Meta-Surrogate Benchmarking for Hyperparameter Optimization

TL;DR

This paper introduces a meta-surrogate model for hyperparameter optimization benchmarks, enabling faster, more realistic, and statistically robust comparisons of HPO methods by generating synthetic problem instances.

Contribution

It proposes a novel meta-surrogate approach that creates inexpensive, realistic HPO tasks, facilitating large-scale, reproducible benchmarking and analysis.

Findings

Benchmarking on generated tasks yields more statistically significant results.

The method accelerates benchmarking by orders of magnitude.

The approach is effective across various HPO methods and problem classes.

Abstract

Despite the recent progress in hyperparameter optimization (HPO), available benchmarks that resemble real-world scenarios consist of a few and very large problem instances that are expensive to solve. This blocks researchers and practitioners not only from systematically running large-scale comparisons that are needed to draw statistically significant results but also from reproducing experiments that were conducted before. This work proposes a method to alleviate these issues by means of a meta-surrogate model for HPO tasks trained on off-line generated data. The model combines a probabilistic encoder with a multi-task model such that it can generate inexpensive and realistic tasks of the class of problems of interest. We demonstrate that benchmarking HPO methods on samples of the generative model allows us to draw more coherent and statistically significant conclusions that can be reached orders of magnitude faster than using the original tasks. We provide evidence of our findings for various HPO methods on a wide class of problems.