Designing dispatching rules with genetic programming for the unrelated machines environment with constraints

Abstract

Scheduling problems constitute an important part in many everyday systems, where a variety of constraints have to be met to ensure the feasibility of schedules. These problems are often dynamic, meaning that changes occur during the execution of the system. In such cases, the methods of choice are dispatching rules (DRs), simple methods that construct the schedule by determining the next decision which needs to be performed. Designing DRs for every possible problem variant is unfeasible. Therefore, the attention has shifted towards automatic generation of DRs using different methods, most notably genetic programming (GP), which demonstrated its superiority over manually designed rules. Since many real world applications of scheduling problems include various constraints, it is required to create high quality DRs even when different constraints are considered. However, most studies focused on problems without additional constraints or only considered them briefly. The goal of this study is to examine the potential of GP to construct DRs for problems with constraints. This is achieved primarily by adapting the schedule generation scheme used in automatically designed DRs. Also, to provide GP with a better overview of the problem, a set of supplementary terminal nodes is proposed. The results show that automatically generated DRs obtain better performance than several manually designed DRs adapted for problems with constraints. Using additional terminals resulted in the construction of better DRs for some constraints, which shows that their usefulness depends on the considered constraint type. Therefore, automatically generating DRs for problems with constraints presents a better alternative than adapting existing manually designed DRs. This finding is important as it shows the capability of GP to construct high quality DRs for more complicated problems, which is useful for real world situations where a number of constraints can be present.