O uso de redes neurais e regressão linear múltipla na engenharia de avaliações : determinação dos valores venais de imóveis urbanos
Resumo
Resumo: O presente trabalho tem por objetivo comparar as técnicas de Redes Neurais Artificiais e Regressão Linear Múltipla, na estimação dos valores venais de imóveis urbanos da cidade de Guarapuava/PR. Para tanto, utilizou-se os dados do Cadastro Imobiliário, fornecido pelo setor de Planejamento da Prefeitura Municipal de Guarapuava. O modelo inicial foi composto por treze variáveis do referido cadastro, sendo elas: bairro, setor, pavimentação, esgoto, iluminação pública, área do terreno, pedologia, topografia, situação, área edificada, tipo, estrutura e conservação. Utilizou-se a técnica de Análise de Componentes Principais para reduzir e transformar as variáveis originais para nove fatores. As Redes Neurais Artificiais desenvolvidas foram do tipo feed-forward, utilizando o algoritmo de treinamento "Levenberg Marquardt" com uma camada oculta, cujos resultados foram comparados com os resultados obtidos através da Regressão Linear Múltipla, para a predição dos valores venais. As duas técnicas se mostraram eficientes, sendo que, para o problema estudado, o desempenho das Redes Neurais Artificiais foi superior ao resultado apresentado pela outra técnica. Abstract: The present work has the objective of comparing the techniques of Artificial Neural Networksand Multiple Linear Regression, in the estimation of venal values of urban real properties ofthe city of Guarapuava. For this purpose, data from the real estate dossier were utilized,supplied by the sector of Planning of the City Hall of Guarapuava. The initial model wascompounded of thirteen variables from the above mentioned dossier: quarter, sector, paving,drain, public illumination, land area, pedology, topography, situation, built area, type,structure and conservation. The technique of Principal Component Analysis was used toreduce and transform the original variables into nine factors. The Artificial Neural Networksdeveloped were of the type feed-forward, using the training algorithm Levenberg Marquardtwith one hidden layer, whose results were compared with the results obtained through theMultiple Linear Regression, for the prediction of venal values. Both techniques resultedefficient, and for this problem, the performance of Artificial Neural Networks was superior tothe result showed by the other technique.
Collections
- Teses & Dissertações [10564]