ENCFIS: An exclusionary neural complex fuzzy inference system for robust regression learning

Robust learning, an emerging research topic in recent years, is a promising branch of advanced artificial intelligence. Robust learning models target mainly noisy and rough datasets, predominantly in situations where noises and outliers are hard to remove. In this paper, the concept of robust learning is combined with complex fuzzy theory for the first time, proposing a novel neuro-fuzzy system ENCFIS with extensive adaptability to numerical regression problems, with or without noise. Simulation results indicate that such architecture has excellent performance on a dataset with massive (45%) label noises and on a distorted time series dataset (25% corrupted). Additionally, experimental results on a metallurgy dataset also show that the approximation performance of ENCFIS is not compromised for the increase in robustness, making it an ideal candidate for general industrial scenarios with weak noise but difficult data characteristics.