The hybrid deep learning GMDH- neo-fuzzy network was suggested and investigated. The application of GMDH based on self-organization enables to build optimal structure of neo-fuzzy system and train weights of neural network in one procedure. As a node of hybrid neo-fuzzy system neo-fuzzy network with small number of adjustable parameters is suggested. This enables to cut training time and accelerate convergence of training. The experimental studies of GMDH-neo-fuzzy network were carried out in the task of forecasting in macro-economy and financial sphere. The forecasting efficiency of the suggested hybrid neo-fuzzy network was estimated and its sensitivity to variation of tuning parameters was investigated. The suggested approach allows to prevent the drawbacks of deep learning such as vanishing or exploding gradient.
[1]
M. MendelJ.,et al.
Fuzzy basis functions
,
1995
.
[2]
Jyh-Shing Roger Jang,et al.
ANFIS: adaptive-network-based fuzzy inference system
,
1993,
IEEE Trans. Syst. Man Cybern..
[3]
A. G. Ivakhnenko,et al.
Inductive sorting-out GMDH algorithms with polynomial complexity for active neurons of neural network
,
1999,
IJCNN'99. International Joint Conference on Neural Networks. Proceedings (Cat. No.99CH36339).
[4]
L X Wang,et al.
Fuzzy basis functions, universal approximation, and orthogonal least-squares learning
,
1992,
IEEE Trans. Neural Networks.