Akademska digitalna zbirka SLovenije - logo
ALL libraries (COBIB.SI union bibliographic/catalogue database)
  • Explainable performance prediction of continuous single-objective optimization algorithms = Razložljivo napovedovanje uspešnosti optimizacijskih algoritmov za zvezno optimizacijo : master's thesis
    Nikolikj, Ana
    Although a large number of continuous single-objective optimization algorithms already exist, none of them can be considered to be dominant over all the others across all optimization problems. We ... have a limited understanding of such algorithm behavior and thus, the algorithms are treated as black-box systems themselves. Therefore, predicting the performance of an optimization algorithm for a new problem is a crucial task. Being able to predict the algorithms’ performance without actually running it on the problem, will further allow us to know if the selected algorithm is appropriate for the instance being solved. To this end, the most widely used performance prediction models based on supervised machine learning approaches are investigated, which try to link the problem landscape features to the algorithms’ performance. With the purpose of explaining algorithm behavior through the problem landscape features, we propose an explainable machine learning pipeline for algorithm performance prediction. The thesis aims to evaluate the usefulness of the problem landscape features using the latest techniques for feature selection and identify features that are the most informative of the algorithm behavior. The experimental results show that the most important feature portfolio is different for different algorithms. Further analysis of the identified portfolios, shows that for algorithms with similar performance on the problems, the regression models identify similar feature portfolios as important for the performance prediction task. Further, we propose a methodology that uses meta-representations that embed the problem landscape properties and the performance of the algorithm on the problems into the same vector space and tries to capture algorithm performance across the problem space. Clustering the meta-representations reveals regions of good and poor algorithm performance that define the algorithm footprint. From the results, we can conclude that the footprints make a clear distinction between good and poor algorithm performance, as meta-representations of similar algorithm behavior are grouped together. Finally, we introduce a method for evaluating the problem landscape feature representation and improved prediction capabilities for algorithm performance prediction. We observe better performance prediction results for a number of problems, however, there are also problems for which similar landscape feature representations are mapped to very different performances of the algorithm. This points out that the current feature representation has more power to discriminate between some of the problems and less for others.
    Type of material - master's thesis ; adult, serious
    Publication and manufacture - Ljubljana : [A. Nikolikj], 2023
    Language - slovenian, english
    COBISS.SI-ID - 165340931

No library within the COBISS.SI system holds a copy of this material.
loading ...
loading ...
loading ...

Shelf entry