A Nonlinear Controller with High Interpretability using Kolmogorov-Arnold Network

Abstract

This paper proposes a data-driven nonlinear controller using Kolmogorov-Arnold Network (KAN). In contrast to a famous Multilayer Perceptron (MLP), KAN utilizes learnable b-spline curve as activation functions. Because of this, KAN can represent complex functions by small neurons, and has high interpretability. By employing KAN as a controller, the proposed controller has both high representational power and interpretability.