A New Deep Complex-Valued Single-Iteration Fuzzy System for Predictive Modelling
Chuan Xue, Mahdi Mahfouf · 2022 IEEE International Conference on Fuzzy Systems (FUZZ-IEEE) · 2022
Numerical prediction is an important application field of machine learning. However, the current mainstream relating to deep network solutions can be computationally taxing, whereas fuzzy systems may also prove to be inefficient for high dimensional systems. Combining the notion of complex-valued neural network and the Wang-Mendel (WM) fuzzy algorithm, we propose a new complex-valued Wang-Mendel (CVWM) method, which reduces the rule-base of fuzzy systems to the scale of its square root. Further, by introducing the concept of a hierarchical fuzzy system, a deep complex-valued single-iteration fuzzy system (DCVSF) that can be trained with only one iteration and can effectively process high-dimensional data is also elicited. In addition, for sparse data, t-distributed stochastic neighbor embedding (t-SNE) dimensionality reduction is introduced to increase data density. The experimental results show that both CVWM and DCVSF exhibit competitive nonlinear modeling capabilities.