Yubin Tian, Professor, Beijing Institute of Technology

Zhongzhi Bai, Professor

Parallel Efficient Global Optimization by Using the Minimum Energy Criterion

10:00-11:00 October 15(Saturday), Tencent Meeting ID: 535-573-704

In optimization problems, the expensive black-box function implies severely restricted budgets in terms of evaluation. Some Bayesian optimization methods have been proposed to solve this problem, such as expected improvement (EI) and hierarchical expected improvement (HEI). However, most of these methods are nonparallel and use a one-point-at-a-time strategy. This study proposes a new parallel framework that uses the minimum energy criterion. This framework can reduce time cost by reducing the number of iterations and avoiding the local optimization trap by encouraging the exploration of the optimization space. We also propose a shrink-augment strategy to correct the local surrogate model for the black-box function by placing more points around the true optima, which could also benefit the optimization. Some numerical experiments are also presented to compare the new method with popular existing methods. The results show the superiority of our proposed method over other Bayesian methods due to delivering better results with fewer iterations.