Create an isotropic (spherical) Gaussian component. Non-positive or non-finite variances are clamped to a small positive value so the covariance is a

Levy N.13. Domain: x1, x2 in [-10, 10]. Min: 0.

Load a study from a checkpoint file, reconstructing the engine from the embedded `StudyConfig`. The checkpoint must have been saved with the new form

Load a checkpoint from a reader with a byte-size cap and a cheap-fail format-version gate. The reader is bounded to [`MAX_CHECKPOINT_BYTES`].

Load checkpoint from a JSON file. Enforces a byte-size cap ([`MAX_CHECKPOINT_BYTES`]) and validates the recorded format version (cheaply, before the

A short, single-line snippet of `text` suitable for a log message.

Current cap on stored trials, if any.

Michalewicz. Domain: xi in [0, pi]. Min: depends on n.

ML-inspired grouped TLP benchmark. Hyper-parameters: n_estimators, max_depth, learning_rate, subsample. Performance group: {accuracy (R^2),

Simulates a model with accuracy/latency trade-off. More layers → better accuracy but higher latency. Larger hidden size → better accuracy but

Create a random strategy with the given seed for reproducibility.

Create a new Sobol' strategy with the given scrambling seed. The seed randomizes the sequence while preserving its low-discrepancy properties. Differ

Create a new Gaussian component with robustness to near-singular covariances. If the covariance matrix is not positive definite, regularization is ap

# Panics Panics if `choices` is empty.

Create a checkpoint and record the observation kind explicitly.

Repair the internal next-ID counter after deserializing or manually modifying a leaderboard.

The observation kind recorded in this checkpoint.

Return trials on the Pareto front, sorted by a scalarizer. Useful for getting an ordered list of Pareto-optimal solutions.

Compute the 0-indexed rank of a single trial without sorting or cloning the whole leaderboard. The rank is the number of stored trials that sort stri

Rastrigin. Domain: xi in [-5.12, 5.12]. Min: 0.

Reconcile a strategy produced by an off-lock refit with the strategy currently installed in the engine, before it is swapped back in. `self` is the f

Rosenbrock. Domain: xi in [-5, 10]. Min: 0.

Delete oldest checkpoint files to keep at most `max` files.

Start a HOLA server and yield its base URL. By default uses a simple 1D continuous space. Pass server_config marker to customize.

Save figure in PDF, PNG, and PGF formats.

Save checkpoint as compact JSON (no pretty-printing). Uses atomic write (write-to-temp + fsync + rename) to prevent data loss.

Schaffer N.2. Domain: x1, x2 in [-100, 100]. Min: 0.

Schaffer N.4. Domain: x1, x2 in [-100, 100]. Min: 0.292579.

Schwefel. Domain: xi in [-500, 500]. Min: 0.

Shubert. Domain: x1, x2 in [-10, 10]. Min: -186.7309.

Fake training function that returns a loss depending on the optimizer choice.

Create a single-component GMM (just a multivariate normal).

Six-Hump Camel. Domain: x1 in [-3, 3], x2 in [-2, 2]. Min: -1.0316.

Access the space configuration.

Synthetic grouped TLP benchmark. 4 raw objectives partitioned into 2 groups of 2. Decision variables: x0..x4 in [0, 1]. Group 1: f1 = ||

Every class in the stubs must exist at runtime.

__all__ must list every expected class.