Abordagem hÃbrida para otimizaÃÃo de redes neurais artificiais para previsÃo de sÃries temporais

AUTOR(ES)

Adrian Lucena Arnaud

DATA DE PUBLICAÇÃO

2007

RESUMO

This thesis proposes a new hybrid approach which combines simulated annealing and standard error backpropagation for optimizing Multi Layer Perceptron Neural Networks (MLP) for time series prediction. This approach named ANNSATS (Artificial Neural Networks and Simulated Annealing for Time Series Forecasting) starts from an initial topology fully connected network with a pre-specified number of neurons and weights. The proposed hybrid system passes through cycles composed of network topology optimization followed by weight optimization. These cycles continue until the optimal topology for the architecture and the optimal weight adjustment for the connections are found, through cross-validation or by reaching a pre-specified maximum number of cycles. At each cycle, firstly, a new candidate network is created, with potentially more adequate topology for solving the specific prediction problem. The new topology is created by enabling or disabling input and hidden neurons. Secondly, this new topology has its weights adjusted by standard error backpropagation for a short and pre-specified number of epochs. After these training epochs, the candidate network is evaluated according to a cost function. This cost function is calculated on the training set and is proportional to the number of neurons currently enabled in the network topology and proportional to the modeling error for the time series. Depending on the cost function and on the simulated annealing current parameters, the candidate topology can be either be accepted or rejected. If it is accepted, then it is used for generating a new candidate network with different topology with the application of simulated annealing, starting a new cycle. If the candidate network is not accepted, then the previous valid topology is used instead. The results produced by experimental tests carried out on one simulated series (HÃnon series) and on eight actual time series have shown that this approach selects the appropriate time series lags and builds an MLP with a number of hidden neurons adequate for achieving, in average, better performance than the MLP trained by the standard error backpropagation algorithm and better results than other available results in related literature

ASSUNTO(S)

time series forecasting decision support systems sistemas de apoio à decisÃo ciencia da computacao previsÃo de sÃries temporais intelligent hybrid systems otimizaÃÃo de redes neurais artificiais simulated annealing sistema hÃbridos inteligentes optimization of artificial neural networks simulated annealing

Documentos Relacionados

• Uma nova metodologia hÃbrida inteligente para a previsÃo de sÃries temporais
• Teoria da perturbaÃÃo em sistemas hÃbridos inteligentes para a previsÃo de sÃries temporais
• Modelo de sÃries temporais para a previsÃo da arrecadaÃÃo tributÃria federal
• PrevisÃo de sÃries temporais usando modelos de composiÃÃo de especialistas locais.
• ModeloshÃbridos baseados em enxame de partÃculaspara previsÃo de sÃries temporais