Re-use the best trial

After tuning, you often want to run the objective once more with the best hyperparameters — to compute extra metrics, to retrain on the full dataset, or to analyze the winning model in detail. Study.BestTrial is a FrozenTrial whose stored parameters can be replayed: passing it to a function that calls Suggest* returns the recorded values instead of sampling new ones.

Tune, then re-evaluate

First a normal optimization:

using OptunaSharp;

double Objective(BaseTrial trial)
{
    double c = trial.SuggestFloat("C", 1e-7, 10.0, log: true);
    return TrainAndScore(c);          // accuracy on the test split
}

var study = Optuna.CreateStudy(direction: StudyDirection.Maximize);
study.Optimize(Objective, nTrials: 10);

Console.WriteLine(study.BestTrial.Value);   // the best accuracy

Note the objective takes BaseTrial, the shared base of both Trial (the live handle) and FrozenTrial (a stored snapshot). That one signature lets the same method be driven by the study during tuning and be replayed from a frozen trial afterward.

Now define a second function that reproduces the winning model and computes more metrics. It shares the Suggest* calls, so replaying study.BestTrial rebuilds exactly the best configuration:

(double acc, double f1, double recall, double precision) DetailedObjective(BaseTrial trial)
{
    double c = trial.SuggestFloat("C", 1e-7, 10.0, log: true);   // replayed from the frozen trial
    var model = Train(c);
    return (model.Accuracy, model.F1, model.Recall, model.Precision);
}

// Pass the frozen best trial — its stored C is returned by SuggestFloat above.
var metrics = DetailedObjective(study.BestTrial);

BestTrial behaves differently from a live trial

A FrozenTrial is not an active trial, so some methods are inert on it — in particular ShouldPrune() always returns false and Report(...) does nothing. That's exactly what you want when replaying: there is no sampler or pruner in the loop, just the stored parameters.

Multi-objective studies

For a multi-objective study there is no single best trial; Study.BestTrials returns the list of FrozenTrials on the Pareto front. Replay any of them the same way:

foreach (FrozenTrial t in study.BestTrials)
    Analyze(DetailedObjective(t));

Relationship to Optuna

Optuna (Python)	OptunaSharp
study.best_trial	study.BestTrial
detailed_objective(study.best_trial) (objective takes a trial)	DetailedObjective(study.BestTrial) (objective takes BaseTrial)
study.best_trial.value	study.BestTrial.Value
study.best_trials (Pareto front)	study.BestTrials

See also

Study and Trials · Multi-objective optimization · Specify hyperparameters manually · the API reference.