Optimizing the Configuration of Intelligent Reflecting Surfaces using Deep Learning

Abstract

We consider a multi-user wireless network, where a single base station intends to communicate with multiple users by means of an intelligent reflecting surface (IRS), and we propose to optimize the IRS configuration using deep learning-based methodologies. In particular, we train a regression deep neural network to predict the communication channel parameters given the IRS configuration vectors. We further re-train this base model using the data of different users in order to maximize a weighted sum-rate objective function. Simulation results demonstrate that our proposed approach is able to optimize the IRS configuration for any unseen test users given their corresponding received signal patterns.