Model Comparative Analysis of Neighborhood Aggregation Levels in Graph Neural Networks for Metaschedulers

In Time-Triggered Systems (TTS), meta-scheduling algorithms are crucial for adjusting to situations that necessitate rescheduling, including hardware malfunctions and changes in operational modes. However, the state-space explosion problem in Cyber-Physical Systems (CPS) causes immense storage demands. This paper extends previous research from 1 2, studying the impact of neibourhood aggregations of Graph Neural Networks (GNN) in meta-scheduling. The focus is on multi-neighbourhood aggregations and their impact on makespan results in task scheduling. The paper presents a comparative analysis between Genetic Algorithms (GA), a GNN-based approach, and traditional Artificial Neural Networks (ANNs). The results highlight the significance of second neighbourhood aggregation as a compromise for event-driven Multi-Schedule Graphs (MSG). Out of the considered machine learning solutions, GNNs are the most feasible one for meta-scheduling in terms of makespan to model parameter cost (number of adjustable weights).