Deep learning based user association in heterogeneous wireless networks

Abstract

Due to the high splitting-gain of dense small cells, Ultra-Dense Network (UDN) is widely regarded as a promising networking technology to achieve high data rate and low latency in 5G mobile communications. In UDNs, user association is an open NP-hard problem due to the high computational complexity. In this paper, we study the user association problem from a deep learning perspective. We propose a U-Net based deep learning scheme aimed at intelligently associating user equipments(UE) to the competing Macro Base Stations (MBS) and small Base Stations (SBS). We formulate the user association problem as a constrained combinatorial optimization problem and employ a cross-entropy algorithm to obtain its asymptotically optimal solution for labelling in supervised learning. We define a differentiable loss function by combining the Mean Squared Errors(MSE) criterion and the fairness and load balancing constraints for the supervised deep learning framework. We first train the U-Net based learning model and then evaluate the accuracy of the proposed scheme. Simulation results show that the proposed U-Net scheme approaches the asymptotically optimum Genetic Algorithm (GA) scheme in terms of minimum rate gain and sum rate gain, whereas outperforms the latter with significantly reduced computation time and robustness to network scales.