| 日時(Date) | 2023年1月19日（木）4限（15:10--16:40) Thu. Jan. 19th, 2023, 4th period (15:10--16:40) |
|---|---|
| 場所(Location) | face-to-face, エーアイ大講義室, AI Inc. Seminar Hall （L1） |
| 司会(Chair) | Bodin Chinthanet |
| 講演者(Presenter) | PASSAKORN PHANNACHITTA (Chiang Mai University) |
| 題目(Title) | A story behind an optimal model-based software effort estimation |
| 概要(Abstract) | Software effort estimation is one of the long-standing debate problems in empirical software engineering involving machine-learning models. The debates were upon what is the best estimation solution. In the early stage, there were no clear answers due to the unavailability of a reliable or trustworthy assessment framework. Numerous research works have nominated widely different techniques as the single best solution. It took a decade until the research community could invent a sufficiently reliable enough assessment framework. The framework suggests using a proper statistically significant testing method and enhancing the studies' generalizability power. However, the discovery of the best estimation method was found to be still a long way. Unlike other research areas at the time, where trends were already to optimizing machine-learning models for best tackling numerous problems, it was still unclear whether machine-learning models were suitable for estimating the software effort. It was even if powerful techniques such as developing a stacked ensemble were assessed. According to the speaker, the main problem was that the choice of algorithms discussed in empirical software engineering circles is still a laggard. Powerful algorithms such as deep learning and gradient-boosting machines had not yet been thoroughly assessed even though they were widely adopted and frequently achieved promising results in many competitions, such as those million-dollar prizes hosted by Kaggle. In a subsequence study, the speaker assessed 14 machine-learning algorithms widely adopted in the data science communities and could provide substantial evidence to hint that the successful algorithms in the data science circle can be the best estimation solution when data are available. To provide conclusion stability for the long debates, the speaker revisited all the state-of-the-art techniques and could suggest a novel strategy to construct an accurate estimator by adopting a combined estimator integrating the significant features of the current state-of-the-art at the mathematical formula level. This combined technique is not as simple as the standard combined techniques, making a stacked ensemble of multiple estimators and claiming a lower variance estimation. As confirmed by the reliable and trustworthy assessment framework, the speaker's adopted technique significantly outperformed state-of-the-art effort estimators, including many hard-to-beat techniques based on a stacked ensemble and that of an automatically transformed linear model. The critical takeaway is that it is crucial to collaborate with human experts in other areas and form multidisciplinary research work. The conclusion stability of the research work of this talk would still be a long way if the speaker stuck only to the empirical software engineering circles and did not try his best to expand his network through those in meteorology, agriculture, bioinformatics, and finance. |
| 講演言語(Language) | English |
| 講演者紹介(Introduction of Lecturer) |
Education: Job records: |