KAT-RSCSA: Routing, Spatial Channel, and Spectrum Assignment in Spatial Channel Networks Using Kolmogorov–Arnold Transformer

Spatial Channel Networks (SCNs) extend the optical bypass concept from wavelength division multiplexing (WDM) to space division multiplexing (SDM), enabling more efficient resource allocation under rising bandwidth demand. However, SCN deployment remains challenging because spatial channel (SCh) types exhibit heterogeneous transmission constraints, and variations in topology and in the size and distribution of service demands further limit utilization and throughput. A central lever is the switching granularity–how SChs are partitioned between bypass and wavelength switching SChs–which trades routing flexibility against guardband overhead and directly affects blocking and fragmentation. To address it, we propose a predict-then-assign framework: first predict the optimal switching granularity for the current network state, then execute routing, SCh selection, and spectrum assignment accordingly. To make reliable predictions, we develop an enhanced Kolmogorov–Arnold Transformer (KAT) that uses convolutional neural networks (CNNs) for feature extraction and replaces the Transformer multilayer perceptron (MLP) with a Kolmogorov–Arnold Network (KAN). Guided by the predictions, the adaptive switching granularity-aware routing, spatial channel, and spectrum assignment (ASGA-RSCSA) algorithm performs joint routing and resource allocation. Experiments show that our proposed KAT streamlines granularity selection, substantially reduces computation time, and achieves nearly 90% top-1 prediction accuracy and over 95% top-2 prediction accuracy. Guided by these predictions, ASGA-RSCSA reduces blocking probability, increases throughput, and lowers spectrum fragmentation compared with existing methods, demonstrating the effectiveness of the integrated approach.