ART-HPO: Adaptive Random Testing for Efficient Hyperparameter Optimization

Abstract

Hyperparameter optimization is critical for achieving optimal machine learning model performance, yet existing methods face significant limitations. Grid search suffers from exponential complexity, random search exhibits poor space coverage due to clustering, and Bayesian optimization introduces substantial computational overhead. We propose ART-HPO, a novel HPO method that adapts Adaptive Random Testing (ART) from software engineering to ensure diverse, systematic exploration of hyperparameter spaces. By selecting configurations that maximize distance from previously evaluated points, ART-HPO achieves superior space coverage compared to random search while maintaining minimal computational overhead. Our method handles mixed-type hyperparameters such as continuous, integer, and categorical through unified distance metrics and requires no surrogate model training. We evaluate ART-HPO across four benchmark datasets and four model families, demonstrating that it finds superior or competitive hyperparameter configurations.

Keywords: Hyperparameter Optimization, Adaptive Random Testing, Machine Learning