Learning-Based Maximum Likelihood Detection for Uplink Massive MIMO with One-Bit ADCs: Biased and Dithering Approaches
The maximum likelihood (ML) detection based on learning is studied in this work for uplink massive multiple-input and multiple-output (MIMO) systems with one-bit ADCs (analog-to-digital converters). Typically, in a learning based detection scheme there is a limitation on the maximum training length. To address the challenge, we provide two one-bit ML detection schemes, a controlled learning scheme and a noise injection scheme. The controlled learning scheme then does not allow learning to erase the information obtained after learning of the zero likelihood functional, which translates to a better detection performance. Bringing this technique further to a scenario with knowledge of received signal to noise ratio and extending the controlled learning approach, the noise injection scheme uses more quantizer input by adding noise to the input to use more likelihood functions. Implementation of the methods has been enhanced by the post detection interference cancellation technique that uses data symbols of correctly decoded code words as training data. The methods developed do not require any channel estimation. Performance analysis through simulation showed that the symbol error rate detection performance for the MLD methods utilized in the present methods has been achieved.