Unsupervised Learning-Based Low-Complexity Integrated Sensing and Communication Precoder Design

This study proposes an unsupervised deep learning-based (DL-based) approach to precoding design for integrated sensing and communication (ISAC) systems. Designing a dynamic precoder that can adjust the trade-off between the sensing performance and communication capacity for ISAC systems is typically highly compute-intensive owing to requiring solving non-convex problems. Such complex precoders cannot be efficiently implemented on hardware to operate in highly dynamic wireless environments where channel conditions rapidly vary. Accordingly, we propose an unsupervised DL-based precoder design strategy that does not require a data set of the optimum precoders for training. The proposed DL-based precoder can also adapt the trade-off between the communication sum rate and sensing accuracy depending on the required communication and/or sensing performance. It offers a low-complexity precoder design compared to conventional precoder design approaches that require iterative algorithms and computationally intensive matrix operations. To further reduce the memory usage and computational complexity of the proposed precoding solution, we have also explored weight quantization and pruning techniques. The results have shown that a quantized and pruned deep neural network (DNN) can achieve 96% of the sum rate achieved by the full DNN while its memory and computational requirements are less than 17% of the full DNN.