Evolutionary Neuro-Fuzzy Random Vector Functional Link with Genetic Optimization

Abstract

Random Vector Functional Link (RVFL) is integrated with a fuzzy inference system to enhance adaptability and interpretability, forming the Neuro-Fuzzy RVFL (NF-RVFL) model. This model is pre-accompanied by randomly initialized parameters, which play a vital role in shaping the error rates, decision boundary of fuzzy rules, and computability of the models. Therefore, we propose to utilize genetic algorithm-based optimization techniques for randomly initialized parameters in the NF-RVFL and call it Evolutionary Neuro-Fuzzy Random Vector Functional Link (ENF-RVFL). The proposed ENF-RVFL processes fuzzified input and optimizes fuzzy rule weight, hidden weight, and bias from the standalone genetic algorithm to process the hidden layer output. The fuzzy inference system utilizes optimized IF-THEN rule weight coefficients to compute defuzzied values. Lastly, the defuzzied values, hidden layer output, and input-to-output link are used to analytically compute the output weight using the Moore–Penrose generalized inverse. The experimentation results obtained on UCI benchmark datasets for binary and multiclass classification tasks show that the proposed model, ENF-RVFL, outperforms NF-RVFL variants regarding accuracy, generalization, and robustness while effectively managing uncertainty.