Publication: Diseño automático de redes de neuronas artificiales para la predicción de series temporales
Loading...
Identifiers
Publication date
2012-01
Defense date
2012-05-07
Authors
Tutors
Journal Title
Journal ISSN
Volume Title
Publisher
Impact
Export
Abstract
El ser humano ha avanzado mucho, tecnológicamente hablando, en el último siglo. La sed por descubrir e innovar no tiene límites y cómo no, aplicar dichas innovaciones para nuestro provecho y bienestar general. Uno de los campos de investigación a los que se ha aplicado dicha innovación es la predicción. Al hablar de predicción, lo primero que nos puede venir a la cabeza son temas no tan científicos como la astrología o la lectura de manos, pero en realidad, diversos métodos estadísticos y matemáticos pueden ayudar a proporcionar información sobre el futuro. Uno de los ejemplos más comunes es la predicción del tiempo meteorológico que podemos observar cada día en televisión. Además de los métodos ya mencionados, en los últimos años ha proliferado el estudio de la predicción mediante técnicas de inteligencia computacional y dentro de las predicciones, aquellas que se ocupan de predecir series temporales. La predicción de series temporales consiste en llevar a cabo aproximaciones o estimaciones de qué valores tendrán los elementos futuros de una serie temporal partiendo de los valores de los elementos previos o ya conocidos. Como veremos en esta tesis doctoral, a lo largo de los años se han usado diferentes técnicas de inteligencia computacional con este propósito, aunque nosotros nos centraremos en las redes de neuronas artificiales. Plantearemos las ventajas y problemas que se pueden dar y nos centraremos en intentar solventar dichos problemas. Uno de los problemas clave que se plantea actualmente a la hora de aplicar redes de neuronas artificiales a cualquier dominio dado, es su correcto diseño. Estudiaremos pues las diferentes soluciones propuestas para el correcto diseño de las redes de neuronas artificiales, aunque terminaremos centrándonos en aquellas que hacen uso de la computación evolutiva. Este modelo de redes es el que se conoce como redes de neuronas artificiales evolutivas. Esta tesis doctoral presenta tres enfoques diferentes para el modelado automático de redes de neuronas artificiales. Cada enfoque irá destinado a solventar cada uno de lo que nosotros consideramos los tres puntos o problemas claves existentes al diseñar una red de neuronas artificial. El primer enfoque consistirá en el tratamiento de los datos que son pasados como patrones a la red para que ésta aprenda y sea evaluada. El segundo enfoque se centrará en las diferentes técnicas evolutivas que pueden ser usadas, cómo obtener un fenotipo a partir de un genotipo (y viceversa) y cómo evaluar una red. El último enfoque que se estudiará, es el tipo de arquitectura de red que debe ser usada para la predicción de series temporales. El objetivo final de esta tesis doctoral es llevar a cabo un sistema automático de diseño de redes de neuronas artificiales para solventar problemas de predicción de series temporales con la mayor exactitud posible y transparente al usuario, es decir, que este no tenga que ser un experto en la materia para poder hacer uso de él. --------------------------------------------------------------------------------------------------------------------------------------------------------------------------
The human being have progressed a lot, technologically speaking, in the last century. The thirst for discovery and innovation has no limits and of course, to apply these innovations to our benefit and general welfare. One of the research areas that have been applied to this innovation is the prediction. When we talk about prediction, the first thing that may come to our minds are not so scientific issues as astrology or hand reading, but in fact, several statistical and mathematical methods can help to provide information about the future. One of the most common examples is the weather forecasting that we can watch every day on television. Besides the methods already commented, in recent years it has proliferated the prediction study using computational intelligence techniques and within these predictions, those consisting of time series forecasting. Time series forecasting consist of carrying out approximations or estimations about which values will have the future elements of a time series starting from the values of the previous already known elements. As we discuss in this PhD thesis, over the years it has been used different computational intelligence techniques for this purpose, although we will focus on artificial neural networks.We will present the advantages and problems that may appear and we will focus on trying to solve these problems. One of the key problems that currently arise in applying artificial neural networks to any given domain, is its correct design. We will study then the different solutions proposed for the proper design of artificial neural networks, although at the end, we will focus on those which use evolutionary computation. This network model is known as evolutionary artificial neural networks. This PhD thesis presents three different approaches for the automatic design of artificial neural networks. Each approach will be dedicated to solve each of what we consider the three points or key problems in designing an artificial neural network. The first approach will consist of treating the data that are passed to the network as patterns to make it learn and be evaluated. The second approach will focus on the different evolutionary techniques that can be used, how to obtain a phenotype from a genotype (and vice versa) and how to evaluate a network. The last approach to be studied, is the type of network architecture to be used for time series forecasting. The ultimate goal of this thesis is to implement an automatic system to design artificial neural networks to solve time series forecasting problems as accurately as possible and transparent to the user, i.e. that the user did not have to be an expert to make use of it.
The human being have progressed a lot, technologically speaking, in the last century. The thirst for discovery and innovation has no limits and of course, to apply these innovations to our benefit and general welfare. One of the research areas that have been applied to this innovation is the prediction. When we talk about prediction, the first thing that may come to our minds are not so scientific issues as astrology or hand reading, but in fact, several statistical and mathematical methods can help to provide information about the future. One of the most common examples is the weather forecasting that we can watch every day on television. Besides the methods already commented, in recent years it has proliferated the prediction study using computational intelligence techniques and within these predictions, those consisting of time series forecasting. Time series forecasting consist of carrying out approximations or estimations about which values will have the future elements of a time series starting from the values of the previous already known elements. As we discuss in this PhD thesis, over the years it has been used different computational intelligence techniques for this purpose, although we will focus on artificial neural networks.We will present the advantages and problems that may appear and we will focus on trying to solve these problems. One of the key problems that currently arise in applying artificial neural networks to any given domain, is its correct design. We will study then the different solutions proposed for the proper design of artificial neural networks, although at the end, we will focus on those which use evolutionary computation. This network model is known as evolutionary artificial neural networks. This PhD thesis presents three different approaches for the automatic design of artificial neural networks. Each approach will be dedicated to solve each of what we consider the three points or key problems in designing an artificial neural network. The first approach will consist of treating the data that are passed to the network as patterns to make it learn and be evaluated. The second approach will focus on the different evolutionary techniques that can be used, how to obtain a phenotype from a genotype (and vice versa) and how to evaluate a network. The last approach to be studied, is the type of network architecture to be used for time series forecasting. The ultimate goal of this thesis is to implement an automatic system to design artificial neural networks to solve time series forecasting problems as accurately as possible and transparent to the user, i.e. that the user did not have to be an expert to make use of it.
Description
Keywords
Series temporales, Predicción, Neuronas artificiales, Diseño automático