Online parameter adaptation methods for population-based metaheurisistics (Doctoral thesis)
Optimization problems lie in the core of scientific and technological development. They appear in almost every decision-making process, under various types and forms. A multitude of algorithms have been proposed in relevant literature to solve optimization problems. However, theoretical evidence suggests that the development of an overall optimal algorithm is impossible. For this reason, problemspecific optimization algorithms have been developed, incorporating a variety of features and ad hoc operations that exploit specific properties of the corresponding optimization problem. Typically, optimization algorithms have control parameters that adjust their dynamic with critical impact on their performance. Thus, proper parameter tuning becomes the cornerstone of efficient problem solving. There is a continuous line of research on parameter tuning methods since the early development of optimization algorithms. The majority of these methods addresses the tuning problem offline, i.e., prior to the algorithm’s execution. Established offline methods are based on statistical methodologies to identify promising parameter configurations, and their results may be reusable in problems of similar type. However, they neglect the algorithm’s feed- back and performance fluctuations during its run. The alternative approach is the use of online methods that dynamically adapt the parameters during the algorithm’s run. These methods exploit real-time performance data and, hence, they can make informative decisions on the parameter adaptation. This usually comes at the cost of non-reusable decisions. The main goal of the present thesis is the development of new online parameter adaptation methods that can be particularly useful for the class of metaheuristic optimization algorithms. The first part of the dissertation comprises the necessary background information on the current state-of-the-art and the optimization algorithms that will be used for demonstration purpose. In the second part of the thesis, two new online parameter adaptation methods are proposed. The first method, called Grid-based Parameter Adaptation Method, is based on grid search in the parameter space. The proposed method can be used on any algorithm and tackles both scalar and discrete parameters (including categorical ones). The new method is demonstrated on two state-of-the-art metaheuristics. For this purpose, two established benchmark suites are also considered. The second proposed method, called Gradientbased Parameter Adaptation Method with Line Search, replaces the grid search with approximate gradient search in the parameter space. The search procedure is further equipped with a recently proposed gradient-free line search technique. These modifications offer additional performance improvement with respect to the grid-based method, as revealed by the relevant performance assessment.
|Institution and School/Department of submitter:||Πανεπιστήμιο Ιωαννίνων. Πολυτεχνική Σχολή. Τμήμα Μηχανικών Ηλεκτρονικών Υπολογιστών και Πληροφορικής|
|Keywords:||Metaheuristics,Optimization,Online parameter adaptation,Parameter control,Differential evolution,Particle swarm optimization,Grid-search,Gradient estimations,Line search,Μεταευρετικοί,Βελτιστοποίηση,Online προσαρμογή παραμέτρων,Ρύθμιση παραμέτρων,Διαφοροεξελικτικός αλγόριθμος,Αλγόριθμος βελτιστοποίησης σμήνους σωματιδίων,Αναζήτηση πλέγματος,Προσεγγιστική αναζήτηση παραγώγων,Ευθύγραμμης αναζήτησης|
|Appears in Collections:||Διδακτορικές Διατριβές|
Files in This Item:
|Δ.Δ. ΤΑΤΣΗΣ ΒΑΣΙΛΕΙΟΣ 2019.pdf||2.1 MB||Adobe PDF||View/Open|
Please use this identifier to cite or link to this item:This item is a favorite for 0 people.
Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.