On the scalability of meta-models in simulation-based optimization of production systems

Optimization of production systems often involves numerous simulations of computationally expensive discrete-event models. When derivative-free optimization is sought, one usually resorts to evolutionary and other population-based meta-heuristics. These algorithms typically demand a large number of objective function evaluations, which in turn, drastically increases the computational cost of simulations. To counteract this, meta-models are used to replace expensive simulations with inexpensive approximations. Despite their widespread use, a thorough evaluation of meta-modeling methods has not been carried out yet to the authors' knowledge. In this paper, we analyze 10 different meta-models with respect to their accuracy and training time as a function of the number of training samples and the problem dimension. For our experiments, we choose a standard discrete-event model of an unpaced flow line with scalable number of machines and buffers. The best performing meta-model is then used with an evolutionary algorithm to perform multi-objective optimization of the production model.