Kolmogorov-Arnold Reservoir Computing

Overview

Kolmogorov-Arnold Reservoir Computing (KARC) is a novel framework for forecasting complex dynamical systems. It addresses the limitations of traditional reservoir computing (RC), which often struggles with hyperparameter sensitivity and sequential dependencies, and next-generation reservoir computing (NG-RC), which suffers from a rapid, combinatorial explosion of feature dimensions. By leveraging the Kolmogorov-Arnold representation theorem, KARC provides a mathematically principled way to project input data into a high-dimensional space using fixed basis functions, allowing for efficient, closed-form training via ridge regression.

The core innovation of KARC is its bridge between reservoir computing and Kolmogorov-Arnold Networks (KANs). While standard KANs use backpropagation to learn univariate functions on network edges, KARC simplifies this by using a fixed dictionary of basis functions (such as Fourier, B-spline, or Chebyshev functions). This transformation allows the model to represent complex multivariate mappings as compositions of univariate functions, which are then combined into a linear readout. Because the basis functions are fixed, the model avoids the iterative, gradient-based optimization typically required by deep learning architectures, retaining the speed and simplicity of traditional reservoir computing.

Performance and Applications

The authors validated KARC across several challenging benchmarks, including the chaotic double-scroll system, the Kuramoto-Sivashinsky equation, and two-dimensional shallow water equations. In these experiments, KARC consistently outperformed traditional RC and NG-RC, demonstrating longer valid prediction times and lower normalized root mean squared errors (NRMSE). Notably, in high-dimensional PDE-governed systems, KARC maintained high accuracy without the need for the complex, localized reservoir partitioning often required by other methods. Beyond dynamical systems, the framework shows promise as a feature-forecasting module for generative diffusion models.