Online Learning for Hierarchical Scheduling to Support Network Slicing in Cellular Networks

Abstract

We study a learning-based hierarchical scheduling framework in support of network slicing for cellular networks. This addresses settings where users and/or service classes are grouped into slices, and resources are allocated hierarchically. The hierarchy is implemented by combining a slice-level scheduler which allocates resources to slices, and flow-level schedulers within slices which opportunistically allocate resources to users/services. Optimizing the slice-level scheduler to maximize system utility is typically hard due to underlying heterogeneity and uncertainty in user channels and performance requirements. We address this by reformulating the problem as an online black-box optimization where slice-level schedulers (parameterized by a weight vector) combined with flow-level schedulers result in user/service level stochastic rewards representing performance fitness; the goal is to learn the best weight vector. We develop a bandit algorithm based on queueing cycles by building on Hierarchical Optimistic Optimization (HOO). The algorithm guides the system to improve the choice of the weight vector based on observed rewards. Theoretical analysis of our algorithm shows a sub-linear regret with respect to an omniscient genie. Finally through simulations, we show that the algorithm adaptively learns the optimal weight vectors when combined with opportunistic and/or utility-maximizing flow-level schedulers.