Deep reinforcement learning and graph neural networks for efficient resource allocation in 5G networks

Abstract

The increased sophistication of mobile networks such as 5G and beyond, and the plethora of devices and novel use cases to be supported by these networks, make of the already complex problem of resource allocation in wireless networks a paramount challenge. We address the specific problem of user association, a largely explored yet open resource allocation problem in wireless systems. We introduce GROWS, a deep reinforcement learning (DRL) driven approach to efficiently assign mobile users to base stations, which combines a well-known extension of Deep Q Networks (DQNs) with Graph Neural Networks (GNNs) to better model the function of expected rewards. We show how GROWS can learn a user association policy which improves over currently applied assignation heuristics, as well as compared against more traditional Q-learning approaches, improving utility by more than 10%, while reducing user rejections up to 20%.