Accelerated Reinforcement Learning for Real-Time Task Scheduling in Edge Computing Environments

Key Speaker

Abstract

As soft real-time applications (SRTAs) grow in complexity, leveraging edge computing (EC) is essential to meet their resource demands. The performance of SRTAs depends heavily on task scheduling algorithms that manage workload offloading to edge-servers. However, task scheduling in EC remains challenging due to exponential search spaces, multi-objective trade-offs, and environmental dynamism. While traditional heuristic and metaheuristic algorithms often struggle with these complexities, reinforcement learning (RL) has emerged as a promising alternative. This study introduces Pruned Reinforcement Learning (PRL), a novel RL-based scheduling method for SRTAs in EC. Standard RL approaches often face scalability issues in medium- and large-scale problems because conventional Markov Decision Processes (MDPs) involve vast state-action spaces that make policy derivation computationally expensive. PRL addresses this by reducing the MDP size to accelerate learning. Experimental results show that compared to current methods, PRL improves the hit-ratio by 12% while reducing runtime overhead by 52%, memory footprint by 27%, and power consumption by 35%.

Authors

1. Amin Avan amin.avan@ontariotechu.net (Ontario Tech University)
2. Akramul Azim Akramul.Azim@ontariotechu.ca (University of Ontario Institute of Technology (UOIT), Canada)
3. Qusay H. Mahmoud qusay.mahmoud@ontariotechu.ca (Ontario Tech University, Canada)

 Overview  Program