UL-CNN: An Unsupervised CNN Model for User Association in Wireless Networks

As the first step of Radio Resource Management (RRM), User Association (UA) is a crucial task. Traditional UA algorithm is either 'complex and impractical' or 'quick and stupid'. The emerging deep learning (DL) methods provide potential solutions to this problem. However, reinforcement learning (RL) formulates the problem as a Markov Decision Process (MDP) and requires long-term interactions with the environment, while supervised learning requires a large amount of high-quality labeled data. Therefore, we propose UL-CNN, a deep unsupervised learning method, using a convolutional neural network (CNN) to solve the UA problem in this paper. A modified loss function and a soft constraint mechanism are employed to use unlabeled data and deal with complex or even infeasible constraints. The experimental results on a multi-cell Orthogonal Frequency Division Multiple Access (OFDMA) network have demonstrated that UL-CNN can achieve promising performance in terms of system throughput with very small computational cost and probability of constraint violations. When the number or the distribution of UEs changes, the model remains scalable and resilient compared to other existing methods.