Adaptive Incremental Learning Stochastic Configuration Network and Its NIR Modeling Application

Near-infrared (NIR) technologies have shown promising advantages in industrial processes. With the continuous increase of spectral dimension, establishing a prediction model with stability and generalization has been a significant research problem in NIR modeling. Compared with traditional methods, stochastic configuration network (SCN) based on randomized approach has demonstrated great advantages in developing nonlinear models due to its effectiveness and efficiency in model structure and parameters learning. However, none of the existing point or block increment learning strategies can obtain a compact network structure while converging fast. To further improve modeling efficiency for high-dimensional NIR analysis, the adaptive incremental learning SCN (SCN-A) algorithm is proposed in this work, in which the size of the increment learning block adaptively adjusts during the iteration process. Comparative results on two measured NIR datasets verify the effectiveness of the proposed method. Compared with the state-of-the-art SCNs, the SCN-A method can obtain similar prediction performance using fewer iteration times.