Optimized stochastic resource allocation using graph neural networks

Stochastic resource allocation is essential in optimizing decision-making processes across various domains. This study examines a representative instance of stochastic resource allocation (heterogeneous resource allocation and task assignment), where efficient utilization and coordination of diverse resources are critical for enhancing system-level performance. Despite extensive research on resource allocation, challenges persist owing to the inherent complexity of the problem, particularly in modeling the intricate relationships among different resource types and tasks. To address these challenges, this study proposes a heterogeneous resource allocation graph neural network with greedy construction algorithm (HRAGNN-GCA), which integrates graph neural networks with a greedy algorithm. The model incorporates resource-task matching probabilities and resource collaboration factors within its message-passing mechanism and employs a greedy algorithm to efficiently construct allocation schemes. Experimental results indicate that compared with existing heuristic methods, the proposed approach achieves notable improvements in both allocation quality and computational efficiency. Furthermore, comprehensive evaluations across various problem scales and scenarios confirmed the effectiveness and generalization capability of HRAGNN-GCA, particularly in large-scale instances.