Development of a real-time learning scheduler using reinforcement learning concepts

A scheme for the scheduling of flexible manufacturing systems (FMS) has been developed which divides the scheduling function (built upon a generic controller architecture) into four different steps: candidate rule selection, transient phenomena analysis, multicriteria compromise analysis, and learning. This scheme is based on a hybrid architecture which utilizes neural networks, simulation, genetic algorithms, and induction mechanism. This paper investigates the candidate rule selection process, which selects a small list of scheduling rules from a larger list of such rules. This candidate rule selector is developed by using the integration of dynamic programming and neural networks. The system achieves real-time learning using this approach. In addition, since an expert scheduler is not available, it utilizes reinforcement signals from the environment (a measure of how desirable the achieved state is as measured by the resulting performance criteria). The approach is discussed and further research issues are presented