Software defects prediction by machine learning models optimized by metaheuristics

[T. Živković ]

146 str.

Elektrotehnika i racunarstvo - Racunarska tehnika i informatika / Electrical and Computer Engineering - Computer Science and Informatics

Testiranje softvera predstavlja kljucnu komponentu razvoja softvera i cesto je ono što pravi razliku izmežu uspešnih i neuspešnih projekata. Iako je izuzetno važno, zbog brzog tempa i kratkih rokova savremenih projekata, cesto se zanemaruje ili nije dovoljno detaljno zbog nedostatka vremena, što može dovesti do potencijalnog gubitka reputacije, podataka privatnih korisnika, novca, pa cak i ljudskih života u nekim situacijama. U takvim situacijama bilo bi od vitalnog znacaja imati mogucnost predvižanja koji softverski moduli su skloni defektima na osnovu skupa metrika softvera i fokusirati testiranje na njih, što je tipican zadatak klasifikacije. Modeli mašinskog ucenja cesto su uspešno korišceni za razlicite probleme klasifikacije, a u ovom radu se predlaže korišcenje Extreme Gradient Boosting (XGBoost) modela za izvršenje zadatka predvižanja softverskih defekata. Predložena je modifikovana varijanta dobro poznatog algoritma za optimizaciju, nazvanog algoritam pretrage reptila (engl. reptile search algorithm, skr. RSA), kako bi se izvršilo fino podešavanje hiperparametara XGBoost modela. Unapreženi algoritam nazvan je HARSA i evaluiran na kolekciji izazovnih funkcija CEC2019 za uporednu analizu (engl. benchmark), gde je pokazao izuzetne performanse. Kasnije je predstavljen XGBoost model koji koristi predloženi algoritam, i evaluiran je na dva skupa podataka za uporednu analizu za testiranje softvera. Rezultati simulacije su uporeženi sa drugim mocnim metaheuristickim algoritmima koji su korišceni u istom eksperimentalnom okruženju, pri cemu je predloženi pristup postigao superiornu tacnost klasifikacije na oba skupa podataka. Nakon toga je izvedena SHAP (engl. Shapley Additive Explanations) analiza kako bi se otkrili uticaji razlicitih metrika softvera na rezultate klasifikacije. Na kraju, razmotrena je i primena ovog rešenja u nastavi, uz osvrt na druga edukaciona okruženja koja se koriste u nastavi iz oblasti testiranja softvera, i uz konkretan primer laboratorijske vežbe koja studentima ilustruje proces razvoja modela za predvižanje softverskih defekata.

Software testing is a pivotal aspect of software development, often determining the success or failure of projects. However, amidst contemporary projects’ rapid pace and stringent deadlines, testing is frequently overlooked or insufficiently detailed due to time constraints. This negligence can result in potential repercussions such as damage to reputation, compromise of user data, financial loss, and in extreme cases, even human casualties. In such scenarios, the ability to anticipate software modules prone to defects based on software metrics becomes crucial, constituting a typical classification task. Machine learning models, renowned for their efficacy in addressing classification problems, offer a promising avenue for predicting software defects. In this dissertation, the utilization of the Extreme Gradient Boosting (XGBoost) model is advocated for this purpose. A modified iteration of the Reptile Search Algorithm (RSA), termed HARSA, is proposed for optimizing the hyperparameters of the XGBoost model. The efficacy of this enhanced algorithm is demonstrated through its exceptional performance on a suite of challenging benchmark functions from CEC2019. Subsequently, an XGBoost model employing HARSA is assessed on two software testing benchmark datasets, showcasing superior classification accuracy compared to other metaheuristic algorithms within the same experimental framework. Furthermore, Shapley Additive Explanations (SHAP) analysis is conducted to elucidate the impact of various software metrics on classification outcomes. Lastly, the educational implications of this solution are explored, contemplating its integration into software testing courses. A practical example of a laboratory exercise illustrates the process of developing a predictive model for software defects to students, fostering a deeper understanding of the subject matter.

testiranje softvera, predvižanje softverskih defekata, XGBoost, reptile algoritam za pretragu, optimizacija metaheuristikama

software testing, software defect prediction, XGBoost, Reptile search algorithm, metaheuristics optimization

004.421:004.85(043.3)

Tekst