Distributed Experimental Design Networks

Abstract

As edge computing capabilities increase, model learning deployments in a heterogeneous edge environment have emerged. We consider an experimental design network, as introduced by Liu et al., in which network routing and rate allocation is designed to aid the transfer of data from sensors to heterogeneous learners. We generalize this setting by considering heterogeneous noisy Gaussian sources, incorporating multicast, but also-crucially-distributed algorithms in this setting. From a technical standpoint, we show that, assuming Gaussian sensor sources still yields an continuous DR-submodular experimental design objective. We also propose a distributed Frank-Wolfe algorithm yielding allocations within a 1-1/e factor from the optimal. Numerical evaluations show that our proposed algorithm outperforms competitors w.r.t. both objective maximization and model learning quality.