End-to-end supervised learning for NOMA-enabled resource allocation: A dynamic and scalable approach

With the integration of wireless networks and artificial intelligence, many network problems have new solving paradigms. This drives us to rethink the classic resource allocation problem in non-orthogonal multiple access (NOMA) systems from a new perspective. Traditional methods often utilize an iterative structure to address this problem, and numerous insights have been uncovered in static environments. However, network environments are inherently dynamic and scalable, rendering resource allocation an independent mixed-integer programming problem at any given time. Therefore, breaking away from traditional optimization planning and employing an end-to-end technique to tackle such problems is highly valuable. To simplify the problem, we first divide the resource allocation problem into two components, namely the channel assignment problem and the power allocation problem. Then we derive the closed-loop expression of the optimal power allocation scheme. The channel assignment problem is optimized by invoking Pointer Network (PTN), whose unique structure offers us the possibility of online solutions. Also, the neighboring context, which is considered to impact decision-making positively, is added by employing Graph Pointer Network (GPN). Due to the short inference time of neural networks, the most exciting advantage of our methods is that, whether the locations of the users change or the number of users changes, our algorithm can give a near-optimal solution immediately. Numerical results show the effectiveness of the proposed methods, especially their strong generalization.