Publication Date

Spring 2026

Degree Type

Master's Project

Degree Name

Master of Science in Computer Science (MSCS)

Department

Computer Science

First Advisor

Genya Ishigaki

Second Advisor

Robert Chun

Third Advisor

Fabio Di Troia

Keywords

Workflow scheduling, Edge data centers, Dependent AI tasks, Reinforcement learning

Abstract

Machine learning workloads have become increasingly common in modern com- puting, driven by the rapid adoption of AI applications across domains such as vision, language, recommendation, and autonomous systems. These workloads are generally pipelines of ordered, dependent computing tasks. Combined with the existing cloud infrastructure, the emerging edge compute is expected to support the low-latency requirements of training and inference of such large machine learning workloads. In the edge-cloud continuum, such offloading requires sophisticated scheduling of task assignments, relying on accurate estimation of fluctuating link delays and execution times. Existing approaches often assume static execution times or use linear approxi- mations based only on the size of the request and do not account for contention among different resource types. To address this challenge, we propose a deep reinforcement learning (RL) based scheduling method that can react to changes in the infrastructure state and efficiently schedule AI workloads across hybrid edge-cloud topologies. Our evaluation shows that the proposed RL scheduling policy consistently outperforms baseline heuristic approaches by achieving a better latency-cost trade-off across diverse edge-cloud environments.

Download

Available for download on Saturday, May 22, 2027

Share

COinS