Mixed-Bit-Width Quantized Neural Network for Coded MISO System in UAV Communications

This letter proposes a low-complexity mixed-bit-width (MBW) - quantized neural network (QNN) for coded multiple-input-single-output (MISO) systems in unmanned aerial vehicle (UAV) communications. Firstly, to solve the NP-hard problem of conventional maximum a posteriori (MAP)-based MISO detection, a learning-based QNN is proposed to fit the mapping of inputs and outputs in the MISO detector, whose computational complexity increases linearly with the number of antennas and modulation order. Secondly, neural networks with a mixture of varying activation and weight bit widths have been proposed. Specifically, the weight bit width has been further compressed to 3 bits, enabling the replacement of multipliers with adders and significantly reducing computational complexity. Finally, the simulation results demonstrate the effectiveness of our proposed MBW-QNN based detection algorithm, whose bit error rate (BER) performance is only 0.8dB away from optimal floating-point MISO detection algorithm.