Dynamic selection of models

From Semantic Portal Wiki

{{#vardefine:category|Publication}}{{#vardefine:templatename|i.publication}}{{#vardefine:package|smwbp_instance_templates}}

Edit

Reference: {{#vardefine:pagename|dynamic selection of models }}

1. [[]]

bibtex

{{#vardefine:pagename|Dynamic selection of models }}{{#vardefine:key| }}

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 ofdecision analysis, statistics, artificial intelligence, andsimulation. I develop and implement a method to solve a time-criticaldecision problem in the domain of critical-care medicine. This methodselects models that balance the prediction accuracy and the need forrapid action. Under a computation-time constraint, the optimal modelfor a model-based control application is the model that maximizes thetradeoff of model benefit (a measure of how accurately the modelpredicts the effects of alternative control settings) and model cost(a measure of the length of the model-induced computation delay). Thisdissertation describes a real-time algorithm that selects, from agraph of models (GoM), a model that is accurate and that is computablewithin a time constraint. The dynamic-selection-of-models (DSM)algorithm is a metalevel reasoning strategy that relies on a DSMmetric to guide the search through a GoM that is organized accordingto the simplifying assumptions of the models. The DSM metric balancesan estimate of the probability that a model will achieve the requiredprediction accuracy and the cost of the expected model-inducedcomputation delay. The DSM algorithm provides an approach to automatedreasoning about complex systems that applies at any level ofcomputation-resource or computation-time constraint.The DSM algorithm solves the model-selection problem for aventilator-management advisor (VMA). A VMA is a computer-based monitorfor patients in the intensive-care unit (ICU); VMAs applypatient-specific prediction models of physiology to interpret ICU dataand to predict the effects of alternative proposedtreatments. VentPlan is a prototype VMA that implements a simplifiedmodel of physiology to monitor postoperative ICU patients; this modelis unable to make accurate predictions for patients with complexphysiologic abnormalities, such as the abnormalities that occur inasthma or pulmonary embolus. I describe the VentSim model, a moredetailed model of cardiopulmonary physiology that makes accuratepredictions for patients with a wide range of physiologicabnormalities. Although VentSim is too computationally complex for useat the inner loop of a real-time VMA, alternative simplifications ofVentSim represent a range of tradeoffs of prediction accuracy andcomputation complexity.I implement the DSM algorithm in Konan, a program that selectspatient-specific models from a GoM of alternative simplifications ofthe VentSim model. Konan demonstrates that the DSM algorithm selectsmodels that balance the competing requirements for high predictionaccuracy and for low computation complexity; these model selectionsallow a VMA to make real-time decisions for the control settings of amechanical ventilator.

download:

• paper:
• slides: