map), to_markdown() and best_result.

Growing a design over time

The sequential module extends finished designs: extend adds new well-placed runs, fill_around concentrates new runs near interesting points, subsample picks a representative pilot subset, run_order orders runs so early stops still leave a balanced design, k_fold_split builds cross-validation folds, and nested creates designs within designs. The workflow is shown in sequential designs.