Time-varying hyperparameter strategies for radial basis function surrogate-based global optimization algorithm

摘要

Radial Basis Function (RBF) surrogate-based global optimization has been shown to be efficient for complex problems with computationally expensive and high-dimensional functions. Based on the DYCORS (DYnamic COordinate search using Response Surface models) algorithm framework, this paper proposes two Time-Varying Hyperparameter DYCORS (TVH-DYCORS) strategies to accelerate RBF surrogate-based optimization algorithms, which include a time-varying perturbation strategy and a time-varying weight pattern strategy. The TVH-DYCORS algorithm is evaluated by a 124-variable benchmark problem from the automotive industry as well as six other high-dimensional optimization test problems. The computational results demonstrate that the proposed algorithm has potential to achieve better solutions, compared with conventional genetic algorithm and two previously proposed RBF surrogate-based optimization algorithms.