Single, multi- and many-objective meta-heuristic algorithms applied to pattern recognition

Abstract

In the last few years, metaheuristic algorithms have been used for solving several problems in engineering, biology, physics, among others, since many of them can be modeled as being optimization tasks. Metaheuristic methods simulate social dynamics and physical phenomena such as the interaction among bats, some species of birds, insects or even gravitational force. Although these metaheuristic techniques are commonly applied to solve single-objective problems, they are also being used to solve multi- and many-objective problems, where the idea of a single global optimal solution is replaced by the concept of Pareto-front. In computer vision and pattern recognition areas, little effort has been dedicated to multi-objective optimization using metaheuristics. As such, this thesis aims at studying and developing new mono, multi- and many-objective versions of metaheuristic techniques in the context of machine learning, which include, among other areas, feature combination and selection, parameter optimization of machine learning techniques and deep learning.