Design of an Iterative Refined Radial Basis Function Neural Network (IRRBFNN) Using Error-based weights

In this study, an Iterative Refined Radial Basis Function Neural Network (IRRBFNN) is proposed for regression tasks. The model leverages error information generated at each inference step to sequentially ⅰ) localize the input space via Weighted Fuzzy C-means(WFCM) and ⅱ) optimize connection weights through Weighted Least Square Estimation (WLSE). At this stage, to optimize the number of local regions, a high-error region search and splitting scheme was applied, and acceptance criteria were used to determine whether a split should be executed. When evaluated on 16 benchmark datasets against state-of-the-art (SOTA) baselines, the proposed method achieves superior performance on 10 datasets. A case study on cement strength prediction further demonstrates the advantages of the IRRBFNN over existing application-specific models, highlighting its practical applicability to forecasting the properties of materials and related regression problems.