Unsupervised Learning for Ultra-Reliable and Low-Latency Communications With Practical Channel Estimation

• Published in: IEEE transactions on wireless communications Vol. 23; no. 4; pp. 3633 - 3647
• Main Authors: Zhang, Litianyi, She, Changyang, Ying, Kai, Li, Yonghui, Vucetic, Branka
• Format: Journal Article
• Language: English
• Published: New York IEEE 01.04.2024; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Algorithms; Artificial neural networks; Benchmarks; Channel estimation; Data communication; Data transmission; Deep learning; Efficiency; Machine learning; Network latency; Optimization; Reliability; Resource allocation; Resource management; Resource utilization; Symbols; Teaching methods; Ultra reliable low latency communication; Ultra-reliable and low-latency communications; unsupervised deep learning; Unsupervised learning
• ISSN: 1536-1276; 1558-2248
• DOI: 10.1109/TWC.2023.3309900

In this paper, we optimize the resource allocation for channel estimation and data transmission and the packet size to maximize the resource utilization efficiency subject to the constraints of ultra-reliable low-latency communications (URLLC). With practical channel estimation, the packet error probability (PEP) does not have a closed-form expression. To solve the problem, we develop novel model-based and model-free unsupervised deep learning algorithms to train a deep neural network for resource allocation and data transmission. Two types of reliability constraints are considered over a wireless link: 1) average PEP constraint; 2) constraint on the probability that PEP is higher than a threshold. The simulation results show that the learning algorithms can guarantee both types of reliability constraints. Compared with a benchmark that maximizes the number of symbols for data transmission and uses the maximum ratio transmission precoding, the learning method with the codebook-based precoding achieves a lower average signal-to-interference-plus-noise ratio (SINR), but improves the resource utilization efficiency by three times. It is because the resource utilization efficiency of URLLC is dominated by the tail distribution of SINR, not the average SINR, and the SINR of the benchmark has a much longer tail distribution than the learning method.