Task Scheduling Research for Cloud Computing Infrastructures: A Reinforcement Learning Method Integrating Resource Occupancy Prediction, Dynamic Node State Encoding, and Placement Profit Maximization

Abstract

To address the challenges of random task arrival, fluctuating resource consumption, continuous node state evolution, and the difficulty of balancing immediate feasibility with long-term placement benefits in cloud computing facilities, this paper proposes a task scheduling method that integrates resource consumption prediction, dynamic node state encoding, and reinforcement learning decision-making. First, a resource consumption prediction mechanism is constructed based on historical task demand information to enhance the scheduling process's ability to perceive future resource pressure. Then, node load, remaining capacity, and runtime context are dynamically represented to improve the accuracy of state modeling in complex environments. Furthermore, an adaptive matching between tasks and nodes is achieved through a reinforcement learning strategy aimed at maximizing placement benefits, thereby improving resource utilization efficiency and scheduling stability. This method unifies prediction, representation, and decision-making within a single framework, enabling a more effective characterization of the coupling relationship between changing task demands and node state evolution. Comparative experimental results demonstrate that the proposed method exhibits superior overall performance in resource utilization, service assurance, energy consumption control, and scheduling response, validating its effectiveness and applicability in cloud computing task scheduling scenarios.