A method for the dynamic selection of models under time constraints

Finding a model of a complex system that is at the right level of detail for a specific purpose is a difficult task. Under a time constraint for decision-making, we may prefer less complex models that are less accurate over more accurate models that require longer computation times. We can define the optimal model to select under a time constraint, but we cannot compute the optimal model in time to be useful. We present a heuristic method to select a model under a time constraint; our method is based on searching a set of alternative models that are organized as a graph of models (GoM). We define the application-specific level of prediction accuracy that is required for a model to be adequate, then use the probability of model adequacy as a metric during the search for a minimally complex, adequate model. We compute an approximate posterior probability of adequacy by applying a belief network to compute the prior probability of adequacy for models in the GoM, then by fitting the models under consideration to the quantitative observations. We select the first adequate model that we find, then refine the model selection by searching for the minimally complex, adequate model. We describe work in progress to implement this method to solve a model-selection problem in the domain of physiologic models of the heart and lungs.