DNN-Based Physical-Layer Network Coding with Low-Complexity Codec

Physical-layer network coding (PNC) is considered as a good option given its ability in doubling the throughput of a two-way relay network (TWRN). In uplink of PNC, two end nodes’ packets are transmitted simultaneously, but the two channel gains are usually irrelevant, hence it is sub-optimal to adopt fixed-size regular QAM-type constellations for PNC system as in conventional designs. Recently, machine learning based wireless communication system becomes an important feature for 6G wireless network. Deep neural networks (DNNs) can generate irregular constellation points accordingly and automatically to deal with the performance damage caused by the different phase shifts in both uplinks. However, DNN-based PNC systems in recent literatures have high complexity to implement, and reducing the computational complexity of DNN-based PNC architecture is of great importance for the system to be practical. In this paper, we proposed a DNN-based low-complexity codec PNC system, named DLC-PNC. Compared with the state-of-the-art DNN-based PNC schemes, DLC-PNC has a much simpler structure and requires less training parameters while having the same bit-error-rate (BER) when using 4-QAM-like modulation and having lower BER rate when using 16-QAM-like modulation when signal-to-noise ratio (SNR) is larger than 15 dB. Next, we designed a sparse training algorithm to prove that the structure of DLC-PNC can be further pruned to decrease the training parameters. Simulation results show that around 50% parameters can be pruned while maintaining the similar system performance, which further reduces the time and space complexity.