Bandwidth Improvement for Patch Antenna via Knowledge-Based Deep Reinforcement Learning

In this letter, knowledge-based deep reinforcement learning (KBDRL) is used to improve patch antenna bandwidth. Conventional optimization algorithms usually require manual configuration of hyper-parameters of the algorithms, which cannot be learned automatically. In contrast, deep reinforcement learning (DRL) employs an end-to-end learning methodology that autonomously learns a policy to achieve optimal bandwidth. Additionally, it excels in efficiently handling nonlinear and nonconvex optimization problems. However, DRL consumes a lot of computing resources when there are many variables. To solve this problem, we propose a KBDRL approach that efficiently combines professional knowledge and machine learning methods. The proposed method is applied to a patch antenna with a size of 25 mm x 25 mm for which the obtained relative bandwidth is 39.18%. Our study shows that KBDRL has remarkable advantages compared to several commonly used local optimization algorithms, global optimization algorithms, and numerical optimization algorithms. Moreover, KBDRL can not only improve the performance of the antenna, but also explore the upper limitation of the performance with a finite-size basic structure.