Publicação
Geometric semantic inspired mutation for M3GP
| Resumo: | One of the most challenging Machine Learning tasks is multiclass classification. Genetic Programming (GP) is not able to achieve a very good performance when applied to classification problems with number of classes bigger than two. However, Multidimensional Multiclass Genetic Programming (M2GP) and Multidimensional Multiclass Genetic Programming with Multidimensional Populations (M3GP), two wrapper-based GP classifiers, have shown to be competitive with state-of-the-art classifiers. The main focus of this work is a new version of M3GP, called Geometric Semantic In- spired M3GP (GSI-M3GP), inspired in geometric semantic operators. GSI-M3GP works in the same way as M3GP, but uses only three operators to create new individuals: add branch, remove branch and a new mutation operator called geometric semantic inspired mutation (gsimutation). In order to test GSI-M3GP and compare it to M3GP, an implementation in Java was developed. Nine different versions of GSI-M3GP were created and tested on eight benchmark problems. For most of the versions of GSI-M3GP, the new algorithm is competitive with M3GP on all the problems. Additionally, it was tested if adding a crossover operator would improve the results, which it did not. A few other alterations were made to the original M3GP algorithm to test the possibility of using the Euclidean distance, instead of the Mahalanobis distance, without harming the quality of the solutions. These alterations do not always maintain the quality of the solutions. |
|---|---|
| Autores principais: | Pinto, Ana Sofia Brás |
| Assunto: | Machine Learning Multiclass Classification Genetic Programming Semantic Genetic Programming Aprendizagem Automática Classificação em mais de duas classes |
| Ano: | 2019 |
| País: | Portugal |
| Tipo de documento: | dissertação de mestrado |
| Tipo de acesso: | acesso aberto |
| Instituição associada: | Universidade Nova de Lisboa |
| Idioma: | inglês |
| Origem: | Repositório Institucional da UNL |
| Resumo: | One of the most challenging Machine Learning tasks is multiclass classification. Genetic Programming (GP) is not able to achieve a very good performance when applied to classification problems with number of classes bigger than two. However, Multidimensional Multiclass Genetic Programming (M2GP) and Multidimensional Multiclass Genetic Programming with Multidimensional Populations (M3GP), two wrapper-based GP classifiers, have shown to be competitive with state-of-the-art classifiers. The main focus of this work is a new version of M3GP, called Geometric Semantic In- spired M3GP (GSI-M3GP), inspired in geometric semantic operators. GSI-M3GP works in the same way as M3GP, but uses only three operators to create new individuals: add branch, remove branch and a new mutation operator called geometric semantic inspired mutation (gsimutation). In order to test GSI-M3GP and compare it to M3GP, an implementation in Java was developed. Nine different versions of GSI-M3GP were created and tested on eight benchmark problems. For most of the versions of GSI-M3GP, the new algorithm is competitive with M3GP on all the problems. Additionally, it was tested if adding a crossover operator would improve the results, which it did not. A few other alterations were made to the original M3GP algorithm to test the possibility of using the Euclidean distance, instead of the Mahalanobis distance, without harming the quality of the solutions. These alterations do not always maintain the quality of the solutions. |
|---|