A Predictive Train Test Benchmark of LSE, MLE, PSO, and Hybrid PSO-LS for NHPP Software Reliability Growth Models
Journal of Computer Science · 01.11.2026 · 22 (11) · 3458-3468
DOI: 10.3844/jcssp.2026.3458.3468 ↗Kurz gefasst
Fachartikel in der Zeitschrift Journal of Computer Science, erschienen am 01.11.2026. Verfasst von Vy Dai Nghia und 2 weiteren Autor:innen.
Abstract (Originalsprache)
Non-homogeneous Poisson process software reliability growth models remain useful for estimating cumulative failures during software testing, but their practical accuracy depends strongly on parameter estimation. This paper presents a predictive train test benchmark of four estimation strategies: Nonlinear least squares estimation, maximum likelihood estimation, particle swarm optimization, and particle swarm optimization with local least-squares refinement. The study evaluates five NHPP mean value functions on twenty public software-failure datasets from the Lyu reliability data repository under chronological 70/30 and 80/20 train-test splits. For each dataset, model, and split, PSO-based methods are run with thirty independent random seeds. The results show a clear separation between fitting and prediction. PSO-LS achieves superior average ranks for training SSE and RMSE, demonstrating an improvement over standalone PSO for SSE-based calibration, while MLE gives the strongest held-out prediction, winning 71 of 100 held-out RMSE blocks under the 70/30 split and 68 of 100 under the 80/20 split. Friedman tests and Holm-corrected Wilcoxon comparisons support the main estimator differences, and the same conclusion holds when held-out negative log-likelihood and dataset-level rank checks are used. Within the tested Lyu-dataset setting, stated parameter bounds, and chronological splits, the evidence supports a narrow recommendation: Hybrid swarm-local search is valuable for SSE-oriented calibration, whereas likelihood-based estimation remains the more reliable predictive baseline under the NHPP assumption. The study also reports wall-clock runtime to support practical cost interpretation.
Zitieren
APA
Nghia, V. D., Nghia, P. T., & Cuong, N. H. (2026). A Predictive Train Test Benchmark of LSE, MLE, PSO, and Hybrid PSO-LS for NHPP Software Reliability Growth Models. Journal of Computer Science, 22(11), 3458-3468. https://doi.org/10.3844/jcssp.2026.3458.3468
Vancouver
Nghia VD, Nghia PT, Cuong NH. A Predictive Train Test Benchmark of LSE, MLE, PSO, and Hybrid PSO-LS for NHPP Software Reliability Growth Models. Journal of Computer Science. 2026;22(11):3458-3468. doi:10.3844/jcssp.2026.3458.3468
BibTeX
@article{Nghia2026,
author = {Nghia, Vy Dai and Nghia, Phung Trung and Cuong, Nguyen Hung},
title = {A Predictive Train Test Benchmark of LSE, MLE, PSO, and Hybrid PSO-LS for NHPP Software Reliability Growth Models},
journal = {Journal of Computer Science},
year = {2026},
volume = {22},
number = {11},
pages = {3458--3468},
doi = {10.3844/jcssp.2026.3458.3468}
}
EUVIMED