There have been several approaches to application of machine learning to scheduling problems. Most of such techniques have focused on improving system performance based on opportunistic selection among multitudes of simple rules rather than the acquisition and analysis of meaningful knowledge as is seen in human expertise.
This study proposes a new method of learning job-dispatching rules that is adherent to the expected way in which human schedulers accumulate their expertise: establishing qualitatively meaningful criteria and quantitatively optimizing the use of them. The job-dispatching criteria used in this study were the length of processing time and the situation of the next workstation. The weighting of these criteria was trained through simulation in response to the system states. The system states are represented by multiple fuzzy partitioning of the values of the same criteria carried by the group of contending items.
The weight set of criteria that was trained through a simulation of a hypothetical FMS environment was examined for its performance and behavior. The performance shows that the proposed method can develop efficient and robust job-dispatching strategies. The behavior of the weight set demonstrates how the strategies may be semantically understood by human experts.
