Implementation of Markov Chain Monte Carlo techniques in parameter estimation for engineering models

A new approach for parameter estimation that gives both accurate point estimates and uncertainty information is explored in this research. This very promising approach is based on an implementation of Markov Chain Monte Carlo (MCMC) methods on Bayesian posterior distributions, which contain full information about uncertainty of the parameter estimates.; The first part of this thesis demonstrates implementations of the MCMC methods in various complex chemical engineering applications including Multiresponse and Error-in-Variables models (EVM). Benefits of the methods are clearly identified and compared with those obtained by classical techniques. Not only familiarity with the MCMC methods can be gained, but also parameter estimation with full uncertainty assessments are obtained for various engineering problems. In particular, this research is the first successful attempt at estimating the parameters with full uncertainty information for a reversible copolymer system using the Kruger model.; The second part of this research investigates improvements to the methods performance and to modifying the methods to suit engineering requirements. As discovered in this research, efficiency of the methods can by improved by combinations of optimal tuning, algorithm selection, and blocking. It was also found that further efficiency improvement is obtained by utilizing parallel computing machines, in that special combination and augmentation techniques for parallel chains were developed. Moreover, two new engineering-specific convergence diagnostics have been successfully developed based on the precision of the estimates and smoothness detection of marginal distributions and Joint Confidence Region (JCR) contours.; The third and most important part in this research is to develop implementation guidelines of MCMC methods for engineers. As mentioned, these methods have excellent benefits over the classical techniques, but implementation procedures have not been very well described, especially for engineering applications. Regardless of great potential of the methods, significant experience, skill and training is needed in order to bring out the full capability of MCMC methods. Our primary goal is to develop and describe implementation guidelines for inexperienced users, so that they can apply the methods without lengthy and expensive training. These guidelines were devised based on the research findings, which are mainly geared toward chemical engineering applications. However, inexperienced users in other fields will also find the guidelines useful. (Abstract shortened by UMI.)...