Computer experiments are becoming popular for their ability to simulate complex physical systems at relatively cheap cost. Since these computer experiments may run very time consuming codes, it is difficult to get many observations and hence statistical models are used to estimate the responses from such codes at untried input points. In the calibration problem in computer experiments, observations from both the computer experiment and the physical experiment are used to find the best setting of the calibration parameters (unknown physical constants in the physical process or inputs to the computer experiment which do not have a counterpart in physical process being studied) so that outputs from the computer experiment are in agreement with the output from the physical experiment.
This thesis discusses a new method to solve the calibration problem. This method coupled with new sequential design strategies for calibration proposed in this thesis is shown to be effective in reducing the number of observations required for the estimates of the calibration parameters to converge to the true value. Sequential designs are proposed for the case when additional observations from only the computer experiment can be observed and the case when additional observations from both physical and computer experiments can be observed.
This thesis also discusses a simple method to tune a computer model so that output from the computer experiment are in agreement with the output from the physical system output when the physical system changes slowly with time. Guidelines are provided for when these methods might be useful.