Publicação

Evolving time-lagged feedforward neural networks for time series forecasting

Ver documento

Detalhes bibliográficos
Resumo:Time Series Forecasting (TSF) is an important tool to sup- port both individual and organizational decisions. In this work, we propose a novel automatic Evolutionary Time- Lagged Feedforward Network (ETLFN) approach for TSF, based on an Estimation Distribution Algorithm (EDA) that evolves not only Artificial Neural Network (ANN) parame- ters but also which set of time lags are fed into the fore- casting model. Such approach is compared with similar strategy that only selects ANN parameter and the conven- tional TSF ARIMA methodology. Several experiments were held by considering six time series from distinct domains. The obtained multi-step ahead forecasts were evaluated us- ing SMAPE error criteria. Overall, the proposed ETLFN method obtained the best forecasting results. Moreover, it favors simpler neural network models, thus requiring less computational effort.
Autores principais:Peralta Donate, Juan
Outros Autores:Cortez, Paulo; Gutierrez Sanchez, German; Sanchis de Miguel, Araceli
Assunto:Connectionism and neural nets Hybrid systems artificial neural networks estimation distribution algorithm forecasting time series
Ano:2011
País:Portugal
Tipo de documento:comunicação em conferência
Tipo de acesso:acesso aberto
Instituição associada:Universidade do Minho
Idioma:inglês
Origem:RepositóriUM - Universidade do Minho
Descrição
Resumo:Time Series Forecasting (TSF) is an important tool to sup- port both individual and organizational decisions. In this work, we propose a novel automatic Evolutionary Time- Lagged Feedforward Network (ETLFN) approach for TSF, based on an Estimation Distribution Algorithm (EDA) that evolves not only Artificial Neural Network (ANN) parame- ters but also which set of time lags are fed into the fore- casting model. Such approach is compared with similar strategy that only selects ANN parameter and the conven- tional TSF ARIMA methodology. Several experiments were held by considering six time series from distinct domains. The obtained multi-step ahead forecasts were evaluated us- ing SMAPE error criteria. Overall, the proposed ETLFN method obtained the best forecasting results. Moreover, it favors simpler neural network models, thus requiring less computational effort.