Evaluation of a Task Offloading Simulator for Edge Resource Management: Comparison of Reinforcement Learning Algorithms

This paper aims to design and evaluate an autonomous task offloading system for edge computing capable of preventing CPU overload and maintaining system stability under highly dynamic workload conditions. The work introduces both a baseline workload and a predictive Non-Homogeneous Poisson Process (NHPP) generator to reproduce realistic patterns of traffic escalation and saturation, allowing a more accurate assessment of algorithm robustness under practical conditions. Within this framework, two RL schedulers are implemented and trained using a reward function designed to encourage the stability of the system. The performance of the system is then assessed along multiple aspects, including task completion, load distribution, migration behavior, and overall operational stability.