Decentralized Wireless Resource Allocation with Graph Neural Networks
Zhiyang Wang, Mark Eisen, Alejandro Ribeiro · 2020
We consider the problem of optimal decentralized resource allocation in wireless networks. We formulate a constrained statistical learning problem to design the optimal resource allocation policy with a localized information structure. To represent the allocation policy, we propose the use of Aggregation Graph Neural Networks (Agg-GNNs), which take a sequence of graph aggregated state information obtained locally at each transmitter from multi-hop neighbors as an input. Moreover, the aggregation sequence with respect to an interference graph is formed naturally through wireless transmission. We demonstrate a permutation equivariance property of the resulting resource allocation policy and train with an unsupervised, model free primal dual method. The proposed strategy is verified by numerical simulations compared with several state-of-the-art methods.