Optimization-Based Quantized Federated Learning for General Edge Computing Systems

This paper investigates optimal implementations of federated learning (FL) in practical edge computing systems with possibly distinct computing and communication resources at the server and workers. First, we present a new random quantization scheme and analyze its properties. Then, we propose a general quantized FL algorithm, namely HQFedWAvg, and analyze its convergence. HQFedWAvg adopts the proposed quantization scheme and a generalized mini-batch stochastic gradient descent (SGD) method and has several adjustable algorithm parameters to maximally adapt to the computing and communication resources at the server and workers. Next, we optimize the algorithm parameters of HQFedWAvg. The resulting challenging non-convex optimization problem is successfully tackled using several optimization techniques. Numerical results demonstrate HQFedWAvg's considerable performance gains over existing FL algorithms and interpret its function principle.