BER-Aware Dynamic Resource Management for Edge-Assisted Goal-Oriented Communications

This paper investigates a BER-aware goal-oriented system design, for digital wireless communications, enabling edge-assisted learning applications. Targeting image classification tasks, we exploit banks of encoders at the transmitter, and classifiers at the edge-server that, trained under different BER conditions and exploiting knowledge of the channel state and computation load, let us to dynamically adapt, in a goal-oriented fashion, the size of the data to be transmitted, the M-QAM scheme, and the system BER, to fulfill task-specific performance. Exploiting Lyapunov optimization, we propose a minimum-energy strategy, which trades information rates for BER, under delay and classification accuracy constraints.