Online time series prediction with meta-cognition
Koshy George, Prabhanjan Mutalik · 2016
Predicting the future course of a sequential collection of an observable has several applications in diverse fields. Traditional techniques assume fixed linear models. In contrast, models based on artificial neural networks are adaptive and nonlinear; however these are typically trained offline. This paper focuses on time-series prediction using a neural network with a single hidden layer that is trained using a sequential variant of the extreme learning machine. The learning process incorporates feedback as the previous predicted values are also used as inputs to the network. The different initialisation of the learning algorithm is shown to improve the prediction performance. This is further improved by including a meta-cognitive component which decides what the predictor should learn and when it should learn.