Avaliação do desempenho de algoritmos de retropropagação com redes neurais artificiais para a resolução de problemas não-lineares

• Ano de defesa: 2018
• Autor(a) principal: Barros, Victor Pedroso Ambiel
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Tecnológica Federal do Paraná; Ponta Grossa; Brasil; Programa de Pós-Graduação em Ciência da Computação
• Palavras-chave em Português: Redes neurais (Computação); Algorítmos; Desempenho; Neural networks (Computer science); Algorithms; Performance; CNPQ::CIENCIAS EXATAS E DA TERRA::CIENCIA DA COMPUTACAO; Ciência da Computação
• Link de acesso: http://repositorio.utfpr.edu.br/jspui/handle/1/3340

Artificial neural networks make possible to work with modeling and resolution of nonlinear problems by training, testing and validating the neural network with a set of input data and an output goal. However, the construction of an artificial neural network is complex and hard-working because there is no neural network model ready to solve any problem, each neural network must be built based on the problem that needs to be solved. One of the main points in the construction of a neural network is the correct choice of the training algorithm for the network to converge correctly, produce good results and correctly solve the problem addressed. Each training algorithm contains its pros and cons that should be taken into consideration. The present work presents the performance comparison between the Levenberg-Marquardt, Bayesian Regularization, Scaled Conjugate Gradient and Resilient Backpropagation algorithms applied to a non-linear case study. The application of the algorithms in the case study was carried out in three different neural network architectures, allowing to evaluate the performance of the algorithms in different architectures. The case study is the learning of the value of the stoichiometric richness in spark ignition engine, which is characterized as a non-linear and high complexity problem. The results obtained with the training of the algorithms in the different architectures showed the importance of the architecture of the neural network used, being that one of the three architectures developed the best result and two algorithms were able to achieve excellent performance rates, whereas other two algorithms did not obtain satisfactory results.