Neural network architectures for network slicing with dynamic offloading rules

Network slicing has been recently suggested as an important building block for future 5G networks, and is envisioned to allow network operators (NOs) to offer the major internet players tailor-made services (e.g., eMBB, URLLC), and in general a more flexible network management. The NOs slice their physical resources vertically from the core to the radio access network (RAN) and allocate slice-tenants a set of resources in edge and RAN. Furthermore, for efficiency reasons, the NOs track the demand that is generated by the users at the RAN, and accordingly re-scale the resources the slice-tenants need. A problem that arises is: Can the NOs decide the resource re-scaling fast and accurately? A recent approach is to place a small Neural Network (NN) in the RAN, to be used as a forecasting device. However, due to space reasons this NN cannot carry out the re-scaling very accurately. Therefore, we propose a split NN architecture: one part of the NN is located at the RAN with a local exit, and a second part, with more resources, is placed at the edge and has the final exit. Each exit is assigned a weight of importance in the training and is expected to forecast the load for all Base Stations covered by the RAN. We train an NN, and then use a reinforcement leaning-based approach, which controls which samples will be inferred locally and which remotely. Our findings validate that we can achieve interesting tradeoffs between the number of samples decided locally and the task accuracy.