University of Cambridge > > British Antarctic Survey's Natural Complexity: Data and Theory in Dialogue > Unnatural Complexity: Towards Achievable Aims for Complicated Simulation Models in Applications with Data (Weather-like) and those without (Climate-like)

Unnatural Complexity: Towards Achievable Aims for Complicated Simulation Models in Applications with Data (Weather-like) and those without (Climate-like)

Add to your list(s) Download to your calendar using vCal

  • UserLeonard A. Smith, London School of Economics, London, U.K.
  • ClockMonday 13 August 2007, 14:00-14:45
  • HouseLaw Faculty, Cambridge.

If you have a question about this talk, please contact Nick Watkins.

Complexity in Nature is seen in the behaviour of systems often viewed as having many interacting parts, while the success of scientific models is most apparent in systems which have been isolated as much as possible. Natural complexity flows from the seamless interaction of these parts as a single whole. While state-of-the-art scientific simulation models are, arguably, unnaturally complex. Rather than aiming to simplify either the system studied or the mathematical framework employed, various important bits of a natural system are modelled explicitly. These components are then bolted together, along with an assortment of parameterisations to handle those other known aspects of the system which are not “resolved explicitly”.

How can these simulation models be gainfully employed to address questions of decision-support in high-priority areas of Earth system science? And how might the dialogue between model and observation be improved to speed the advancement of both? Can we find a more coherent approach interpreting data in a field often characterised by what it is not (non-linear, non-Brownian, non- Gaussian tails, ...)? In particular, how might we develop faith that these unnaturally complex models can extrapolate beyond observed conditions to provide a “comprehensive, physically consistent, prudent projection” of future likely conditions? Does Bayesian emulation provide a decision-relevant alternative to naive realism? Do our simulation models have a well-defined mathematical target? Do we have grounds to believe that unnaturally complex models will be structurally robust?

These questions will be addressed within the framework of nonlinear dynamical systems. While modern computing power allows us much more complicated models, similar issues concerned Maxwell and Poincare. We will ask what achievable aims we might target given the mathematical nature of our models; weather-like applications where we have a historical archive of forecast-observation pairs will be contrasted with climate-like applications where by definition no such archive exists. Resource allocation (in terms of experimental design) for model improvement will be contrasted with that for decision support. Our ability to forecast, and in particular to construct decision-relevant probability forecasts from ensembles of simulations, will be examined. Probabilistic Similarity and Shadowing are suggested as primary tools to evaluate the decision-support relevance of ensemble simulations with imperfect models. Applications in weather forecasting and climate modelling are contrasted with those involving Near-Earth Objects to suggest necessary conditions for deploying probability forecasts (as such) from imperfect models in practice. This appears to be somewhat more difficult than using state-of-the-art simulation models to improve our understanding of natural complex systems.

This talk is part of the British Antarctic Survey's Natural Complexity: Data and Theory in Dialogue series.

Tell a friend about this talk:

This talk is included in these lists:

Note that ex-directory lists are not shown.


© 2006-2020, University of Cambridge. Contact Us | Help and Documentation | Privacy and Publicity