Adaptive sampling with application in environmental studies and computer experiments

Adaptive sampling, which select samples sequentially, is known to be more efficient than traditional non-adaptive sampling and fixed design procedures. However, most of the methods are developed based on relatively small and well-defined regions. These assumptions are often violated in environmental studies we faced today because they invariably involve populations distributed over a large space with irregular sampling frame. A new sampling plan is proposed which enhances the estimation efficiency by taking into account the shape of the sampling region and incorporating a novel adaptive procedure. Unbiased estimators, an optimal sampling criterion, and a heuristic search algorithm is introduced. Applications to real examples are presented, which show remarkable improvement in estimation efficiency using the proposed plan over existing methods. Unlike environmental studies, design of computer experiments has been widely investigated, however, most of the designs are chosen in advance without utilizing any information from the response, which results in insufficient information. We introduce a new class of sequential designs for computer experiments. It is model-free and constructed based on space-filling designs. The construction procedure, design-unbiased estimators, and some improvements using Rao-Blackwellization are proposed. More importantly, we introduce a refinement that provides better control over sample size and avoids replicates in the final sample. We demonstrate this new class of sequential designs are sampling-wise efficient by a simulation study and a IBM data center thermal management example.Ph.D.Includes bibliographical referencesby Huijuan L