Comparison of experimental designs for simulation-based symbolic regression of manufacturing systems

In this article, an empirical analysis of experimental design approaches in simulation-based metamodelling of manufacturing systems with genetic programming (GP) is presented. An advantage of using GP is that prior assumptions on the structure of the metamodels are not required. On the other hand, having an unknown structure necessitates an analysis of the experimental design techniques used to sample the problem domain and capture its characteristics. Therefore, the study presents an empirical analysis of experimental design methods while developing GP metamodels to predict throughput rates in a common industrial system, serial production lines. The objective is to identify a robust sampling approach suitable for GP in simulation-based metamodelling. Experiments on different sizes of production lines are presented to demonstrate the effects of the experimental designs on the complexity and quality of approximations as well as their variance. The analysis showed that GP delivered system-wide metamodels with good predictive characteristics even with the limited sample data.

► Empirical study of genetic programming metamodelling in discrete event simulation.
► Genetic programming gave an accuracy of 98% on average in different problem sizes.
► Genetic programming shown to be robust when space filling designs are used.