Title: Sequential experiment design for inverse problem from complex dynamic computer codes
Authors: Devon Lin - Queen's University (Canada) [presenting]
Abstract: The inverse problem in computer experiments refers to finding an optimal design input that a field observation or a prespecified target as closely as possible. We consider the inverse problem of dynamic computer experiments in which both the field observation and the computer simulator outputs are time series valued. We take a sequential design approach by first fitting a statistical emulator on an initial design set, selecting the follow-up design points using a novel expected improvement criterion, and updating the emulator. The optimal design input is extracted by the proposed least expected square discrepancy method. The proposed approach is shown to be effective in an illustrative example. Moreover, it achieves better accuracy in extracting the optimal design input compared with existing alternatives in the simulation study and a real application.