Age of information-aware multi-tenant resource orchestration in network slicing

To satisfy diverse services from mobile users (MUs) over a common network infrastructure, network slicing is envisioned as a promising technology. This paper considers radio access network (RAN)-only slicing, where the physical RAN is judiciously tailored to accommodate computation and communication functionalities. Multiple service providers (SPs, a.k.a., tenants) compete for a limited number of channels across the discrete scheduling slots in order to serve their respective subscribed MUs. From a MU perspective, the age of information of data packets from traditional mobile services and the energy consumption at mobile device are of practical importance. We characterize the interactions among the SPs via a stochastic game, in which a SP selfishly maximizes its own expected long-term payoff. To approximate the Nash equilibrium solutions, we build an abstract stochastic game exploring the local information of SPs. Furthermore, the decision-making process at a SP can be much simplified by linearly decomposing the per-SP Markov decision process, for which we derive a deep reinforcement learning based scheme to find the optimal abstract control policies. TensorFlow-based experiments validate our studies and show that the proposed scheme outperforms the three baselines and yields the best performance in average utility.