Empirical Analysis of Fidelity in Multi-Fidelity Hyperparameter Optimization
Year: 2026
Type: article
Abstract: The performance of many machine learning algorithms depends on their hyperparameter settings. To reduce the high cost of evaluating hyperparameter configurations, various multi-fidelity hyperparameter optimization (MFHPO) methods have been developed. However, most existing methods either select low-fidelity levels uniformly or rely on expert knowledge, often overlooking the relationship between low- and high-fidelity performance and the associated evaluation costs, both of which are important for optimization efficiency. Given a machine learning algorithm with controllable fidelity parameters, how does its performance–cost relationship vary across fidelity levels, and do hyperparameter importance and the distribution of promising configurations change with fidelity? To address these questions, we analyze three machine learning algorithms—support vector machines, XGBoost, and multilayer perceptrons—across multiple datasets. Our experiments reveal that, as fidelity increases, the performance correlation, hyperparameter importance, and distribution of promising configurations at lower fidelities become increasingly similar to those at the highest fidelity. These findings suggest that selecting an appropriate fidelity level can reduce computational costs while maintaining a reasonable degree of consistency with high-fidelity evaluations. Moreover, an appropriate fidelity level may provide hyperparameter-importance estimates and promising-configuration distributions that are more consistent with those obtained at the highest fidelity.
Source: Algorithms
DOI: 10.3390/a19110997
Authors: Fan Li, Shifeng Huang, Jingye Hai, Ziyuan Zhou
Institutions: Central South University, Hunan University
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $1,930
OpenAlex ID: W7225875941
