An Improved SARSA Learning Hyper-Heuristic Algorithm For Task Scheduling in Cloud Computing

Abstract

The increasing computational demands in modern cloud environments highlight the critical need for efficient task scheduling strategies. Task scheduling, a well-known NP-complete problem, seeks to optimize the allocation of tasks to resources while meeting predefined objectives such as minimizing completion time. This paper introduces a hyper-heuristic algorithm, Improved State Action Reward State Action (ISARSA), which uses reinforcement learning to dynamically regulate a pool of meta-heuristics, including Genetic Algorithm (GA), Cuckoo Search (CS), and Particle Swarm Optimization (PSO). SARSA, is an on-policy, high-level decision maker that is trained to prefer the most appropriate low-level heuristic based on the present system state and realized rewards, and the chosen meta-heuristic carries out the actual task-VM mapping optimization. The study is carried out in two phases: the first part involves simulation-based experiments demonstrating that ISARSA has a smoother learning behavior, adaptability, and more stable convergence in comparison to the traditional deterministic, heuristic, and Q-learning-based schedules; the second part is the implementation of the framework in a Hadoop cluster, where the same performance trends are observed and prove the robustness and practical applicability of the proposed framework.