Deep Learning-Based Decision Region for MIMO Detection

In this work, a deep learning-based symbol detection method is developed for multi-user multiple-input multiple-output (MIMO) systems. We demonstrate that the linear threshold-based detection methods, which were designed for AWGN channels, are suboptimal in the context of MIMO fading channels. Furthermore, we propose a MIMO detection framework which replaces the linear thresholds with decision boundaries trained with neural network (NN) classifiers. The symbol error rate (SER) performance of the proposed detection model is compared against conventional methods under state-of-the-art system parameters. Here, we report to up to a 2 dB gain in SER performance using the proposed NN classifiers, allowing for exploiting higher-order modulation schemes, or transmitting with reduced power. The underlying gain in performance may be further enhanced from improvements to the NN architecture and hyper-parameter optimization.