Self rising tri layers MLP for time series forecasting

Abstract

Time series forecasting is an attractive and heavily researched area. A very popular approach in this field is the usage of artificial neural networks. Some artificial neural network are oriented to deep learning as training algorithm. In this study instead of hidden layers number extension the size of input layer of tri layers multilayer perceptron is extended. The network starts with 1-1-1 topology. The input layer rise to n, according the size of input time series. In parallel hidden layer goes to m by application of pruning algorithm. Achieved topology n-m-1 is trained with classical backpropagation of the error.