Μπεϋζιανή βελτιστοποίηση για τη ρύθμιση υπερπαραμέτρων σε αλγορίθμους μηχανικής μάθησης
View/
Keywords
Μπεϋζιανή Βελτιστοποίηση ; Μηχανική μάθηση ; Υπερπαράμετροι ; Gaussian Process ; Γκαουσιανές ΔιαδικασίεςAbstract
This master’s thesis investigates Bayesian Optimization for handling computationally expensive black-box functions. The core research question focuses on evaluating custom implementations —based on Gaussian Processes with various kernels and acquisition functions— against established automated machine learning libraries, specifically Optuna and HyperOpt.
The main scientific contribution lies in the development of a fully customizable optimization frame-work and its extensive performance analysis. The methodology was evaluated on mathematical bench-mark functions of varying dimensions under noisy conditions, as well as on a real-world hyperparam-eter tuning task for the Agglomerative Clustering algorithm.
The results reveal a distinct trade-off between accuracy and computational efficiency. The custom im-plementations consistently outperform standard libraries in terms of solution precision, locating values remarkably close to the global optimum. Conversely, standard libraries offer vastly superior execution speed and scalability as the problem dimensionality increases.
The application domains of these outcomes span across Data Science for automated model tuning, as well as industrial optimization problems where evaluating the objective function is financially or com-putationally prohibitive.
Postgraduate Studies Programme
Πληροφοριακά Συστήματα και ΥπηρεσίεςDepartment
Σχολή Τεχνολογιών Πληροφορικής και Επικοινωνιών. Τμήμα Ψηφιακών ΣυστημάτωνNumber of pages
63Language
GreekCollections
Except where otherwise noted, this item's license is described as
Αναφορά Δημιουργού-Μη Εμπορική Χρήση-Όχι Παράγωγα Έργα 3.0 Ελλάδα
Αναφορά Δημιουργού-Μη Εμπορική Χρήση-Όχι Παράγωγα Έργα 3.0 Ελλάδα