A Distributed Learning Algorithm for Power Control in Energy Efficient IRS Assisted SISO NOMA Networks

This paper proposes a novel framework for energy efficiency maximization in an intelligent reflecting surface (IRS) aided single-input, single-output (SISO) non-orthogonal multiple access (NOMA) network through distributed learning based power control. A two-timescale based algorithm is presented to jointly optimize the transmit power of the user equipments (UEs) and reflection coefficients of the IRS elements, while ensuring a minimum rate of transmission for the users. The joint optimization problem is solved at two levels by employing two learning algorithms where the action choice updations in the learning algorithms are performed at two different timescales. The base station (BS) assists the IRS to learn its reflection coefficient matrix. The problem is formulated as an exact potential game with common payoffs and a stochastic learning algorithm (SLA) is proposed. During each iteration of SLA, corresponding to a particular reflection coefficient matrix of the IRS, the UEs learn the minimum transmit power required to satisfy their SINR requirements by employing a distributed learning for pareto optimality (DLPO) algorithm. The proposed learning algorithms are fully distributed since the UEs and the BS need to know only their own utilities and need not have the global channel state information (CSI).