Selection Heuristics on Semantic Genetic Programming for Classification ProblemsMaterial type: BookPublisher: Aguascalientes, 2020Description: 150 páginas; ilustraciones a color, 22 x 17 cm.Subject(s): Ciencia de Datos | Programación genética | Tecnologías de la información y comunicación
|Item type||Location||Collection||Call number||Status||Date due||Barcode|
|Tesis||Sección Tesis||Trabajos de Titulación INFOTEC||INFDCCD00002 S26 2020 (Browse shelf)||No para préstamo (Consulta únicamente en sala)||AGS20090120|
|Tesis||Sección Tesis||Trabajos de Titulación INFOTEC||INFDCCD00002 S26 2020 (Browse shelf)||No para préstamo (Consulta únicamente en sala)||DF-TLALPAN20090119|
Tesis que para obtener el grado de Doctora en Ciencias en Ciencia de Datos
In this dissertation, three heuristics for parent selection in Genetic Programming (GP) based on functions’ properties and individuals’ semantics have been proposed. Semantics have recently used for guiding the learning process when GP is used to solve supervised learning problems. However, to the best of our knowledge, this is the first time that functions’ properties are used for guiding the learning process in GP. The Pheuristics are tailored to the function , and the classifiers Naive Bayes and Nearest Centroid. The first heuristic based on cosine similarity promotes the selection of parents whose semantics are as perpendicular as possible between them in the semantics space. The second one, based on Pearson’s correlation coefficient, searches parents whose semantics’ vectors are uncorrelated. Finally, the last one, based on accuracy, tries to select parents whose semantics are different between them. In addition, we
analyze the use of completely random selection for parents and negative selection. These selection techniques were implemented on EvoDAG, a Semantic Genetic Programming system. For comparing the different selection schemes, we use 30 classification problems with a variable number of samples, variables, and classes. The results indicate that the combination of our heuristic based on accuracy for parent selection and negative random selection produces the best combination, and the difference in performances between this combination and the classical selection based on fitness is statistically significant. Furthermore, we compare our heuristics with state-of-the-art
schemes, Angle-Driven Selection, and Novelty Search. Besides, EvoDAG with the proposed selection heuristics was compared against 18 classifiers that included traditional approaches as well as auto-machine-learning techniques. We conclude that the use of our proposed heuristics significantly improves the learning process of EvoDAG.