RACFIS: A new rapid adaptive complex fuzzy inference system for regression modelling

The theory of complex fuzzy sets has made great breakthroughs in recent times. Complex fuzzy theory (CFT) allows a fuzzy set to include more information with the help of its two-dimensional rule-base, which is of great potential for improving the related fuzzy system performance while managing the size of the associated rule-bases. In this paper, a new rapid adaptive complex fuzzy inference system (RACFIS) is proposed by redesigning the optimization policy of the earlier complex neuro-fuzzy system (CNFS) algorithm. Improvements include a new three-parameter Quasi-hyperbolic momentum (QHM) optimization method to replace the original particle swarm optimization (PSO), and unsupervised learning is introduced, for the first time, into the complex neuro-fuzzy model to pre-train the antecedent parameters for a better global optimum as well as a faster convergence. Experimental results show that RACFIS performs very well on all data sets, obtaining excellent prediction accuracies with on average 10 times lower epoch numbers (as compared with all benchmark models) and a reduction in the size of the rule-bases by nearly 20 % ∼30% (as compared with non-complex fuzzy models). RACFIS also possesses a strong generalization capability that outperforms all the benchmark algorithms employed in the simulation experiments. In addition, a mean impact value (MIV) algorithm based on radial basis function (RBF) network is also employed to select variables with higher relevance in order to mitigate the drawbacks caused by the higher dimensionality of the data.