A Novel Deep-Q-Network-Based Fine-Tuning Approach for Planar Bandpass Filter Design

Abstract

This letter proposes a novel fine-tuning approach of microstrip bandpass filters (BPFs) with a deep Q-network (DQN). In conventional works, reinforcement learning using DQN has been investigated for the automation of screw tuning in cavity BPFs. However, cross couplings appearing in planar BPF require a more complicated tuning process. To consider all the cross couplings, the proposed method introduces two neural-network-based surrogate models expressing the relationship between a coupling matrix and structural parameters. The two models also enable to drastically speed up reinforcement learning. As an example, a DQN is constructed for the design of the fifth-order microstrip BPF. The effectiveness of the DQN is numerically demonstrated through successful structural adjustments.