Research on Multi-Agent Task Scheduling Optimization Based on Deep Reinforcement Learning

Abstract

For the task scheduling problem in multi-agent systems, this paper proposes a collaborative optimization method based on Graph Neural Network and Reinforcement Learning. Firstly, a heterogeneous graph structure is constructed to uniformly model the temporal dependencies, resource competition, and agent capability differences among tasks, and multi-dimensional node features are designed to fully describe the scheduling state. Secondly, the Proximal Policy Optimization algorithm is adopted to achieve efficient training and stable convergence of the policy network based on graph embedding, supporting rapid decision-making for large-scale instances. To verify its effectiveness, 30 test cases are generated for each of the three scales, totaling 90 cases. It is compared with Genetic Algorithm and Gurobi's exact solver. Through a large number of simulation experiments, the effectiveness and advantages of this method in solving the studied problem have been verified.