Publication Details
Overview
 
 
Seppe Renty, Avalos Martinez de Escobar, Raphaël, Andries Rosseau,
 

Proceedings of the Benelux Conference on Artificial Intelligence

Contribution To Journal

Abstract 

Job scheduling is a central element of our society. While this problem has been actively researched by the field Operations Research (OR), yielding good algorithms, these solutions are usually not generalizable across different problem instances. They also tend to behave poorly when there is uncertainty during execution. In this thesis, we research whether dispatching rules that generalize to groups of similar problem instances can be learned through Reinforcement learning. To leverage the complex structure of a Job scheduling problem we designed a graph representation of the problem. We then used a combination of Graph Neural Networks (GCN) and Multi-Agent Reinforcement Learning (MARL) to improve on existing dispatching rules. We found that the technique generates shorter schedules on average than popular dispatching rules for multiple families of problems.

Reference 
 
 
Link