Aggregation of Multiple Adaptive Sampling Criteria in Computer Experiments

• Activity Number: 463
• Type: Contributed
• Date/Time: Wednesday, August 1, 2012 : 8:30 AM to 10:20 AM
• Sponsor: Section on Physical and Engineering Sciences
• Author(s): Vera Bulaevskaya*+ and Gardar Johannesson and David Domyancic
• Companies: Lawrence Livermore National Laboratory and Lawrence Livermore National Laboratory and Lawrence Livermore National Laboratory
• Keywords: computer experiments ; design of computer experiments ; adaptive sampling ; aggregation of multiple sampling criteria

Abstract

Codes involved in computer experiments are often computationally expensive, so only a small number of points in the input space is sampled with such codes, and the response in the remainder of the space is fit with a statistical emulator. It is thus essential to produce an efficient design, i.e., one that maximizes the information about the response with a small number of design points. Adaptive sampling is one way to accomplish this: given a set of input points already sampled with the code, additional points are chosen from a large candidate set according to some user-specified criterion that reflects the amount of information about the response gained by sampling a given point. There are many such criteria, and different choices of these will yield different designs, obviating the need to aggregate their individual preferences. The talk will discuss the performance of some established aggregation approaches, such as Expected Improvement and voting systems used in elections, and compare these to a learning-based method that dynamically builds a statistical model to predict a candidate point's utility from its individual criterion scores.