Efficient approximation in decision models using Monte Carlo simulation

In this dissertation, I develop an approach to solving complex decision-theoretic models using Monte Carlo simulation techniques. I take advantage of recent results in optimal stopping rules for Monte Carlo simulation to develop methods that can estimate the expected utility of decisions considered by the user of a normative system, and the expected value of information of potential tests. These methods do not require the user to know about the Monte Carlo process; the stopping rule theory is used to translate uncertainty in the estimates into terms that have meaning for the user of a system, such as the risk in acting now versus continued solution refinement.;Complexity is a problem for many real-world systems, especially in the context of medical decision making. Reducing complexity by model simplification will yield unpredictable errors in solutions. The approach taken in this dissertation provides approximate answers to decision problems with a precise bound in the error of the estimated solution in terms of utility. I have developed strategies that improve the efficiency of the simulation process by selectively focusing computational resources on distinguishing among candidates for maximum expected utility actions.;I evaluate the methods using a very large medical belief network, the Quick Medical Reference-Decision Theoretic (scQMR- scDT) model. This model is one of the largest probabilistic models available. I describe a simple utility model for a medical triage decisions that I have built for the scQMR- scDT network. I use the resulting decision model to evaluate the performance and scalability of the algorithms described in this work. The extreme size and complexity of the scQMR- scDT model required the further development of adaptive strategies to better guide the expected value of computation decision making process. I demonstrate that the methods introduced in this work allows approximate solutions for a very large decision model in cases are not amenable to exact computation by any known algorithms.