Empirical studies on software effort estimation have employed hyper-parameter tuning algorithms to improve model accuracy and stability. While these tuners can improve model performance, some might be overly complex or costly for the low dimensionality datasets used in SEE. In such cases a method like random search can potentially provide similar benefits as some of the existing tuners, with the advantage of using low amounts of resources and being simple to implement. In this study we evaluate the impact on model accuracy and stability of 12 state-of-the-art hyper-parameter tuning algorithms against random search, on 9 datasets of the PROMISE repository and 4 sub-datasets from the ISBSG R18 dataset. This study covers 2 traditional exhaustive tuners (grid and random searches), 6 bio-inspired algorithms, 2 heuristic tuners, and 3 model-based algorithms. The tuners are used to configure support vector regression, classification and regression trees, and ridge regression models. We aim to determine the techniques and datasets for which certain tuners were 1) more effective than default hyper-parameters, 2) more effective than random search, 3) which models(s) can be considered "the best" for which datasets. The results of this study show that hyper-parameter tuning was effective (increased accuracy and stability) in 862 (51%) of the 1,690 studied scenarios. The 12 state-of-the-art tuners were more effective than random search in 95 (6%) of the 1,560 studied (non-random search) scenarios. Although not effective in every dataset, the combination of flash tuning, logarithm transformation and support vector regression obtained top ranking in accuracy on the highest amount (8 out of 13) of datasets. Hyperband tuned ridge regression with logarithm transformation obtained top ranking in accuracy on the highest amount (10 out of 13) of datasets. We endorse the use of random search as a baseline for comparison for future studies that consider hyper-parameter tuning.
Tipo de publicación: Conference Paper
Publicado en: International Conference on Predictable Models and Data Analytics in Software Engineering (PROMISE 21).Autores
- Leonardo Villalobos-Arias
- Quesada, Christian
Proyecto asociado a la publicación