Abstract
Near-infrared (NIR) sorting systems in waste sorting plants operate under multiple settings, creating distinct input-output relationships that challenge predictive modeling. Conventional neural networks, such as multilayer perceptron (MLP), often suffer from catastrophic forgetting under continual training, limiting reliability across settings. This study evaluates Kolmogorov-Arnold Networks (KAN) for continual regression modeling of multi-setting NIR systems. KAN assign nonlinear transformations to network edges using localized spline grids, enabling structural isolation between input regions. We introduce controlled input-space manipulations (shifting successive settings to adjacent or non-overlapping grid regions) and compare KAN performance with MLPs of comparable parameter count. We also examine single-input versus multi-input configurations to assess dimensionality effects. Results show that KANs with sufficient input-space separation maintain previously learned knowledge with perfect resistance to forgetting, whereas overlapping inputs induce interference. MLP forgetting depends on learning rate and training duration and cannot be fully avoided without additional methods. Single-input KANs achieve comparable accuracy to multi-input models in this system, suggesting limited benefit from additional inputs. These findings demonstrate that KAN's structural locality, combined with controlled input-space alignment, provides a practical and robust approach for continual learning in industrial NIR systems.