The Development of Improved Incremental Models Using Local Granular Networks with Error Compensation
Chan-Uk Yeom, Keun-Chang Kwak · Symmetry · 2017
In this paper, we use the fundamental idea of the incremental model (IM) and develop the design framework. The design method of IM is composed of two steps. In the first step, we perform a linear regression (LR) as the global model. In the second step, the errors obtained by the global model are predicted by fuzzy if-then rules generated through a local linguistic model. Although the effectiveness of IM has been demonstrated in various prediction examples, we propose an improved incremental model (IIM) to deal with complex nonlinear characteristics. For this purpose, we employ adaptive neuro-fuzzy networks (ANFN) or radial basis function networks (RBFN) to create local granular networks in the design of IIM. Furthermore, we use quadratic regression (QR) as a global model, because linear relationship of LR may not hold in many settings. Numerical studies concern four datasets (automobile data, energy efficiency data, Boston housing data and computer hardware data). The experimental results demonstrate that IIM outperformed the previous models.