Uma anÃlise de otimizaÃÃo de redes neurais MLP por exames de partÃculas

AUTOR(ES)

Marcio Ribeiro de Carvalho

DATA DE PUBLICAÇÃO

2007

RESUMO

This work presents a methodology for global optimization of MLP artificial neural networks. The main objective here is the simultaneous optimization of architectures and connection weights of MLP networks, aiming good classification performance for most data sets. The simultaneous adjustment of architectures and connection weights of MLP networks represents an interesting approach to obtain efficient networks with good generalization performance, as it creates a compromise between low architectural complexity and low training errors. This application was investigated in earlier works using genetic algorithms, simulated annealing, tabu search and combinations of them. Other global search technique less applied for the optimization of architectures and connection weights of neural networks is the Particle Swarm Optimization (PSO) which has received more attention from the community due to good results obtained for numerical optimization problems. The methodology developed in this work consists in the application of two PSO algorithms for the optimization of MLP neural networks, the first one is applied in the search for good multilayer architectures, while the latter is specialized in the adjustment of connection weights for each one of the architectures generated in the first PSO. These two sub-processes are interleaved in a main process for a determined number of iterations. This work also presents results from the application of the proposed methodology in three well known pattern classification benchmark problems of the medical field. In the hardest problems, the presented methodology obtained satisfactory results and provided networks with low generalization error and complexity. Those results are important to show that the particle swarm optimization global search technique is an effective alternative to handle the optimization of architectures and connection weights of MLP artificial neural networks for classification problems

ASSUNTO(S)

tÃcnicas de busca global pso global search techniques mlp generalization generalizaÃÃo neural networks optimization pso otimizaÃÃo de redes neurais mlp ciencia da computacao

Documentos Relacionados

• OtimizaÃÃo estrutural utilizando o algoritmo evolucionÃrio do enxame de partÃculas
• Reconhecimento de objetos em imagens digitais utilizando otimizaÃÃo por enxame de partÃculas
• Optimization methods for the definition of MLP neural network architectures and weights
• Uma metodologia para otimizaÃÃo de arquitetura e pesos de redes neurais
• OtimizaÃÃo da geometria de aglomerados de silÃcio via redes neurais.