Dynamic Resource Allocation for Multi-User Goal-oriented Communications at the Wireless Edge

This paper proposes a wireless, goal-oriented, multi-user communication system assisted by edge-computing, within the general framework of Edge Machine Learning (EML). Specifically, we consider a set of mobile devices that, exploiting convolutional encoders (CE), namely the encoder part of the convolutional auto-encoders (CAE), send compressed data units to an edge server (ES) that performs a specific learning task, such as image classification. The training of both the CEs and the ES classification networks is performed in a off-line fashion, employing a cross-entropy loss, regularized by the mean squared error of the CAE expanded output. Then, exploiting such goal-oriented architecture, and employing a Lyapunov optimization framework, we considered the joint management of computation and transmission resources for the overall system. In partic-ular, we considered a Multi-User Minimum Energy Resource Allocation Strategy (mu-MERAS), which provides the optimal resource allocation for both the devices and the ES, in a energy-efficient perspective. Simulation results highlight a classical EML trade-off between energy, latency, and accuracy, as well as the effectiveness of the proposed approach to adaptively manage resources according to wireless channels conditions, computing requests, and classification reliability.