Combined Usage of Convex Optimization and Neural Network for Resource Allocation

Coping with the continually increasing volumes of data being handled by mobile networks requires the use of enabling technologies, such as small cell deployments and additional higher frequencies. The network needs to carefully associate the users to the base stations (BSs) and frequency bands (we call resource allocation in the paper), since users can connect multiple BSs and multiple frequency bands in such deployments. Convex optimization can achieve the optimum resource allocation, although its application in systems is challenging. This is because the optimizer located at the network needs to have prior knowledge of the achievable data rates for all users to all BSs. Meanwhile, the modification of the user association criteria, in which different offset values are applied to different cells, can realize similar optimization without the knowledge of the achievable data rate. Therefore, this study combined these two schemes and used a neural network-based optimizer to optimize cell-specific offsets. A good trade-off between performance and user feedback is confirmed by simulation results in small cell deployments.