Comparison and Contrast of two Machine Learning Assisted In-Device Scheduling Applications

Abstract

Managing offloaded tasks efficiently is vital for keeping devices fast and responsive. The main hurdle is finding a way to schedule and run these incoming workloads without starving local applications of critical CPU, RAM, and energy. This study evaluates how Federated Learning (FL) and Reinforcement Learning (RL) can be applied directly at the device level to optimize task scheduling within IoT ecosystems. Devices learn to optimize tasks execution locally while trying to maintain its primary functionality, thus benefiting stability of the entire system. An experimental comparison indicated that RL achieved lower task roundtrip and waiting time, within the range 15-30%, however at a significantly higher CPU/RAM/storage overhead. At the same time FL showed to be viable on constrained devices due to much lower resource consumption. The contribution of this research lies in the investigation of ML-assisted frameworks for optimizing in-device scheduling of offloaded tasks in heterogeneous IoT environment.