An Enhanced Neuro-Fuzzy Random Vector Functional Link with Affine Transformations

Abstract

The Random Vector Functional Link (RVFL) architecture has proven its efficiency compared to traditional neural networks, which suffer long training times and have suboptimal solutions. Neuro-Fuzzy based RVFL (NF-RVFL) models have further enhanced the interpretability while maintaining the efficiency and generalization capabilities of RVFL. However, the practical challenges of input scaling, poor data distribution, and initialization issues persist. To address these issues and enhance information retention and generalization, we propose introducing an affine transformation (AT) to the input of NF-RVFL. We collectively term this novel model as ATNF-RVFL. This AT would enhance efficiency, making the model more transparent, and strengthen its ability to handle linear transformations. Input data is initially fuzzified and later enhanced by AT before passing through a fuzzy layer in the ATNF-RVFL model. These components are further processed by a hidden layer through random projections and are defuzzified to make them more interpretable. The defuzzified values, hidden layer outputs, and the original input data are combined to determine the final prediction. We tested the ATNF-RVFL model with the benchmark datasets, covering binary and multiclass classification tasks. The outcomes highlight the efficiency of the proposed model, outperforming other approaches.