Design Methodologies of Physical Layer Artificial Intelligence Technologies

Artificial intelligence (AI) has been introduced into the design of physical layer technologies and showed great potential in improving system performance. However, there are many open issues and challenges to be addressed. In this paper, the design methodologies of physical layer AI are investigated. The evaluation methodology and evaluation key performance indicators (KPIs) are first presented, which include the performance-related and cost-related KPIs. Furthermore, examples are given to illustrate how to integrate AI and expert knowledge, including dataset acquisition, AI model selection, data preprocessing based on expert knowledge, and auto-encoder design based on partial channel reciprocity and other available information. Some designs to enhance the generalization capability of AI models are also proposed, including low overhead auto-encoder design, transfer learning, and distributed machine learning. Finally, the standardization roadmap of physical layer AI is investigated.