Module belief 

What a run has learned about the function, as the page draws it (D95).

Three views, each computed from public atune state only:

• Belief fits atune’s Gaussian process ([atune::gp::Gp], the model behind GP-EI, with its closed-form median-heuristic hyperparameters) to the run’s completed trials, collected exactly as GP-EI collects them. For a GP-EI run it is the model GP-EI fits for its next proposal; for any other sampler it is the same model applied to that run’s results, not the sampler’s own. The page maps its posterior mean, standard deviation, or expected improvement.
• tpe_good is the set of trials TPE’s density l(x) is fitted to: the min(⌈0.1·n⌉, 25) lowest values, by TPE’s own (linear) split rule.
• ellipse traces CMA-ES’s search distribution from the sampler’s own mean, step size and covariance (Controller::cmaes_distribution).

Structs

Belief

A Gaussian process fitted to a run’s completed trials.

Enums

ModelMap

What the model map shows.

Constants

GP_WARMUP

GP-EI’s warmup: below this many completed trials it draws space-filling points instead of consulting its model (the earlier search’s floor for two parameters, which the page builds GP-EI with).

MIN_POINTS

The fewest completed trials the page fits a model to.

TPE_SPLIT_CAP

The cap on TPE’s linear split.

TPE_SPLIT_DIVISOR

TPE’s linear split: n_good = min(⌈n / TPE_SPLIT_DIVISOR⌉, TPE_SPLIT_CAP) (Tpe::new’s default, Optuna’s default_gamma).

TPE_STARTUP

TPE’s uniform startup trials before its model takes over (Tpe::new’s default).

Functions

ellipse

The ellipse of points sigmas standard deviations from the mean of CMA-ES’s search distribution, as segments unit-square points.

tpe_good

The trial numbers in TPE’s “good” set for its next proposal, lowest value first, or None when the run is not an unscheduled TPE run past its startup.

training_data

GP-EI’s training pairs for the controller’s study (see Belief::new).