Learning to Optimize Resource in Dynamic Wireless Environment via Meta-Gating Graph Neural Network

Generally speaking, artificial intelligent (AI) models are trained under special learning hypotheses, especially the one that statistics of the training data are static during the training stage. However, the distribution of the channel state information (CSI) is constantly changing in real wireless communication environments. Therefore, it is essential to study the dynamic deep learning (DL) technology for wireless communications. In this paper, we investigate a beamforming design problem by maximizing the weighted sum rate in episodically dynamic wireless environment, where the CSI distribution changes over periods and maintains stationary within each period. In order to effectively solve this problem, a novel framework named meta-gating network is proposed, which can achieve three important goals, i.e., seamlessly, quickly and continuously. Specifically, the proposed framework consists of an inner network and an outer network, both of them are implemented by graph neural networks (GNNs). To achieve the former two goals, we propose a training method by combining the model-agnostic meta learning (MAML) algorithm with the unsupervised training. Following this training method, the outer network can help the inner network learn good initialization and then fast adapt to the different channels. As for the goal of ‘continuously’, we design an element-wise gating operation to multiply the outputs of the inner and outer networks, aiming at the selection activation of the inner network. Simulation results demonstrate that the proposed meta-gating GNN can well achieve the three important goals compared with the existing state-of-the-art algorithms.