Comparative Analysis of Different Approaches towards Multilayer Perceptron Training.pdfVIP

COMPARATIVE ANALYSIS OF DIFFERENT APPROACHES TOWARDS * MULTILAYER PERCEPTRON TRAINING A. Valishevsky Keywords: neural net; time-series prediction; adaptive learning algorithms; backpropagation; QuickProp; Rprop 1. Introduction Error Backpropagation algorithm appeared in mid-80’s and caused a true revolution in the field of neural nets, as it allowed to train multilayer neural networks, which are capable of solving non-linear classification tasks. Just after the introduction of the algorithm, articles showing its weaknesses started to appear. The principal weakness of backpropagation is that a comparatively big amount of iterations has to be made during the learning process. Three fundamentally different approaches towards decrease of the number of iterations are considered in this paper: arbitrary error steepness, employment of the second derivative and an adaptive learning rule. Share price forecasting task is being considered in order to compare the above-mentioned algorithms. This task can be considered as a non-linear classification task. There are several approaches towards forecasting of the behaviour of a given system. The first one is based on exact knowledge of regularities that underlie a system. If the regularities can be expressed in terms of precise equations, which can in principle be solved, it’s possible to predict the future behaviour of the system. However, the knowledge of the rules governing the behaviour of the system usually is not available. This is particularly true for most macroeconomic problems. The main cause is the non-stationary character of the system parameters. The second approach is based on the creation of statistical models, which approximate t