Crowding Game and Deep Q-Networks for Dynamic RAN Slicing in 5G Networks

Fifth generation (5G) mobile networks do not solely support services with very high throughput, but also answer the requirements of other heterogeneous services, characterized by different Quality of Service (QoS) criteria such as maximum tolerated delay, minimum guaranteed throughput, or capacity constraints. This heterogeneity necessitates the partitioning of the available radio resources into multiple slices, where each slice is characterized by specific QoS and isolation constraints. Nonetheless, a static resource allocation scheme among slices might not be an effective solution to face the dynamic and sporadic nature of traffic and users' service types. Therefore, a dynamic slicing scheme that leverages on the ease of slice selection per user in the fully virtualized open Radio Access Networks (RAN), is paramount to accommodate the various demands and load conditions. The main contribution of this paper is to apply traffic engineering in the scope of RAN slicing. In fact, instead of solely re-dimensioning the slices by injecting more resources when congestion hits, we also allot users to slices that may differ from their service type if their performance target is met. This is a novel definition of dynamic slicing, the dynamicity not being in re-dimensioning on the fly but in accommodating active users to existing resources for higher resource utilization efficiency without hindering users' performances. Additionally, the available bandwidth is dynamically adjusted among slices based on their load and affected users' performances. To reach that goal, two Dynamic RAN Slicing schemes are proposed: a centralized scheme based on Deep Reinforcement Learning (DRL) and a distributed scheme based on non-cooperative game theory. Exhaustive numerical simulations demonstrate the high efficiency of our proposed approach in comparison with the state-of-the-art where users are allocated to slices according to their service type and required Key Performance Indicators (KPIs).