Hybridization of the SGTM Neural-Like Structure Through Inputs Polynomial Extension

摘要

In this paper, a new approach for increasing the approximation accuracy with the use of computational intelligence tools is described. It is based on the compatible use of the neural-like structure of the Successive Geometric Transformations Model and the inputs polynomial extension. To implement such an extension, second degree Wiener polynomial is used. This combination improves the method accuracy for solving various tasks, such as classification and regression, including short-term and long-term prediction, dynamic pricing, as well as image recognition and image scaling, e-commerce. Due to the use of SGTM neural-like structure, the high speed of the system is maintained in both training and using modes. The simulation of the described approach is carried out on real data, the time results of the neural-like structure work and the accuracy results (MAPE, RMSE, R) are given. A comparison of the operation of the method with existing ones, such as Support vector regression, Classic linear SGTM neural-like structure, Linear regression (using Stochastic Gradient Descent), Random Forest, Multilayer Perceptron, AdaBoost are made. The advantages of the developed approach, in particular with regard to the highest accuracy among existing ones were experimentally established.