Two-Timescale Beamforming for IRS-Assisted Millimeter Wave Systems: A Deep Unrolling-Based Stochastic Optimization Approach

We study the problem of joint active and passive beamforming for intelligent reflecting surface (IRS)-assisted millimeter Wave (mmWave) communication systems. To relieve from frequent instantaneous channel state information (I-CSI) acquisition, we consider a two-timescale beamforming protocol, in which the reflecting coefficients at the IRS are designed according to the long-term (i.e. statistical) CSI, and the transmit beamforming matrix is devised based on the instantaneous equivalent channel in a short-term scale. Such a problem is formulated as a stochastic optimization problem. To tackle this highly-coupled non-convex problem, we propose a combined effective channel gain maximization criterion for long-term passive beamforming. Also, a deep unrolling-based method is developed to provide a unified framework to address the stochastic optimization problem. Simulation results show that the proposed two-timescale beamforming methods can approach the performance of the beamforming approach that is based on the I-CSI.