Low-Complexity MIMO Detection Based on Reinforcement Learning With One-Bit ADCs

Abstract

This paper proposes a low-complexity reinforcement learning detection (RLD) algorithm for multi-input multi-output systems with one-bit analog-to-digital converters. The proposed algorithm exploits pairs of quantized received signals and detected symbols as training examples to train the likelihood function (LF) of the system. A major challenge in optimizing the RLD algorithm is to determine the optimal policy that decides whether to exploit the training examples based on their reliabilities. Determining the optimal policy inherently involves huge complexities in reflecting all possible transitions among candidate symbols. Thus, we simplify the optimal policy by considering only the most probable candidates among all possible decisions to reduce this complexity. Another major challenge in applying the RLD algorithm is that it requires high computational complexity to produce soft information for detection. Thus, we define new branch and path metrics derived from the LF and then remove the candidate symbols whose path metrics are smaller than a pre-defined value to alleviate the complexity. Moreover, we analyze the complexity of the proposed algorithm by deriving the expected number of surviving candidates. Simulation results show that the proposed algorithm provides a better performance-complexity tradeoff than the conventional RLD algorithm.