Evict the oldest trials until at most `cap` remain. Rebuilds the id index only when at least one trial is removed. With a single eviction per push th

Convert a `BTreeMap<String, f64>` to a JSON object, representing infinity as `"inf"`.

Generate the filename for a checkpoint at the given trial count.

Fit GMM parameters from normalized samples using EM algorithm. # Arguments `samples` - Normalized samples in [0, 1]^n, each inner Vec is one sample `

Get a single completed trial by ID, computing its rank and Pareto front.

Derive a stable group key from an objective's group label. Falls back to the field name when no explicit group is set.

Deserialize a checkpoint JSON file with a byte-size cap and a cheap-fail format-version gate. The version is validated from the raw bytes before the

Load all multi-objective results.

Load all single-objective results.

Perform NSGA-II non-dominated sorting including infeasible trials.

Whether observation `a` dominates `b` (no worse in any group, strictly better in at least one), assuming minimization. Mirrors the leaderboard's own d

Compute the Pareto front including infeasible trials.

Get the Pareto front with crowding distances. Returns feasible trials on the Pareto front along with their crowding distances, sorted by crowding dis

Bar plot of mean normalized score per optimizer, grouped by budget.

Get all trials (including infeasible) with NSGA-II ranking information.

Read up to `MAX_CAPTURE_BYTES` from `reader`, discarding the rest so a runaway producer cannot stall the worker. Returns the captured text and a flag

Re-compute observations for all trials that have raw_metrics. This enables mid-run objective changes: update the scoring function, call `rescalarize`

Run optimization and track convergence.

Save a full checkpoint (leaderboard + strategy state). This is the stable persistence API. Use `save` / `load` for checkpointing.

Start the server with explicit host, auth, CORS, and checkpoint options.

Sort a Vec<CompletedTrial> by the given criterion.

Get parameter metadata for dashboard auto-configuration.

Best-effort fsync of the directory containing `path` so a preceding rename is durable. Opening and syncing a directory is unsupported on some platform

Target-Limit-Priority (TLP) scalarization for a single value. Direction is inferred from the relationship between `target` and `limit`: - `target < l

Return top-k trials as (candidate, scalarized_score) for strategy refit.

Monotonic count of trials ever pushed, forwarded from the inner `Leaderboard`. Increments on every push and is never decremented by eviction, so for a

Reconstruct a component from on-disk data, or `None` if the data is malformed (zero-length, mismatched covariance dims, or non-finite values). `new` r

Two priority-group objectives that force a vector leaderboard.

Classify a code block as 'exec', 'syntax', or 'skip'.

Yield pytest params for all Python code blocks in docs.

Extract (line_number, code) tuples from a markdown file.

Build a namespace with all hola exports pre-imported.

Execute one multi-objective run.

Execute one single-objective run. Must be top-level for pickling.

Produce metrics with occasional extreme values.

Mean normalized score per (optimizer, budget), averaging across problems. This is the "mean of means" metric used in the paper.

Return the best trial including infeasible ones.

Get the best feasible trial for a single objective (ignoring others).

Return the k worst trials, including infeasible.

Branin-Hoo. Domain: x1 in [-5, 10], x2 in [0, 15]. Min: 0.397887.

Canonicalize the nearest existing ancestor of `path` and re-append the remaining (not-yet-created) components, resolving any symlinks along the existi

Validate that a recorded format version matches the version this build understands. Returns a clear error on mismatch.

Return set of (problem, optimizer, budget, run_id) already completed.

Return set of (problem, optimizer, budget, run_id) already completed.

Compute hypervolume dominated by front relative to reference point. Higher is better. Returns 0.0 if front is empty.

Compute IGD from true front to approximation. Lower is better. Returns inf if front is empty.

Compute per-priority-group aggregated scores from raw metrics.

Compute spacing metric (std of nearest-neighbor distances). Lower is better (more uniform). Returns NaN for fewer than 2 points so that degen

Return the path to the bundled dashboard static files. Useful for passing to ``Study.serve(dashboard_path=str(dashboard_dir()))``. Returns a

Format-agnostic objective scalarization math. These are the pure mathematical operations that map raw metric values to optimization scores. They have

Enforce `max_size` after a push, evicting the oldest trial if needed. A push adds exactly one trial, so at most one eviction is required. The `cap >

Optimize the 1D Forrester function.