Sequential Parameter Optimization of an Evolution Strategy for the design of Mold Temperature Control Systems

Sequential Parameter Optimization (SPO) is a popular model-assisted approach for tuning the parameters of metaheuristics, which is based on models from the Design and Analysis of Computer Experiments (DACE). Despite the indisputable success of SPO, some of the assumptions behind DACE, such as deterministic output and stationary covariance, do not hold for parameter optimization. Thus, an analysis of enhanced covariance kernels for the consideration of noise is performed. Furthermore, the effects of different sequential sampling strategies and an increasing number of replicates of each design on the quality of the models are discussed. To accomplish this, an Evolution Strategy (ES) is tuned on the real-world optimization problem of designing Mold Temperature Control Systems. Based on the results, recommendations for the ES parameters are provided, insights about the robustness of DACE with respect to the violations are made, and recommendations for appropriate combinations of sampling strategies and covariance kernels are derived.