Spatial domain prediction of optimal MIMO beam alignment pairs in D2D networks

The problem of resource-efficient beam alignment is a long standing one within massive MIMO (mMIMO) enabled wireless communication networks. In device-to-device enabled networks, the beam alignment problem is repeated for every new pair that appears and wishes to communicate, leading to a seemingly unbounded resource expenditure as the network grows dense. In this paper, we develop a new approach that uses the implicit geometric structure of such networks to break the spell. Instead of combatting it, our method exploits densification to facilitate alignment with minimal resources. Assuming a static (or slow varying) network, the intuition behind our approach is to utilize the beam alignment solutions at prior device pairs to predict optimal alignment in future pairs at independent new locations. We show the equivalence between this problem and a non-linear matrix completion (MC) problem under some sparsity condition. In order to solve it, we design a MC technique based on attention-based graph neural network (GNN) which proves effective to predict optimal beam pairs with little side-information.