Reference: Rutledge, G. W. Dynamic Selection of Models. Ph.D. Thesis, Stanford University, 1995.
Abstract: In this dissertation, I develop an approach to high-stakes, model-based decision making under scarce computation resources, bringing together concepts and techniques from the disciplines of decision analysis, statistics, artificial intelligence, and simulation. I develop and implement a method to solve a time-critical decision problem in the domain of critical-care medicine. This method selects models that balance the prediction accuracy and the need for rapid action. Under a computation-time constraint, the optimal model for a model-based control application is the model that maximizes the tradeoff of model benefit (a measure of how accurately the model predicts the effects of alternative control settings) and model cost (a measure of the length of the model-induced computation delay). This dissertation describes a real-time algorithm that selects, from a graph of models (GoM), a model that is accurate and that is computable within a time constraint. The dynamic-selection-of-models (DSM) algorithm is a metalevel reasoning strategy that relies on a DSM metric to guide the search through a GoM that is organized according to the simplifying assumptions of the models. The DSM metric balances an estimate of the probability that a model will achieve the required prediction accuracy and the cost of the expected model-induced computation delay. The DSM algorithm provides an approach to automated reasoning about complex systems that applies at any level of computation-resource or computation-time constraint. The DSM algorithm solves the model-selection problem for a ventilator-management advisor (VMA). A VMA is a computer-based monitor for patients in the intensive-care unit (ICU); VMAs apply patient-specific prediction models of physiology to interpret ICU data and to predict the effects of alternative proposed treatments. VentPlan is a prototype VMA that implements a simplified model of physiology to monitor postoperative ICU patients; this model is unable to make accurate predictions for patients with complex physiologic abnormalities, such as the abnormalities that occur in asthma or pulmonary embolus. I describe the VentSim model, a more detailed model of cardiopulmonary physiology that makes accurate predictions for patients with a wide range of physiologic abnormalities. Although VentSim is too computationally complex for use at the inner loop of a real-time VMA, alternative simplifications of VentSim represent a range of tradeoffs of prediction accuracy and computation complexity. I implement the DSM algorithm in Konan, a program that selects patient-specific models from a GoM of alternative simplifications of the VentSim model. Konan demonstrates that the DSM algorithm selects models that balance the competing requirements for high prediction accuracy and for low computation complexity; these model selections allow a VMA to make real-time decisions for the control settings of a mechanical ventilator.