Real-time adaptive wireless communication system design via meta-learning

Abstract

This paper proposes AI-based wireless communication system that can adapt to new environment in real-time fashion. It is well known that when there is model deficiency, data-driven approach can result in better performance compared to model-based approach. However, this data-driven approach requires sufficient amount of training data to achieve (near-)optimal solution. This restriction is crucial especially when we deal with wireless communication system that rapidly varies in real-time. In this paper, we propose to solve this problem via meta-learning, or sometimes called, learning to learn. Given a particular wireless communication environment, e.g., channel realization, meta-learning first experiences about how 'to learn' in order to achieve (near-)optimal solution

repeat the same procedure for multiple environments to eventually 'learn' some good way of 'learning' to any new, unseen, but related environments. In this work, we have considered two aspects in wireless communication system: 1) neural network design of receiver and 2) neural network design of both transmitter and the receiver. In both cases, sufficient training data, or pilot data, is needed with enough training time to ensure a near-optimal neural network design. In this work, with the aid of meta-learning, we propose a scheme to leverage these constraints: overhead in pilot data and training time. Extensive numerical results validate the advantage of meta-learning in wireless communication systems both in performance and overhead perspective compared to conventional communication schemes and conventional machine learning schemes.