Publication: Analisis de técnicas para medir la influencia de las variables del vector de entrada de un perceptron multicapa
Loading...
Identifiers
Publication date
2012-05
Defense date
2012-07-16
Authors
Advisors
Tutors
Journal Title
Journal ISSN
Volume Title
Publisher
Impact
Export
Abstract
El objetivo primordial de cualquier proyecto relacionado con la simulación es obtener resultados fiables a partir de experimentos planteados. La principal cualidad que debe cumplir cualquier simulación reside en lograr resultados fidedignos con un menor coste de recursos (tiempo, personal, monetario . . .) que los que acarrearía la realización del experimento propiamente dicha. Centrándonos en el ámbito que nos ocupa, el problema que se presenta en el entorno de los impactos en materiales es su no linealidad, y el elevado coste que requiere su estudio en modo experimental. Por ello, se trata de resolver estos problemas mediante una herramienta capaz de obtener resultados válidos mediante simulaciones, lo que reduce de forma drástica los costes asociados al análisis de los impactos. Para este proyecto se ha creado una herramienta, en adelante Simulador, que automatiza el entrenamiento y testeo de una RNA, así como la generación de informes y resúmenes que permiten extrapolar datos del comportamiento de una red en base a su entrenamiento y datos de entrada. Con ayuda de esta información se ha buscado la mejor configuración de la RNA para el problema que nos atañe. Uno de los principales objetivos de este proyecto es ayudar al lector a comprender el funcionamiento de la red mediante el análisis de los resultados que esta proporciona. Es decir, como influyen en los resultados la cantidad y variedad de los datos de entrada, así como la configuración de los diferentes métodos de entrenamiento. Otro de los objetivos de todo proyecto de simulación es conocer la influencia que tienen cada una de las variables de entrada en los resultados obtenidos por la RNA. En este proyecto se utilizan dos métodos de estudio para la valoración de la criticidad de las variables de entrada “Fortaleza frente a valores atípicos” y “Sensitivity about the mean”. Se va a analizar y confrontar los resultados de ambos métodos para ver si sus resultados coinciden, y de no ser así, valorar cuál de los métodos parece más fiable. ____________________________________________________________________________________________________________________
The primary objective of any project related to the simulation is to obtain reliable results from experiments presented. The best quality to be met by any simulation lies in achieving reliable results with a lower cost of resources (time, personnel, money ...) that would lead to the realization that the experiment itself. Focusing on the subject who concerns us, the problem presented around of impacts on materials is their non-linearity, and the high cost necessary for his experimental study. Therefore, the question is resolve this problems with a tool to be able to have a good result through simulations, which dramatically reduces costs associated with the analysis of impacts For this project has been made a tool, hereinafter referred to as "Simulator", which automates the training and the testing of an ANN, as well as the generation of simulation reports and summaries that allow extrapolating the behavior of a network based on its training and input data. Using this information we have tried to achieve for the best ANN configuration to the problem that concerns us. One of the project main objectives is to help the reader to understand network's behavior by analyzing the results that it provides. That is, how influence in the results the quantity and the variety of the input data, as well as the configuration of the distinct training methods. Another goal of any simulation project is to determine the influence of each of the input variables on the results obtained by ANN. This project will study two methods for assessing the criticality of the input variables "Fortress before outliers" and "Sensitivity about the mean". It will analyze and compare the results of both methods to see if their results match, and if not, assess which method seems more reliable.
The primary objective of any project related to the simulation is to obtain reliable results from experiments presented. The best quality to be met by any simulation lies in achieving reliable results with a lower cost of resources (time, personnel, money ...) that would lead to the realization that the experiment itself. Focusing on the subject who concerns us, the problem presented around of impacts on materials is their non-linearity, and the high cost necessary for his experimental study. Therefore, the question is resolve this problems with a tool to be able to have a good result through simulations, which dramatically reduces costs associated with the analysis of impacts For this project has been made a tool, hereinafter referred to as "Simulator", which automates the training and the testing of an ANN, as well as the generation of simulation reports and summaries that allow extrapolating the behavior of a network based on its training and input data. Using this information we have tried to achieve for the best ANN configuration to the problem that concerns us. One of the project main objectives is to help the reader to understand network's behavior by analyzing the results that it provides. That is, how influence in the results the quantity and the variety of the input data, as well as the configuration of the distinct training methods. Another goal of any simulation project is to determine the influence of each of the input variables on the results obtained by ANN. This project will study two methods for assessing the criticality of the input variables "Fortress before outliers" and "Sensitivity about the mean". It will analyze and compare the results of both methods to see if their results match, and if not, assess which method seems more reliable.
Description
Keywords
Inteligencia artificial, Redes neuronales, Perceptrón multicapa, Simulación por ordenador