Deep Learning-Based Low Complexity MIMO Detection via Partial MAP

Abstract

In multiple-input multiple-output (MIMO) communication systems, signal detection plays a crucial role in achieving reliable and high-performance wireless communication. However, the complexity of optimal detection methods, such as maximum likelihood (ML) detection, grows exponentially with the number of transmit antennas when exhaustive search is used, hindering practical implementation. To address this challenge, suboptimal algorithms such as successive interference cancellation (SIC)-based detection have been developed, but they suffer from error propagation. To mitigate error propagation in SIC detectors, a soft-decision based partial maximum a posteriori (MAP) method has been derived to enhance performance. Since the partial MAP method allows MIMO detection to be divided into multiple stages, detection of each layer can be approached as a regression problem, and can be carried out by deep learning (DL)-based method to reduce computational overhead. Therefore, in this paper, we propose PMAP-Net, which integrates deep neural networks (DNNs) into partial MAP method for MIMO systems. We derive the soft log-likelihood ratios (LLRs) for single and multiple signals and design the input sets of DNNs. To further reduce the number of inputs in DNNs, we decrease input dimensionality by deriving extended input sets, which alleviates computational burden to be linear with respect to the number of antennas. Simulation results demonstrate that our proposed DL-based detection algorithm can provide near-optimal performance with relatively low complexity and outperforms other DL-based detectors in various MIMO scenarios.