Contributions to projection pursuit learning / Contribuições para o aprendizado por busca de projeção
AUTOR(ES)
Leonardo de Moraes Holschuh
DATA DE PUBLICAÇÃO
2008
RESUMO
The production of parsimonious models is a common demand on a wide variety of engineering problems, as in the design of embedded systems. Constructive algorithms for supervised learning have shown to be effective methodologies for the synthesis of parsimonious artificial neural networks, with high levels of accuracy and generalization capability, though requiring more computational resources during the training phase. Even being one of the most frequently adopted constructive learning methods, the projection pursuit learning algorithm still presents some limitations, and three of them will be treated here: the initialization of the direction of projection, the use of a bias term at the input of the hidden-layer neurons, and the selection of input variables as a form of reducing the number of inputs, i.e. the dimension of the vector of projection. The variable selection technique adopted here is denoted wrapper, and a genetic algorithm was considered as the search engine. The performance analysis has been carried out by experiments involving time series prediction, indicating that the three propositions suggested to deal with limitation of projection pursuit learning contribute favorably to the process of constructive learning
ASSUNTO(S)
genetic algorithms projection modelos não-lineares (estatistica) inteligencia artificial redes neurais (computação) neural networks models non-linear (statistical) projeção artificial intelligence algoritimos geneticos
ACESSO AO ARTIGO
http://libdigi.unicamp.br/document/?code=000438705Documentos Relacionados
- Contribuições da teoria da atividade para m-learning
- Contributions to topic taxonomy construction in a specific domain using statistical learning
- Projeção multidimensional aplicada a visualização de resultados de busca textual
- Contributions to monitoring process for attributes.
- Relational reinforcement learning to control sociable robots