Application of Kolmogorov-Arnold Neural Network Architecture for Microwave Filter Tuning

Over the past few decades, the requirements for wireless communications equipment have increased dramatically. Each component of the microwave device must meet parameters provided in the terms of reference with minimum deviations. However, the variation of the component parameters (e.g., permeability of the dielectric ceramics) causes deviation of the microwave filter characteristics. To solve this problem, special constructive elements are added for tuning. These elements change the position of the microwave device components relative to each other, thereby affecting its characteristic. The process of tuning microwave devices is a rather time-consuming and consuming process, so in order to automate this process, one could use artificial neural networks. This article is dedicated to the application of a new type of neural network architecture - Kolmogorov-Arnold Networks (KAN) - for the microwave filter tuning task. Kolmogorov-Arnold neural networks showed that they have enormous potential in comparison to Multi-Layer Perceptron (MLP). Due to the new learning type, such neural networks can be trained faster and with fewer computing resources because they need far fewer parameters in contrast to MPL, and they can be even more accurate with a smaller number of parameters. The results of the KAN application are discussed.