Feature Selection in Multi-Layer Perceptrons (MLP) for Forecasting Using Self Organization Maps (SOM)

Nowadays multilayer perceptrons (MLPs) are one of the most important and widely-used neural networks used as continuous measurable function with a desired accuracy. The second benefit is nonparametric data-driven nature of them which meansmultilayer perceptrons impose few prior assumptions on the underlying process. Being adaptive is the third advantage of MLPs. The adaptation of MLPs implies that in a nonstationary environment the accuracy and robustness of results are still countable. Utilizing fewer parameters is the fourth benefit of MLPs. Despite all these unique advantages of multilayer perceptrons, they suffer from some limitations such as negative relationship between number of inputs and the achieved performance, though using hybrid methods to overcome the limitations by means of a method alone and improving forecasting performance is achievable. Literature review suggests that by utilizing disparate and unrelated methods, we can obtain a new hybrid scheme capable of less variance or error. Hybridization of dissimilar methods can reduce the risk of using an inappropriate method. Usually, this is done based on this fact that the underlying process cannot easily be determined. The motivation behind using hybrid method is two folds: either single method cannot identify the true data generating process or cannot identify all the characteristics of the time series. In this paper, a new hybrid method of multilayer perceptrons is proposed which uses the self-organizational maps. The self-organizational maps are one of the most accurate tools in recognizing and analyzing the nonlinear multidimensional spaces. In the proposed method, inputs of the multilayer perceptron are firstly clustered by using a self-organizational map, and then variables in each cluster are combined together according to their effectiveness values. Empirical results of steel price forecasting in TehranMetal Exchange indicate that the efficiency of the proposed method is comparable to other methods.