Predictive State-aware Deep Reinforcement Learning with Hyper-Heuristic for Resolving Conflicting Objectives in Scientific Workflow Scheduling

Scientific workflows in cloud environments are highly complex and dynamic, necessitating intelligent and flexible scheduling solutions to handle important factors, including resource heterogeneity, budget constraints, deadlines, and ever-changing workload requirements. In contrast to traditional scheduling methods, a model with predictive ability and adaptability is required to manage these complex challenges effectively. Hence, this work intends to address the challenges associated with scheduling scientific workflows, balancing conflicting objectives, such as cost and deadline, managing intricate inter-task workflow dependencies, and dynamic resource availability. To accomplish this goal, this work presents a predictive state-aware deep reinforcement learning with a hyper-heuristic for deadline and budget-aware workflow scheduling optimization. Initially, the proposed system applies a Multihead Graph Attention Network (MGAN) to describe complex interactions between workflow tasks and cloud resources in order to predict the state for modeling an accurate environment. Moreover, the design of hyper-heuristic generation with Deep Q-Network (DQN) improves deadline and budget-aware decision-making in the uncertain workflow scheduling environment. Experiments show that the proposed method outperforms state-of-the-art approaches in terms of deadline and cost-effectiveness, providing a reliable and intelligent strategy for scheduling scientific workflows.