Deep Reinforcement Learning-based Auto-tuning Algorithm for Cavity Filters

Abstract

Over the past few decades, the tuning of cavity filters has often been done by trial and error, using human experience and intuition, due to the imprecision of the design and manufacturing tolerances, which often results in detuning the filters and requiring costly post-production fine-tuning. Various techniques using optimization and machine learning have been investigated to automate the process. The superiority of a deep reinforcement learning approach, which can properly explore various possibilities and operate them in the desired way according to the well-defined reward, has motivated us to apply it to our problem. To meet the demand for an automatic tuning algorithm for cavity filters with high accuracy and efficiency, this study proposes an automatic tuning algorithm for cavity filters based on the deep reinforcement learning. For the efficiency of the tuning process, we limit the order of the elements to be tuned, inspired by the experience of experts based on domain knowledge. In addition, the coarse tuning process is performed first, followed by the fine tuning process to improve the tuning accuracy. The proposed method has demonstrated the ability of the deep reinforcement learning to learn the complex relationship between impedance values of equivalent circuit elements and S-parameters to effectively satisfy filter design requirements within an acceptable time range. The performance of the proposed automatic tuning algorithm has been evaluated through simulation experiments. The effectiveness of the proposed algorithm is demonstrated by the fact that it is able to tune a detuned filter from random starting point to meet its design requirements.