Unfolding WMMSE Using Graph Neural Networks for Efficient Power Allocation

We study the problem of optimal power allocation in a single-hop ad hoc wireless network. In solving this problem, we depart from classical purely model-based approaches and propose hybrid method that retains key modeling elements conjunction with data-driven components. More precisely, put forth neural network architecture inspired by algorithmic unfolding iterative weighted minimum mean squared error (WMMSE) method, denote unfolded WMMSE (UWMMSE). The learnable weights within UWMMSE are parameterized using graph networks (GNNs), where time-varying underlying graphs given fading interference coefficients These GNNs trained through gradient descent approach based on multiple instances problem. show proposed is permutation equivariant, thus facilitating generalizability across topologies. Comprehensive numerical experiments illustrate performance attained along its robustness to hyper-parameter selection unseen scenarios such as different densities sizes.