On estimation of nonlinear black-box models: how to obtain a good initialization
Jonas Sigfrid Sjoberg · 2002
An algorithm to define and initialize nonlinear recurrent neural net models using linear models is described. From a modeling point of view it is natural to try linear models first and then continue with nonlinear models. The suggested method gives such an algorithm and the nonlinear recurrent model is defined as an extension of the linear model. This gives less problems with local minima compared to a random initialization. Also, the stability of the model and its derivative with respect to the parameters can be guaranteed which is a requirement for the prediction-error estimation method (sometimes called back-propagation through time) to be applicable.