A hyper-solution framework for classification problems via metaheuristic approaches

Abstract

This is a summary of the author’s PhD thesis, supervised by Prof. Domenico Conforti and defended on 26-02-2010 at the Universitá della Calabria, Cosenza. The thesis is written in Italian and a copy is available from the author upon request. This work deals with the development of a high-level classification framework which combines parameters optimization of a single classifier with classifiers ensemble optimization, through meta-heuristics. Support Vector Machines (SVM) is used for learning while the meta-heuristics adopted and compared are Genetic-Algorithms (GA), Tabu-Search (TS) and Ant Colony Optimization (ACO). Single SVM optimization usually concerns two approaches: searching for optimal set up of a SVM with fixed kernel (Model Selection) or with linear combination of basic kernels (Multiple Kernel Learning), both issues were considered. Meta-heuristics were used in order to avoid time consuming grid-approach for testing several classifiers configurations and some ad-hoc variations to GA were proposed. Finally, different frameworks were developed and then tested on 8 datasets providing reliable solutions.