Contextual Bandit-Based Amplifier IBO Optimization in Massive MIMO Network

Abstract

Massive Multiple-Input Multiple-Output (MMIMO) is one of the 5G key enablers. Though, most of the works consider MMIMO under assumptions of ideal hardware. It has been shown that Power Amplifiers (PAs) introduce nonlinear distortion while operating close to their saturation power. Moreover, these distortions are in some cases beamformed toward the user, preventing antenna array gain from solving this problem. One of the possible solutions is an adaptive adjustment of the PA operating point, measured by Input Back off (IBO), to find a balance between wanted signal power and nonlinear distortion power. This work proposes a Contextual Bandit-Based IBO Optimization (COBBIO) algorithm to find rate-maximizing IBO for a given user’s radio conditions using learning through interaction. The proposed solution is tested in a realistic analog beamforming MMIMO cell simulator with multiple functional blocks, e.g., precoder, user scheduler, and utilizing an accurate 3D Ray-Tracing radio channel model. COBBIO provides throughput gains both over fixed-IBO schemes and state-of-the-art analytical IBO adjustment algorithms. The highest gains were observed for the so-called cell-edge users, where up to 83% improvement over the state-of-the-art algorithm was observed for the proposed COBBIO algorithm.

Keywords

• Massive MIMO
• 5G
• machine learning
• nonlinear distortion
• input back-off (IBO)