- 1 - Genetic Operators Based on Constraint Repair * Jürgen Dorn and Mario Girsch Christian Doppler Laboratory for Expert Systems Vienna University of Technology Paniglgasse 16 A-1040 Vienna In this paper we describe an approach to solve a scheduling problem in the steel making industry with a combination of a constraint repair approach and genetic algorithms. Because several, sometimes conflicting objectives exist in the production of steel, we have proposed a representation of con- straint violations and their importance by fuzzy sets (Dorn and Slany 1994). A weighted aggregation of the violations gives us a means to compare schedules. Furthermore, as pointed out by Minton et al. (1990) the strategy to repair constraints in order to achieve better schedules is a good heuristic for large applications. We have therefore developed domain independent genetic operators that apply knowledge of constraint violations. We report on experiments that show the improvement of the combination for our application and draw some conclusions. 1 Introduction Many years of research in scheduling were dominated by the search for techniques to achieve optimal schedules. Especially for flow shops there was the hope of finding ef- ficient algorithms because flow shops are structured more strictly than general job shops (Dudek et al. 1992). With the rise of complexity theory (Garey and Johnsson 1979), it was found out that even for flow shops with a number of jobs and machines typical for real applications, the problem is intractable. It must be stated that the complexity of the task stems from the sequencing of operations and not from the assignment of times to operations (French 1982). Consequently, problem solving methods often apply heuristics to find solutions that are not necessary optimal. We can distinguish between heuristic domain knowledge and general heuristics. Although domain heuristics are often more efficient, one tries to find general heuristics to enable an easy reuse of scheduling methods. Recently, some new scheduling procedures occurred that are based on the idea to improve an existing schedule iteratively. By simple modifications that do not cost too much a schedule is changed and evaluated anew. If this new schedule is better than the old, this procedure is repeated. By iteration a search for a good schedule is performed. * This paper was published in the Proceedings of the ECAI’94 Workshop on Genetic Algorithms and other Evolutionary Algorithms, Amsterdam and is considered for publication in LNCS.