Low Complexity Deep Learning-Assisted Golden Code Sphere-Decoding with Sorted Detection Subsets

Abstract

Golden code is a space-time block coding (STBC) scheme that has spatial multiplexing gain over the Alamouti STBC which is widely used in modern wireless communication standards. Golden code has not been widely adopted in modern wireless standards because of its inherent high detection complexity. However, detection algorithms like the sphere-decoding with sorted detection subsets (SD-SDS) have been developed to lower this detection complexity. Literature indicates that the SD-SDS algorithm has lower detection complexity relative to the traditional sphere-decoding (SD) algorithm, for all signal-to-noise ratio (SNR) values. The SD-SDS algorithm exhibits low detection complexity at high SNR; however, at low SNR the detection complexity is higher. We propose a deep neural network (DNN) aided SD-SDS algorithm (SD-SDS-DNN) that will lower the Golden code's SD-SDS low SNR detection complexity, whilst maintaining the bit-error-rate (BER) performance. The proposed SD-SDS-DNN is shown to achieve a 75% reduction in detection complexity relative to SD-SDS at low SNR values for 16-QAM, whilst maintaining the BER performance. For 64-QAM, the SD-SDS-DNN achieves 99% reduction in detection complexity relative to the SD-SDS at low SNR, whilst maintaining the BER performance. The SD-SDS-DNN has also shown to achieve low detection complexity comparable to that of the Alamouti linear maximum likelihood (ML) detector for a spectral efficiency of 8 bits/s/Hz. For a spectral efficiency of 12 bits/s/Hz, the SD-SDS-DNN achieves a detection complexity that is 90% lower than the Alamouti linear ML detector.