Terminators

A terminator stops the whole study once more trials are unlikely to help — as opposed to a pruner, which stops a single unpromising trial. It compares the potential for further improvement against the noise in your evaluation, and calls it done when improvement no longer clears the noise. OptunaSharp ports optuna.terminator into OptunaSharp.Terminator.

Deprecated in Optuna. The optuna.terminator module is deprecated as of Optuna v4.9.0 and is scheduled for removal in Optuna v6.0.0, with its functionality planned to move to hub.optuna.org. OptunaSharp ports it for parity and reproduces the deprecation: Terminator, TerminatorCallback, and ReportCrossValidationScores emit a warning when used. It remains functional; prefer a fixed trial budget or a pruner for new code.

How termination is decided

A Terminator combines two evaluators:

an improvement evaluator — estimates the remaining room to improve (a regret bound); and
an error evaluator — estimates the statistical noise floor of your objective.

When the estimated improvement drops below the estimated error (after a minimum number of trials), the study should stop.

public Terminator(
    BaseImprovementEvaluator? improvementEvaluator = null,  // default: RegretBoundEvaluator
    BaseErrorEvaluator? errorEvaluator = null,              // default: CrossValidationErrorEvaluator
    int minNTrials = 20);

public abstract bool ShouldTerminate(Study study);         // the decision

Using it in a run

The usual way is TerminatorCallback, which checks after each trial and stops the study:

using OptunaSharp;
using OptunaSharp.Terminator;

var callback = new TerminatorCallback();   // wraps a default Terminator
study.Optimize(Objective, nTrials: 400, callbacks: new[] { callback.Invoke });
// The study stops early once the terminator is satisfied, even though nTrials allows 400.

Or query a Terminator directly in your own loop:

var terminator = new Terminator(minNTrials: 50);
while (!terminator.ShouldTerminate(study))
{
    Trial trial = study.Ask();
    study.Tell(trial, Evaluate(trial));
}

Improvement evaluators

Evaluator	Idea	Constructor
`RegretBoundEvaluator` (default)	Gaussian-process regret bound over the top trials	`(topTrialsRatio: 0.5, minNTrials: 20, seed: null)`
`BestValueStagnationEvaluator`	Stop after N trials with no new best	`(maxStagnationTrials: 30)`
`EMMREvaluator`	Expected-minimum model regret (a GP criterion)	`(deterministicObjective: false, delta: 0.1, minNTrials: 2, seed: null)`

RegretBoundEvaluator and EMMREvaluator build on the same Gaussian-process stack as the GPSampler; their output tracks Optuna to a documented tolerance (see below).

Error evaluators

Evaluator	Idea	Constructor
`CrossValidationErrorEvaluator` (default)	Scale of the best trial's cross-validation scores	`()`
`StaticErrorEvaluator`	A fixed error floor you supply	`(double constant)`
`MedianErrorEvaluator`	Median improvement × a ratio	`(BaseImprovementEvaluator pairedImprovementEvaluator, warmUpTrials: 10, nInitialTrials: 20, thresholdRatio: 0.01)`

To use cross-validation error, report the per-fold scores during the trial:

TerminatorFunctions.ReportCrossValidationScores(trial, new[] { 0.91, 0.89, 0.90, 0.92, 0.88 });

Relationship to Optuna

Optuna (Python)	OptunaSharp
`Terminator(min_n_trials=50)`	`new Terminator(minNTrials: 50)`
`study.optimize(obj, callbacks=[TerminatorCallback()])`	`study.Optimize(obj, callbacks: new[] { new TerminatorCallback().Invoke })`
`RegretBoundEvaluator(top_trials_ratio=0.5)`	`new RegretBoundEvaluator(topTrialsRatio: 0.5)`
`report_cross_validation_scores(trial, scores)`	`TerminatorFunctions.ReportCrossValidationScores(trial, scores)`

Divergence. The GP-based evaluators (RegretBoundEvaluator, EMMREvaluator) inherit the GP stack's tolerance boundary: the managed kernel fit drifts the hyperparameters by ~1e-3, which propagates through the criterion, so their values match Optuna to a few 1e-3 relative rather than bit-for-bit. The termination decision agrees with Optuna in tests within that tolerance.