A Further Investigation to Improve Linear Genetic Programming in Dynamic Job Shop Scheduling

Abstract

Dynamic Job Shop Scheduling (DJSS) is an important problem with many real-world applications. Genetic programming is a promising technique to solve DJSS, which automatically evolves dispatching rules to make real-time scheduling decisions in dynamic environments. Linear Genetic Programming (LGP) is a notable variant of genetic programming methods. Compared with Tree-based Genetic Programming (TGP), LGP has high flexibility of reusing building blocks and easy control of bloat effect. Due to these advantages, LGP has been successfully applied to various domains such as classification and symbolic regression. However, for solving DJSS, the most commonly used GP method is TGP. It is interesting to see whether LGP can perform well, or even outperform TGP in the DJSS domain. Applying LGP as a hyper-heuristic method to solve DJSS problems is still in its infancy. An existing study has investigated some basic design issues (e.g., parameter sensitivity and training and test performance) of LGP. However, that study lacks a comprehensive investigation on the number of generations and different genetic operator rates, and misses the investigation on register initialization strategy of LGP. To have a more comprehensive investigation, this paper investigates different generations, genetic operator rates, and register initialization strategies of LGP for solving DJSS. A further comparison with TGP is also conducted. The results show that sufficient evolution generations and initializing registers by diverse features are important for LGP to have a superior performance.