Strategies for reducing the size of the search space in semantic genetic programming

• Ano de defesa: 2018
• Autor(a) principal: Luis Fernando Miranda
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Universidade Federal de Minas Gerais
• Palavras-chave em Português: Programação Genética ( Computação); Computação; Supervised Learning; Instance Selection; Geometric Semantic Genetic Programming; Symbolic Regression
• Link de acesso: https://hdl.handle.net/1843/ESBF-B5UMN3

Genetic programming (GP) uses bio-inspired operations to find adequate solutions to a diverse range of problems, including symbolic regression. Recent works introduced semantic awareness into these operations, enhancing the search. In particular, Geometric Semantic GP (GSGP)-the method taken as reference in this context-exploits geometric properties describing the relationship between possible solutions in an n-dimensional semantic space, where n is the number of training instances. However, in problems where n is high-a common scenario in real applications-the search can become excessively complicated. We tackle this problem by reducing the dimensionality of the semantic space through instance selection (IS) methods. We present two distinct strategies: the first is implemented as a pre-processing step while the second incorporates IS to the search. Our results showed that IS methods can improve the search performed by GSGP, resulting in a faster search for models with smaller errors.