A Machine Learning Approach to User Association in Enterprise Small Cell Networks

Abstract

Enterprise small cell networks often need to balance serving the prioritized corporate users and offering open access to guest users. Hybrid access is an efficient means to achieve this goal but how to design a user association policy to optimally achieve this tradeoff is a difficult problem. Furthermore, system dynamics in practical networks are often unknown a priori but must be learned online. To address these challenges, we first model the dynamic user association as a Markovian process, and then train two neural networks to develop a near-optimal algorithm that redistributes controllable corporate users among base stations (the "Policy Network") and learns the behavior of uncontrollable external guest user dynamics (the "Prediction Network"). The resulting system relies on the convergence of both networks and jointly they lead to a near-optimal user allocation policy. System simulations are provided to compare the performance advantages of the proposed algorithm over existing solutions. Specifically, we observe that the proposed policy can be easily applied to different reward functions.