Experimental setup for latency optimization for reliable edge computing

Abstract

Edge computing is increasingly vital for latency-sensitive and mission-critical applications in domains such as autonomous vehicles, smart cities, and healthcare. However, maintaining consistently low average and tail latency in heterogeneous, dynamic edge environments remains a significant challenge. This thesis presents an experimentally validated framework that integrates real-world workload modeling, deep reinforcement learning (RL)-based scheduling, and predictive modeling to optimize latency and reliability in edge computing. Three representative workloads-object detection, instance segmentation, and speech-to-text conversion-are deployed on a heterogeneous edge testbed to emulate realistic computational demands and infrastructure. A key contribution is the development of an intelligent controller that uses deep RL to adaptively manage task offloading and resource allocation, informed by real-time system and network conditions. Predictive models, based on multi-layer perceptrons, estimate per-task computation and waiting times, enabling proactive bottleneck mitigation and adaptive redundancy. Experimental results demonstrate substantial reductions in both average and tail latency, as well as improved reliability across diverse workloads. This work highlights the effectiveness of combining RL-based orchestration and predictive modeling for next-generation edge systems, and suggests future directions including multi-objective optimization and scaling to large, distributed deployments.