Learning-Assisted Optimization for Energy-Efficient Scheduling in Deadline-Aware NOMA Systems

In this paper, we study a class of minimum-energy scheduling problems in non-orthogonal multiple access (NOMA) systems. NOMA is adopted to enable efficient channel utilization and interference mitigation, such that base stations can consume minimal energy to empty their queued data in presence of transmission deadlines, and each user can obtain all the requested data timely. Due to the high computational complexity in resource scheduling and the stringent execution-time constraints in practical systems, providing a time-efficient and high-quality solution to 5G real-time systems is challenging. The conventional iterative optimization approaches may exhibit their limitations in supporting online optimization. We herein explore a viable alternative and develop a learning-assisted optimization framework to improve the computational efficiency while retaining competitive energy-saving performance. The idea is to use deep-learning-based predictions to accelerate the optimization process in conventional optimization methods for tackling the NOMA resource scheduling problems. In numerical studies, the proposed optimization framework demonstrates high computational efficiency. Its computational time is insensitive to the input size. The framework is able to provide optimal solutions as long as the learning-based predictions satisfy a derived optimality condition. For the general cases with imperfect predictions, the algorithmic solution is error-tolerable and performance scaleable, leading the energy-saving performance close to the global optimum.