Evolutionary algorithm using progressive Kriging model and dynamic reliable region for expensive optimization problems
Suprayitno Suprayitno, Jyh‐Cheng Yu · 2016
Surrogate based optimization provides an efficient approach for expensive optimization requiring either costly experiments or time consuming simulations. However, a “good” surrogate model requires lots of training data which is impractical in applications. This work proposes an evolving algorithm, POSER, combining a progressive Kriging model and a constrained search in the dynamic reliable regions to reduce the experiments/function calls required in optimum search. The proposed algorithm starts from a Kriging model from a small sample size and improves progressively using sequential infilling samples, which often has limited generality. In general, the prediction accuracy is worse for a design farther away from the training samples. The prediction error of the Kriging model is applied to establish the reliable region to guide the evolutionary searches in the neighboring region of samples for a quasi-optimum. A hybrid infilling strategy switches between exploitation and exploration to improve sample efficiency. The reliable regional surrogate evolves and refines at the most promising regions of optimum. The process iterates until the convergence of optimum. Optimization of two benchmark numerical functions and an engineering case study are shown and compared with previous literatures. The proposed algorithm outperforms the literature results with a much smaller sample, which demonstrates the robustness and efficiency in the future applications of expensive optimization.