Goal-Oriented Water-Filling for Dynamic Management of Edge-Assisted OFDM Communications

Goal-oriented communications aim to deploy efficient learning architectures at the wireless edge, by parsimoniously using transmission resources, while possibly guaranteeing prescribed learning accuracy and delay. Aim of this paper is to propose a practical goal-oriented communications framework, which exploits an OFDM physical layer to cope with frequency selective fading channels. Targeting an image classification task, where images are scheduled according to an unknown arrival process, we propose a dynamic and goal-oriented data-loading of the wireless OFDM link that, exploiting banks of convolutional encoders at the user equipment, is capable to extract data features, with time-adaptive low dimension, that are meaningful to the classification task performed at the edge-server. The transmission of these features to the edge-server is dynamically optimized by exploiting knowledge of the channel state, computation load and images backlog, resorting to a Lyapunov stochastic optimization framework, which jointly optimizes the communication resources, the computation energy, and the encoders to be used. Specifically, we propose a minimum energy strategy under prescribed delay and accuracy of the edge-server learning task. Simulation results testify the effectiveness of the proposed strategy.