# A network of autoregressive processing units for time series modeling

Research output: Contribution to journal › Article

### Standard

**A network of autoregressive processing units for time series modeling.** / Lehtokangas, M.; Saarinen, J.; Kaski, Kimmo; Huuhtanen, P.

Research output: Contribution to journal › Article

### Harvard

*Applied Mathematics and Computation*, vol. 75, no. 2-3, pp. 151-165. https://doi.org/10.1016/0096-3003(96)90057-0

### APA

*Applied Mathematics and Computation*,

*75*(2-3), 151-165. https://doi.org/10.1016/0096-3003(96)90057-0

### Vancouver

### Author

### Bibtex - Download

}

### RIS - Download

TY - JOUR

T1 - A network of autoregressive processing units for time series modeling

AU - Lehtokangas, M.

AU - Saarinen, J.

AU - Kaski, Kimmo

AU - Huuhtanen, P.

PY - 1996

Y1 - 1996

N2 - A new approach for time series modeling has been developed. This has a neural network structure with three layers: the input, switch, and processing layers. The processing layer contains processing units, whose structure is defined by the user. These units follow the structure of a general autoregressive model. A hybrid training method that combines self-organized and supervised training is used for parameter estimation. The model selection of this network structure is studied by an information theory criterion that is based on the predictive minimum description length principle. The aim is to determine the size and complexity of the model. This network is found to perform as well as or even better than the multilayer perceptron network, the radial basis function network, the threshold autoregressive model, and the autoregressive model in nonlinear time series modeling experiments. Other benefits of the network are its simple architecture and fast processing ability in the training and recalling phases.

AB - A new approach for time series modeling has been developed. This has a neural network structure with three layers: the input, switch, and processing layers. The processing layer contains processing units, whose structure is defined by the user. These units follow the structure of a general autoregressive model. A hybrid training method that combines self-organized and supervised training is used for parameter estimation. The model selection of this network structure is studied by an information theory criterion that is based on the predictive minimum description length principle. The aim is to determine the size and complexity of the model. This network is found to perform as well as or even better than the multilayer perceptron network, the radial basis function network, the threshold autoregressive model, and the autoregressive model in nonlinear time series modeling experiments. Other benefits of the network are its simple architecture and fast processing ability in the training and recalling phases.

KW - stochastic complexity

U2 - 10.1016/0096-3003(96)90057-0

DO - 10.1016/0096-3003(96)90057-0

M3 - Article

VL - 75

SP - 151

EP - 165

JO - Applied Mathematics and Computation

JF - Applied Mathematics and Computation

SN - 0096-3003

IS - 2-3

ER -

ID: 10192081