Graph-Guided Data Generation for Deep Learning Based Scheduling in Wireless Access Networks

We consider the problem of utility-optimal link scheduling in a time-slotted multiple access point (AP) wireless network. In each slot, given the received signal strengths (RSS) from all potential transmitters to all receivers, and the link-wise scheduling weights, the objective is to select a subset of communication links such that the sum of the weighted long-term average link throughputs is maximized. Exhaustive search for the optimal configuration, in each slot, is computationally infeasible even with a few APs and stations, due to exponential complexity.We introduce a two-stage approach that combines graph-based approximation algorithms with deep learning models. In the data generation stage, in each slot, the scheduling problem is approximately reduced to finding a maximal independent set on an interference graph, which, in practice, yields an order of magnitude reduction in schedule search time. We parametrically unify three heuristics for generating the interference graph in each slot and obtain the best parameter set. The generated dataset is used to train a deep neural network (DNN) to predict link-sets when given as input the RSS values (we assume equal link weights in this paper) during the actual operation of the wireless network.The graph-based data generation method achieves over 98% of the optimal network utility with a per slot computation that is at least an order of magnitude smaller than optimal data generation but still requires milliseconds per slot, while the DNN trained on this data, in online usage, achieves over 94% of the optimal utility with about 100−200 µsec computation time per slot on commodity hardware. These results confirm that the hybrid approach offers a practical trade-off between computational efficiency and scheduling performance, making it well-suited for scaling to practical wireless network environments.