Meta-Learn TPE: Task Similarity-Based Warm-Starting for TPE

Abstract

This package provides a meta-learning extension of the Tree-structured Parzen Estimator (TPE) that leverages previously completed Optuna studies on related tasks to accelerate optimization.

The algorithm is based on the approach described in:

• Speeding Up Multi-Objective Hyperparameter Optimization by Task Similarity-Based Meta-Learning for the Tree-Structured Parzen Estimator (IJCAI 2023)

The key idea is to compute task similarity between the target task and source tasks based on the overlap of their promising regions, and then use a weighted mixture of TPE models to guide the search. When source tasks are similar to the target, the sampler converges faster by exploiting shared structure.

Class or Function Names

• MetaLearnTPESampler

Arguments

Installation

$ pip install optuna

No additional dependencies are required beyond Optuna.

Example

import optuna
import optunahub


def objective(trial: optuna.Trial) -> float:
    x = trial.suggest_float("x", -5.0, 10.0)
    y = trial.suggest_float("y", 0.0, 15.0)
    return (x - 2) ** 2 + (y - 3) ** 2


# 1. Run source studies on related tasks.
source_study = optuna.create_study()
source_study.optimize(objective, n_trials=30)

# 2. Create the meta-learning sampler.
sampler = optunahub.load_module(
    package="samplers/meta_learn_tpe"
).MetaLearnTPESampler(
    source_studies=[source_study],
    seed=42,
)

# 3. Optimize a new (related) target task.
target_study = optuna.create_study(sampler=sampler)
target_study.optimize(objective, n_trials=50)
print(target_study.best_params)

How It Works

1. Build TPE models: For each source study and the target study, a TPE model is fitted, splitting trials into “below” (promising) and “above” (non-promising) groups.
2. Compute task similarity: The overlap between the target’s promising region and each source’s promising region is measured using Total Variation distance.
3. Weight tasks: Source tasks with higher similarity receive larger weights. The target task weight ensures that as optimization progresses, the sampler increasingly relies on the target data.
4. Weighted acquisition: Candidates are sampled from all tasks’ below distributions, and scored using a weighted mixture of all tasks’ TPE likelihoods.

Bibtex

@inproceedings{watanabe_meta_learn_tpe_2023,
    title={Speeding Up Multi-Objective Hyperparameter Optimization by Task Similarity-Based Meta-Learning for the Tree-Structured Parzen Estimator},
    author={Watanabe, Shuhei and Awad, Noor and Onishi, Masaki and Hutter, Frank},
    booktitle={International Joint Conference on Artificial Intelligence},
    year={2023}
}