Joint user grouping and power allocation for MISO systems: Learning to schedule

In this paper, we address a joint user scheduling and power allocation problem from a machine-learning perspective in order to efficiently minimize data delivery time for multiple-input single-output (MISO) systems. The joint optimization problem is formulated as a mixed-integer and non-linear programming problem, such that the data requests can be delivered by minimum delay, and the power consumption can meet practical requirements. For solving the problem to the global optimum, we provide a solution to decouple the scheduling and power optimization. Due to the problem's inherent hardness, the optimal solution requires exponential complexity and time in computations. To enable an efficient and competitive solution, we propose a learning-based approach to reduce data delivery time and solution's computational delay, where a deep neural network is trained to learn and decide how to optimize user scheduling. In numerical study, the developed optimal solution can be used for performance benchmarking and generating training data for the proposed learning approach. The results demonstrate the developed learning based approach is able to significantly improve the computation efficiency while achieves a near optimal performance.