Resource Management for Edge-Assisted Learning with Deterministic Reliability Constraints

We consider a novel resource allocation framework designed to achieve optimal resource management for edge-assisted inference tasks. This is obtained by a new optimization approach, addressed as conformal Lyapunov optimization, which integrates online conformal risk control (O-CRC) with conventional Lyapunov optimization (LO). Unlike traditional LO, this approach ensures compliance with deterministic long-term reliability constraints. Simulation results, based on an edge-assisted segmentation task, demonstrate the effectiveness of the proposed method in balancing energy consumption and inference performance, while maintaining strict control over deterministic long-term constraints, related to the false negative segmentation rate.