Batch Bayesian Active Learning For Feasible Region Identification by Local Penalization

Identifying all designs satisfying a set of constraints is an important part of the engineering design process. With physics-based simulation codes, evaluating the constraints becomes considerable expensive. Active learning can provide an elegant approach to efficiently characterize the feasible region, i.e., the set of feasible designs. Although active learning strategies have been proposed for this task, most of them are dealing with adding just one sample per iteration as opposed to selecting multiple samples per iteration, also known as batch active learning. While this is efficient with respect to the amount of information gained per iteration, it neglects available computation resources. We propose a batch Bayesian active learning technique for feasible region identification by assuming that the constraint function is Lipschitz continuous. In addition, we extend current state-of-the-art batch methods to also handle feasible region identification. Experiments show better performance of the proposed method than the extended batch methods.