Nuevas variantes de sistemas difusos genéticos para resolver problemas de regresión de alta dimensionalidad
Archivos
Fecha
2013-07-04
Autores
García Martínez, Victor Manuel
Título de la revista
ISSN de la revista
Título del volumen
Editor
Universidad Central “Marta Abreu” de Las Villas
Resumen
Los sistemas difusos genéticos surgen como una fusión entre la Lógica Difusa y la
Computación Evolutiva. Un sistema difuso genético es básicamente un sistema difuso con un
proceso de aprendizaje incorporado basado en algoritmos genéticos. Se considera que la
interpretabilidad (medida frecuentemente como el número de reglas generadas) de los
sistemas difusos genéticos es su principal ventaja competitiva en comparación con otras
técnicas por lo cual ha recibido atención especial. La obtención de un alto grado de
interpretabilidad es contradictoria ante el objetivo de disminuir la precisión (obtenida como el
error) por lo que es necesario alcanzar un buen balance entre ambos objetivos.
Los sistemas difusos genéticos han sido utilizados para resolver diversos problemas, entre
ellos, los problemas de regresión de alta dimensionalidad. La resolución de problemas de alta
dimensionalidad puede influir en el decremento de la interpretabilidad de los modelos
obtenidos. La alta dimensionalidad en los datos implica una mayor utilización de los recursos
computacionales en la solución del problema.
En este trabajo se realiza un estudio del algoritmo Embedded Genetic Learning of Highly
Interpretable Fuzzy Partitions (EGLFP) diseñado para resolver problemas de regresión de alta
dimensionalidad y se analizan las partes del mismo que provocan mayor número de
evaluaciones. En base a este estudio se proponen tres variantes que mejoran el tiempo de
ejecución del algoritmo original, logrando un buen balance entre interpretabilidad y precisión.
Genetic fuzzy systems emerge as a fusion between Fuzzy Logic and Evolutionary Computation. A genetic fuzzy system is basically a fuzzy system with a learning process incorporated based on genetic algorithms. The interpretability (often measured as the number of rules generated) of genetic fuzzy systems is their main competitive advantage in comparison to other techniques, so it has received special attention. Obtaining a high degree of interpretability is contradictory to the aim of reducing the accuracy (measured as the error), so it is necessary obtain a good trade of between these two objectives. Genetic fuzzy systems have been used to solve several problems, including high-dimensional regression problems. Solving high-dimensional problems may affect the interpretability of the obtained models. The high dimensionality in the data implies greater use of computational resources in the solution the problem. In this work we make a study of Embedded Genetic Learning of Highly Interpretable Fuzzy Partitions (EGLFP) algorithm, designed to solve high-dimensional regression problems, and we analyze the parts of it that cause the largest number of evaluations. Based on this study, we propose three alternatives that improve the execution time of the original algorithm, achieving a good trade of between interpretability and accuracy.
Genetic fuzzy systems emerge as a fusion between Fuzzy Logic and Evolutionary Computation. A genetic fuzzy system is basically a fuzzy system with a learning process incorporated based on genetic algorithms. The interpretability (often measured as the number of rules generated) of genetic fuzzy systems is their main competitive advantage in comparison to other techniques, so it has received special attention. Obtaining a high degree of interpretability is contradictory to the aim of reducing the accuracy (measured as the error), so it is necessary obtain a good trade of between these two objectives. Genetic fuzzy systems have been used to solve several problems, including high-dimensional regression problems. Solving high-dimensional problems may affect the interpretability of the obtained models. The high dimensionality in the data implies greater use of computational resources in the solution the problem. In this work we make a study of Embedded Genetic Learning of Highly Interpretable Fuzzy Partitions (EGLFP) algorithm, designed to solve high-dimensional regression problems, and we analyze the parts of it that cause the largest number of evaluations. Based on this study, we propose three alternatives that improve the execution time of the original algorithm, achieving a good trade of between interpretability and accuracy.
Descripción
Palabras clave
Sistemas Difusos Genéticos, Problemas de Regresión, Interpretabilidad, Alta Dimensionalidad