Application and Challenges of Machine Learning-Assisted Antenna Design

Publikation: Bidrag til bog/antologi/rapport/konference proceedingKonferenceartikel i proceedingForskningpeer review

Abstract

Machine learning-assisted antenna design has attracted a lot of attention due to the advancement of machine learning techniques. Despite its success, there are many challenges that need to be solved. Firstly, the quality of training data has an important effect on the accuracy and efficiency of the machine learning model. Research on high-quality training data generation is very necessary. Secondly, most existing machine learning methods rely on human experts to collect data, which is a time-consuming and tedious process. Studying how to automate antenna design is another challenge. Finally, while reinforcement learning (RL)-based antenna design can achieve fully automated design, their learning performance can be poor if they do not have sufficient prior knowledge, leading to slow convergence or even divergence. Determining how to improve learning efficiency is meaningful. This paper investigates the above three problems, illustrated by using three examples. For the first problem, we use domain knowledge stemming from the understanding of EM problems to guide the generation of training dataset so that the generated dataset are more relevant to the design goals, which reduces the solution space and improves model training efficiency. For the second problem, we use an RL algorithm to fully automate the antenna design. Unlike supervised learning training, which requires the use of human-supplied data, RL allows agents to learn from their own experience gained from interaction with the environment. The entire process does not require human involvement, relieving designers' burden. For the third problem, we take advantage of both imitation learning (IL) and RL. IL allows the agent to imitate the behavior of a human expert, and then RL autonomously explores the environment to boost generalization performance. This combination of IL and RL exhibits better performance compared to pure RL or IL. In addition, we also propose some future directions for machine learning-based antenna design. The machine learning-aided design approach is showing great potential in solving complex antenna design problems, which may become an indispensable tool for humans to design antennas in the future.
OriginalsprogEngelsk
TitelPhotonIcs and Electromagnetics Research Symposium
StatusAccepteret/In press - 2024

Fingeraftryk

Dyk ned i forskningsemnerne om 'Application and Challenges of Machine Learning-Assisted Antenna Design'. Sammen danner de et unikt fingeraftryk.

Citationsformater