5.4 Machine Learning-Enabled 5G Network Slicing

In this contribution, the different parts of the end-to-end network slicing concept are presented, including the Core Network (CN) and the Radio Access Network (RAN), while highlighting the differences and similarities of both domains. Further, prototypical implementations and empirical evaluations of 5G network slicing are discussed, deepening the understanding of network slicing and identifying possible advantages and challenges. The predictability of user traffic in the respective network slices poses such a challenge, as resources in the RAN-in contrast to resources in the CN-are subject to fluctuations based on channel quality. Critical infrastructures typically require very low latencies in the single-digit milliseconds range and are thus considered ultra-Reliable Low Latency Communication (uRLLC) . To mitigate latencyintensive scheduling requests and grant operations, resources in the RAN have to be pre-allocated for uRLLC slices. This operation, also known under the term Configured Grants (CGs), pre-allocates resources for, say, high-priority slices, so that User Equipments (UEs) are able to send data without asking for resources, which reduces the scheduling latency down to zero. The simplest method for calculating CGs is based on static allocations, which has one major drawback: unused resources are wasted, and thus, can not be used by remaining slices, effectively lowering spectral efficiency. Here, we present SAMUS (Slice-Aware Machine Learning-based Ultra-Reliable Scheduling), a data-driven method to predict resources in the future based on real data, e.g., solar activity in smart grid slices, to reduce latencies while maintaining high spectral efficiency.