Mytnyk O. Information technologies for robust neurofuzzy stochastic process model fusion

In this paper we investigate the approaches to forecasting of stochastic processes based on observed data as an ill-posed stochastic dependence recovery problem. The regularization methods are studied to pose the problem well. The generalized neurofuzzy network of C.Harris is considered as the state-of-the-art approach to construct the polynomial complexity transparent models. We present an inductive method to build balanced robust neurofuzzy model in Bernstein form based on Bayesian SVR in feature space spanned by Bezier-Bernstein polynomial functions (PRIAM – Polynomial Regression Inductive Algorithm Modeling) for stochastic dependence recovery problem. It combines the precedence of Bayesian inference, robustness of the support vector approach and transparency of the high end neurofuzzy modeling. Dual model conception allows PRIAM both be competitive with modern machine learning algorithms and be convenient for knowledge representation in expert systems. The complexity and convergence properties of PRIAM have been analysed. We conduct experiments on synthetic reference data sets as well as on real world economic, ecologic, meteorologic models and compare PRIAM forecasts with results of group method of data handling (GMDH), fuzzy GMDH, recurrent neural networks and ANFIS. Our experiments show that PRIAM outperforms the methods listed above having parsimony model construction logic.