A Pattern Recognition Method for Solving Large Scale Mathematical Programming Problems with Applications to Scheduling

Abstract This paper presents a technique for making the selection of the dynamic decisions involved in scheduling through a pattern recognition method. This is done in order to provide the best possible decisions for the scheduling problem without requiring the solution of the large scale optimization problem of a model on line. The decisions to be used for the system are obtained by off-line calculations. The resulting sets of possible decisions are then divided into several groups in such a way that each member within a group shares a set of common characteristics. Then, a statistical pattern classification technique is applied to classify the groups of decisions on the basis of the features describing the state of the scheduling system. Finally the results obtained by each of these groups of decisions are compared with plant data to show which set of decisions should be made for any given plant situation. Such a pattern recognition approach to the scheduling of ingot processing in soaking pits and slabbing mills in steel production is used as an example in this paper. At present the required optimization model for scheduling the system demands far too much computing time to be of use in on-line decision making. However once the classifier described herein has been determined, the pattern recognition approach may be exploited to overcome the dimensionality and computational requirements in such a massive dynamic scheduling problem.