Joint Mode Selection And Power Allocation for NOMA Systems With D2D Communication

In this paper, we study the integration of device-to-device communication into a non-orthogonal multiple access system. To deal with the complex co-channel interference resulting from the dense spectral reuse, we aim to maximize the sum proportional bit rate by jointly optimizing mode selection (MS) and power allocation (PA). Considering the high complexity of the original problem and the dynamics of the wireless environment, we propose an online mechanism with a double-layer structure by efficiently combining machine learning with optimization theory. In particular, when the MS scheme is given, the remaining nonconvex PA problem can be equivalently transformed into a convex one under certain manipulations. Based on the above optimum, a deep reinforcement learning-based online mechanism is designed and it constantly refines the output MS scheme generated from a deep neural network by utilizing the recent historical experiences via reinforcement learning. Finally, simulations are conducted to validate the superiority of the proposed mechanism in balancing the fundamental tradeoff between performance and online computational time.