Principal Investigator
|
Project Title
| Sensitivity Analysis of Complex Simulation Models for Environmental Management |
|---|
Brief Description for General Publications
Integrated simulation models are widely used to help guide environmental-management decisions. For example, in the Namoi catchment in Northern NSW, irrigation policy is guided by a model which combines hydrology (to forecast river flows in given climate and irrigation scenarios), land-use and cropping, and economics (to assess farming incomes and other economic consequences). Another example is the modelling of salinity in the lower Murray to guide choice of flow targets in the Murray-Darling Basin. There is a pressing problem to assess the uncertainty in the predictions from such models. Uncertainty arises in the relations modelled, in the estimated parameter values of the model, in the inputs applied and in the boundary conditions (e.g. groundwater flows), and from the aggregation over space and time necessary in constructing a usable model. Sensitivity assessment is the first stage in assessing uncertainty. It determines how strongly the model outputs are influenced by variation in all (in principle) combinations of uncertain items. Two traditional approaches are (i) to change one or two uncertain items (say parameter values) at a time and see the effects on the outputs, then express them as derivatives. Clearly this is limited and computationally expensive; (ii) Monte Carlo trials, applying sample changes to all uncertain items and drawing conclusions about the relation between those items and the outputs. This may be informative about which combinations have the largest and smallest influence, but there is a risk that too few trials can be afforded to reveal everything that matters, or that the results are too complicated to interpret usefully. The new approach proposed is to list a collection of requirements about the outputs and find over what set of values of the uncertain items those requirements are met. At the simplest, we ask over what ranges of the items the output variables remain in given ranges, but this problem formulation allows much more flexibility than that. You can think of the process as mapping a set in output space to a feasible set in the space of given uncertain items. The crucial problem is then to characterise the feasible set, which is usually very complicated, so as to pick out features which are informative for particular purposes. For example, we might want to know directions in which the feasible set of some model parameters is large (by some norm) so as to identify parameter combinations which have little effect and can be deleted, simplifying the model. In the project, we have first to decide what features are useful, by consulting with modellers and users of model predictions, then devise ways to find and evaluate them. We shall have to exploit existing techniques for set-to-set mapping (which involve approximation and some insight into dynamics), but we shall also have to develop search and approximation methods. To test the new approach, a comparison with the traditional sensitivity analysis approaches is necessary. As noted, this requires considerable computational resources, and would not be possible without the resources of the National Facility. |