Training a neural network to select dispatching rules in real time

Dispatching rules are often suggested to schedule manufacturing systems in real-time. Numerous dispatching rules exist. Unfortunately no dispatching rule (DR) is known to be globally better than any other. Their efficiency depends on the characteristics of the system, operating condition parameters and the production objectives. Several authors have demonstrated the benefits of changing dynamically these rules, so as to take into account the changes that can occur in the system state. A new approach based on neural networks (NN) is proposed here to select in real time, each time a resource becomes available, the most suited DR. The selection is made in accordance with the current system state and the workshop operating condition parameters. Contrarily to the few learning approaches presented in the literature to select scheduling heuristics, no training set is needed. The NN parameters are determined through simulation optimization. The benefits of the proposed approach are illustrated through the example of a simplified flow-shop already published. It is shown that the NN can automatically select efficient DRs dynamically: the knowledge is only generated from simulation experiments, which are driven by the optimization method. Once trained offline, the resulting NN can be used online, in connection with the monitoring system of a flexible manufacturing system.