Deep frameworks based on residual dense network and reinforcement learning for mmWave massive MIMO systems

Abstract

In millimeter wave (mmWave) massive multiple-input-multiple-output (MIMO) systems, hybrid precoding plays a pivotal role in reducing complexity and cost while providing a good spectral efficiency. However, implementation of digital precoders with large number of antennas is difficult due to hardware constraints, while analog precoders offer confined performance. This leads to high computational complexity and cannot fully exploit the spatial information. Previous studies on hybrid precoding were based on exhaustive search solutions or greedy schemes, which result in higher complexity system performance. To face these challenges, this paper proposes deep hybrid precoding framework with phase quantization and residual dense network to design the matrix of analog and digital precoders. The proposed deep hybrid precoding technique consists of offline training stage and online deployment stage. In offline training stage, hybrid precoding is obtained assuming the approximate phase quantization. While in the online deployment stage, the matrix of analog precoding is calculated by exchanging approximate phase quantization with ideal phase and grouping the analog precoding vectors. In this paper, we also propose a deep reinforcement learning-based hybrid precoding. It consists of a deep reinforcement learning with employing convolutional neural network and long short-term memory (LSTM) methods. In our proposed frameworks, structures of proposed techniques are trained for maximum spectral efficiency.Our proposed techniques are compared with other precoding techniques. Results illustrate that the proposed techniques outperforms the other precoding techniques in terms of the spectral efficiency.