Data analysis, model uncertainty propagation, and decision analysis are often integral components of the research we conduct.
We have developed innovative tools for analyzing data from experiments, propagating uncertainties of model inputs to output
predictions, and decision making when faced with uncertain outcomes. Some of the computational tools we apply are Monte Carlo
simulation, Bayes Monte Carlo uncertainty analysis, first-order uncertainty methods, and conditional and regression tree analysis
(CART).
Quantifying Model Prediction Uncertainty
Monte Carlo methods
Uncertainty analysts routinely apply Monte Carlo simulation techniques to characterize uncertainty in predictions from computer
fate and transport models. In a Monte Carlo approach, rather than make a single prediction based on the "best guess" for each
input parameter, multiple predictions are made using different values for the input parameters, where each parameter value is
selected from a distribution of possible values. Monte Carlo simulation provides very good description of possible model
outcomes. We are creating tools to facilitate the use Monte Carlo methods to quantify uncertainty in predictions from a program
called "Comis" that predicts indoor air flows and pollutant transport.
First-order analysis
If model predictions are highly variable or if model execution is very time-consuming, a complete Monte Carlo uncertainty analysis
may be computationally infeasible. For example, in one of our studies, predictions for twenty-four hours of air flow and
pollutant transport in a five-floor commercial building required more than 10 minutes of real time for each realization using
a desktop personal computer. Five hundred simulations required more than three days to complete, though we could have run the
simulations in parallel on several computers to reduce the real time. For longer pollutant transport predictions, an analysis
could be very burdensome. This motivated us to explore simpler methods of estimating the uncertainty in model outcomes.
Currently, we are testing first-order uncertainty analysis methods. Using a dataset from experiments in a three-floor building,
we are comparing uncertainty estimates from Monte Carlo simulation and first-order uncertainty analysis and determining in what
conditions, if any, a simpler method is acceptable.
