High-Dimensional Bayesian Optimization Using Greedy Variable Selection Strategy

Bayesian optimization (BO) is a popular method for solving expensive black-box problems. However, when the dimension of the problem increases, the performance of BO decreases dramatically. In this work, we propose a new greedy variable selection (Gr-VS) approach to extend BO to solve high-dimensional problems. The proposed Gr-VS approach first selects one variable among all the variables based on their effectiveness, then removes the selected variable from the variable set and selects the second variable from the remaining variables. The Gr-VS approach processes in this way until a subset of variables are selected. After that, the proposed approach performs BO on the selected subspace instead of the original design space. The proposed Gr-VS has shown very competitive performance when compared with the BO and five state-of-the-art high-dimensional BOs.