Statistics Seminars: Comparing multiple simulators using Intermediate Variable Emulation
11 February 2013 14:00 in CM221
Complex systems are often modelled by several different simulators, each with its own strengths and weaknesses. Comparing these simulators as functions over their input spaces is very difficult, as their input
variables cannot in general be linked.
By creating a set of `intermediate variables', representing the sub-processes within these simulators, we develop a framework using Bayesian emulation that enables the simulators to be better understood in terms of the processes they model. This method is then extended to multiple simulators, using common sets of intermediate variables. The input spaces can be more thoroughly linked, and the treatment of the sub-processes within each simulator compared. The intermediate variables enable a direct comparison of the two simulators, using emulators from this space to the output space.
We demonstrate this method using two ocean carbon cycle simulators. We are able to refine both input spaces, link several input parameters and discover which sub-processes are treated similarly and which differently by the two simulators.
Contact email@example.com for more information