Objective satisfaction assessment using neural nets for balancing multiple objectives

In today's production systems, improving the use of the manufacturing resources and reacting efficiently to disturbances leads to schedules more and more adapted to the considered workshop. A generic software can hardly take into account the specificity of each workshop: in that context, it is not sufficient anymore to provide a feasible schedule, and human expertise becomes necessary in order to improve the provided solution. This improvement requires the definition of synthetic performance indicators allowing us to assess a schedule before choosing improvement actions. Many performance indicators have been defined; however, they are seldom structured in order to supply a complete and progressive assessment framework. We suggest in this paper a parametrable hierarchic structure of performance indicators allowing us to aggregate the degree of satisfaction of elementary objectives, thus allowing the definition of a compromise between these elementary objectives. Neural networks have been tested in order to emulate the expertise involved in the definition of such compromises. Neural networks enable us to express the satisfaction provided by a schedule in a synthetic way, then to describe the satisfaction of the elementary objectives in order to select improvement actions. Using the same indicator values, several aggregation strategies can be considered and stored in order to adapt the assessment phase to the global situation of the workshop (e.g. in the presence of overloads, under loads, rush orders, lateness, bottlenecks, etc.). The implementation of this method in an industrial scheduler, called IO, is in progress.