Topology-Aware Decision Making in Distributed Scheduling via Multi-Agent Reinforcement Learning

Abstract

This study focuses on the problem of multi-node task scheduling and resource optimization in distributed systems. A collaborative optimization method based on multi-agent reinforcement learning is proposed. By modeling the system as a partially observable multi-agent Markov decision process, each agent is enabled to make autonomous decisions based on local observations, while a global reward mechanism ensures overall optimization. In terms of algorithm design, a graph attention mechanism is introduced to enhance the modeling of inter-agent dependencies. In addition, a value function decomposition framework is adopted to improve the stability of joint policy convergence. The experimental setup is built on the real-world Cluster Trace dataset. A simulated environment is constructed to evaluate the proposed method across multiple metrics, including average task completion time, resource utilization, and system throughput. The performance is compared with traditional scheduling strategies and representative reinforcement learning algorithms. Results show that the proposed method achieves significant improvements in scheduling efficiency and resource usage. It effectively enhances the operational performance and intelligent coordination level of distributed systems.