Sequential Sampling With Kernel-Based Bayesian Network Classifiers

Abstract

Complex design problems are typically decomposed into smaller design problems that are solved by domain-specific experts who must then coordinate their solutions into a satisfactory system-wide solution. In set-based collaborative design, collaborating engineers coordinate themselves by communicating multiple design alternatives at each step of the design process. Previous research has demonstrated that classifiers can be a communication medium for facilitating set-based collaborative design because of their ability to divide a design space into satisfactory and unsatisfactory regions. The proposed kernel-based Bayesian network (KBN) classifier uses a set of example designs of known acceptability, called the training set, to create a map of the satisfactory region of the design space. However, previous implementations used deterministic space-filling sampling sequences to choose the training set of designs. The shortcoming of deterministic space-filling sampling schemes is that they do not adapt to focus the samples on regions of interest to the design team (exploitation) or, alternatively, on regions in which little information is known (exploration). In this paper, we introduce the use of KBN classifiers as the basis for sequential sampling strategies that can be exploitive, exploratory, or any combination thereof.