摘要
In bit-interleaved coded modulation (BICM) systems, the conventional demodulation assumes equal a priori probabilities for all constellation points, inherently leading to performance degradation. To enhance the performance of BICM, BICM with iterative decoding (BICM-ID) was developed. We aim to apply neural networks to BICM to achieve joint modulation and decoding and overcome the performance degradation, with the goal of surpassing the performance of BICM-ID. In this paper, we propose a neural demodulator that incorporates an additional probability layer to mitigate performance degradation in BICM systems. Furthermore, we introduce Joint Model-1, which integrates this neural demodulator with a belief propagation (BP) decoder for 16-QAM polar-coded BICM. To further improve performance, we use Joint Model-1 as a pre-trained model and extend it by adding an additional dense layer, resulting in Joint Model-2. Experimental results show that Joint Model-2 surpasses both BICM and BICM-ID systems under Ungerboeck labeling.