An automatic tuning method for multi phased probability of rule application in tabu search for a job shop problem

The target of this research is a job shop problem with group constraints where jobs are grouped and processed together. This problem is evaluated using the average manufacturing time needed for the completion of all jobs, the number of group changes, and the number of line rests. It is a large-scale combinatorial problem where about 2000 jobs are processed. Our research group has proposed a scheduling method for repeatedly modifying solutions by applying rules to modify each evaluation factor based on the application probability of each rule using tabu search. However, the parameters set for tabu search have to be fixed using manual tuning. It takes a long time to tune the rule application probabilities because of the complex operations required for tuning and verification. Moreover, the effects of some rules are not performed often enough when the search is advancing the constant rule application probabilities. Two methods are proposed in this research. The first is dividing the application probabilities into different multi-phases depending on the progress of the search, so that an efficient improvement of the search is achieved. The second is automatically tuning the probabilities of the rule applications in each phase, analyzing the effect of the rules, and increasing the application probabilities of those rules that display high improvement ability and that indirectly leads to better solutions. The proposed method has been applied to a real large-scale job shop problem. It is confirmed that an application probability tuned by the method for 18 hours can generate the solutions that are almost equal to those obtained using an application probability tuned manually for about a week