Workshop Series on Uncertainty in Computer Models
University of Sheffield, October 2014 - January 2015
Do you use computer modelling in your research? Are you concerned about the accuracy of your model predictions? In most applications of computer modelling, we can expect to have uncertainty about the choice of model input parameters, and more general uncertainty about how well the model represents the physical, biological or social system. How can we quantify uncertainty in a model prediction, and can we identify the most effective way to reduce it?
If you are interested in these issues then you are very welcome to come along to a series of four workshops that will be exploring methods for quantifying and managing uncertainty in computer models. The purpose is to bring together researchers with methodological interests in uncertainty quantification, and computer model users. Model users will have the opportunity both to learn about existing uncertainty quantification methods, and to shape the agenda for future research that will best meet their needs.
The sessions will be open to all researchers who use computer models regardless of which discipline they are in or which faculty they are hosted by. A basic knowledge of probability and statistics will be assumed. The workshops are intended to be mostly self-contained, so participants do not have to sign up for all four of the workshops and are welcome to get involved in just one or a couple of those sessions they think will be of interest. Regardless of career stage, sessions are open to all. Each workshop will consist of a presentation, with participant discussion encouraged throughout.For further reading, see this discussion in the MUCM Toolkit.
Workshop 1. An introduction to uncertainty quantification for computer models: problems and methods
22 October 2014, 14:00-16:00, Council Room in Firth Court
The aim of this workshop is to give an overview of the field, including a short introduction to each topic that will be covered in the following workshops.
Workshop 2. Emulator methods for working with computationally expensive models
19 November 2014, 14:00-16:00, BMS Conference room
We consider the problem of working with a computer model that takes a long time to run, to the extent that we cannot evaluate the model output at all the different input values we wish to try. In the emulator method, we use a relatively small number of model runs, and then interpolate to fill in the gaps. A key feature of the emulator method is that we also quantify uncertainty in the interpolation, so that we are able to assess how the computer model analysis might change if we could do further model runs..
Workshop 3. Sensitivity analysis for investigating the effects of model input uncertainty
17 December 2014, 14:00-16:00, Information Commons Collaboratory Space
Typically, a computer model will have many uncertain input parameters. However, it is rarely the case that all uncertain inputs are equally 'important'; some inputs may have very little effect on the output uncertainty. We discuss how to measure the importance of an uncertain input, and computational tools for doing the measurements. These analyses can help us to decide which uncertain inputs to learn more about, to best reduce output uncertainty.
Workshop 4. Calibration, inverse problems, and the problem of imperfect models
21 January 2015, 14:00-16:00, BMS Conference room
In the calibration or inverse problem, we have observations related to the computer model outputs, and we wish to find computer model input values to that the model outputs fit the observed data. We discuss tools for doing this, as well as the problem of calibrating an imperfect model and how to avoid overconfident model predictions.