Learning Constrained Network Slicing Policies for Industrial Applications

The proliferation of private 5G campus networks into industrial domains has increased the need for network solutions capable of simultaneously meeting a wide range of technical demands and diverse Quality-of-Service (QoS) requirements. Therefore, network-slicing (NS) is envisioned to be a key technology that allows multiple virtual networks to run on top of a common physical infrastructure. From a network management point of view, the main objective of NS is to optimally allocate network resources to different network slices. To ensure a reliable operation of industrial systems, performance constraints of industrial applications have to be considered in the resource allocation. Industrial campus networks offer an unprecedented opportunity to access the instantaneous states and performance metrics of various industrial applications. In this paper, we propose to incorporate such performance metrics as performance constraints into the optimization of network slice resource allocations. To reduce the complexity of the resulting optimization problem, we recast the resource allocation problem into a statistical learning problem and leverage reinforcement learning (RL) techniques to learn a parametrized NS policy that requires neither the knowledge of application dynamics nor exact communication models. The performance constraints are handled by reformulating the learning problem in the dual domain which we solve via a primal-dual policy gradient algorithm. We conclude the paper with numerical simulations in which we evaluate the performance of the learned network slicing policy on simulated closed-loop-control applications with varying dynamics and performance requirements.