Evaluating Stochastic Computer Experiments using High-Performance Computing Systems, by Dominik Dahlem of Trinity College

Large-scale simulation studies are necessary where mathematical
closed-form solutions do not exist. This entails a statistical analysis
of the random quantities to meaningfully answer questions about optimality in system responses or dynamic behaviour of the involved
entities. Yet, limited computing resources often force researchers to
take short-cuts to obtain results. This talk will present two parts to
address the tasks of simulating and analysing stochastic computer
simulations. First, a practical approach to designing computer
simulations for high-performance computing clusters is
presented. Second, a statistical regression methodology (stochastic
Kriging) is introduced to analyse the global response of a system over a
range of tunable simulation parameters. Kriging metamodelling is a
spatial interpolation technique which permits the prediction of any
random quantity that is not yet known exactly within a specified design
domain. The uncertainty of these predictions can be assessed and
simulation studies can be guided towards those locations with the
highest uncertainty. The iterative nature of kriging can be exploited to
reduce the computational burden significantly while at the same time
delivering a global picture of the system measurements with respect to
the simulation parameters.


Thursday, April 8, 2010, 17:30