ADAPTIVE RETRAINING ALGORITHM WITH SHAKEN INITIALIZATION

The paper presents new specific aspects that could improve the adaptive retraining procedure of artificial neural networks (ANNs) for time series predictions. Usually, a retraining step starts from proportionally reduced values of the parameters (weights) used in the previous version of the ANN model. This time, the initial configuration of the weights is randomly “shaken” in order to further improve the model. The present results are promising and show a better adaptation of the forecasting system in a nonstationary environment.