Joint Research Conference
June 24-26, 2014
A Unified Framework for Uncertainty and Sensitivity Analysis of Computational Models with Many Input ParametersComputational models have found wide applications in simulating physical systems. As the physical system becomes more complicated, the number of input parameters for the model can be very large. Two major tools, uncertainty analysis (UA) and sensitivity analysis (SA), are often employed. Existing methods for UA and SA require separate designs, and most of them work only for small numbers of input parameters. They become inefficient when the number of input parameters is large. We propose a unified framework for both UA and SA. Using the same samples for Monte Carlo based UA, a screening method designed for large numbers of input parameters, which follows the Hamada-Wu strategy, is proposed for selecting significant linear, nonlinear and interaction effects. Then the sensitivity ranking of the selected significant effects is computed based on a decomposition of variance. Because the procedure requires only one design, it is economical in run size and computationally efficient. Illustration of the method is given with the building energy model of the Cherry L. Emerson building at Georgia Tech.