Toward Intrinsically Interpretable AI in Optical Networks Using KAN-based Symbolic Regression

Artificial intelligence (AI) and machine learning (ML) have demonstrated remarkable performance in addressing multiple optical network tasks. However, the lack of interpretability and transparency in black-box models impedes understanding of model behaviors and thereby leads to subpar generalization ability and hinders operators from trusting model decisions. In this paper, we target building intrinsically interpretable AI models for optical networks using Kolmogorov–Arnold network (KAN)-based symbolic regression (SR). Specifically, KAN-based SR parameterizes intrinsically interpretable formulas by learnable activation functions (i.e., spline functions) and approximates the true mappings by formula fitting and pruning. We perform a case study on KAN-based SR for QoT estimation and demonstrate successful learning of interpretable formulas via a fully transparent training process. The results also show superior generalization ability of the proposed design compared with the baselines, achieving a mean squared error (MSE) of below 0.035 dB for optical signal-to-noise ratio (OSNR) prediction under different distributions of training and testing data.