NeuroFAST: high accuracy neuro-fuzzy modeling
Spyros G. Tzafestaş, Konstantinos C. Zikidis · 2003
Most fuzzy modeling algorithms rely either on simplistic (grid type) or off-line (trial-and-error type) structure identification methods. The proposed neurofuzzy modeling architecture, NeuroFAST, is an on-line, structure and parameter learning algorithm, featuring high function approximation accuracy. It is based on the first order Takagi-Sugeno-Kang (TSK) model (functional reasoning), where the consequence part of each fuzzy rule is a linear equation of the input variables. Fuzzy rules are allocated as learning evolves by a modified Fuzzy ART (Adaptive Resonance Theory) mechanism, assisted by fuzzy rule splitting and adding procedures (structure learning). The well known /spl delta/-rule continuously tunes learning weights on both premise and consequence parts (parameter identification). Tested on the Box-Jenkins gas furnace process modeling and the Mackey-Glass chaotic time series prediction, NeuroFAST yields very good results in terms of approximation accuracy, outperforming all known approaches.