A Bi-level Scheduling Framework for Scientific Research Resource Allocation Using MADDPG and NSGA-III

Abstract

To address key challenges in scientific research resource allocation, such as inefficient allocation, multi-objective trade-offs, poor dynamic adaptability to disruptions (e.g., sudden tasks and equipment failures), and cross-institutional unfairness, this paper proposes an intelligent scheduling mechanism that integrates MADDPG (Multi- Agent Deep Deterministic Policy Gradient) and NSGAIII. This mechanism employs a two-layer architecture: the MADDPG layer employs centralized training and distributed execution, enabling agents to collaboratively handle dynamic resource competition through local observations and achieve real-time decision-making. The NSGAIII layer utilizes reference points to find Pareto optimal solutions, balancing task time, resource utilization, cost, and fairness. A closed-loop "perception-decision-scheduling-optimization" mechanism enables large-scale allocation at the second layer. Experiments on both simulated and real-world datasets (resource-constrained, interdisciplinary tasks) demonstrate that, compared to classic algorithms (DDQN and MOEAD), it achieves a 97.5% improvement in allocation success rate, a 32.8% reduction in average task time, and a 38.6% improvement in multi-objective performance, providing an innovative solution for intelligent, efficient, and fair resource management.