TPE Sampler with Categorical Distance

Class or Function Names

• TPESampler

APIs

• TPESampler(...)
  • This is Optuna’s TPE sampler with support for categorical_distance_func.
  • categorical_distance_func: A mapping from categorical parameter names to distance functions. Each function receives two choices and returns a non-negative distance.

Example

import optuna
import optunahub


module = optunahub.load_module(package="samplers/tpe_cat_distance")


def hamming_distance(x: str, y: str) -> float:
    return float(x != y)


sampler = module.TPESampler(
    categorical_distance_func={"optimizer": hamming_distance},
)


study = optuna.create_study(sampler=sampler)


def objective(trial):
    optimizer = trial.suggest_categorical("optimizer", ["adam", "sgd"])
    return float(optimizer != "adam")


study.optimize(objective, n_trials=20)

Categorical distance

This package is based on the Optuna 5.1.0.dev TPESampler and restores categorical-distance support using the Optuna 4.9 implementation as a behavioral reference.

Others

See the documentation for more details.