Adaptive Resource Scheduling in Distributed Computing via Multi-Agent Reinforcement Learning and Graph Convolutional Modeling

Abstract

This paper proposes an adaptive scheduling method based on multi-agent reinforcement learning to address the complexity and uncertainty of resource dynamic scheduling in distributed computing environments. The method centers on a multi-agent collaboration mechanism, modeling computing nodes as agents with local perception and autonomous decision-making capabilities. Through a centralized training and decentralized execution (CTDE) framework, it achieves a dynamic balance between global optimization and local autonomy. The model incorporates a graph convolutional network to capture topological connections and resource dependencies among nodes, while a joint value function enables global coordination of multi-agent policies, improving cooperative perception and policy convergence efficiency. A multi-level state representation and temporal modeling mechanism are designed, combined with a dynamic reward strategy to handle environmental non-stationarity, enabling adaptive optimization for task load fluctuations, network bandwidth changes, and node failures. Comparative experiments on real distributed task data demonstrate significant improvements in Average Completion Time, system throughput, and resource utilization. Further sensitivity analysis shows that factors such as learning rate, bandwidth limit, task length distribution skewness, and historical window size greatly influence model performance, indicating strong stability and generalization in complex distributed computing scenarios. The results confirm that the proposed framework effectively achieves task-resource matching optimization in highly dynamic environments, providing a practical learning-based scheduling solution for the efficient operation of intelligent distributed systems.