Multivariate metamodelling: a new way to understand your mathematical model and how it relates to measurements (Video) 
(Seminars)
7 February 2012
4 - 5pm
Venue: Ground Floor Seminar Room (G010), UniServices House, 70 Symonds Street, Auckland
A Bioengineering special Seminar by Professor Harald Martens, Centre for Integrative Genetics (CIGENE), Norwegian University of Life Sciences /NOFIMA, Aas, Norway
Mathematical models are essential to modern bioengineering. But high-dimensional nonlinear spatiotemporal models can be difficult to analyze theoretically, difficult to overview, difficult to compare to each other and difficult to fit to empirical data.
Multivariate metamodelling represents a generic tool that can help the modeler to understand a model’s behavioral repertoire, to assess its global sensitivity patterns, to visualize the consequences of modifying the model, as well as to simplify the fit of a model to data and to attain computational compaction.
Multivariate metamodelling of a model consists of extensive, statistically designed computer simulations to cover the relevant input combinations, followed by graphically oriented statistical approximation analysis of the input-output relationships, and results in simple, interpretable predictive models of the type Outputs=f(Inputs) or Inputs=f(Outputs).
The lecture will explain generic multivariate tools that we have developed in CIGENE and at NOFIMA in order to make the metamodelling suitable for bioengineering purposes. It will demonstrate how multivariate metamodelling methods based on subspace regression can reveal a model’s underlying “harmonies and rhythms”, and how qualitatively different models can be mapped together.
The lecture will also briefly cover chemometric techniques for removing selectivity problems in quantitative measurements in e.g. biospectroscopy (multivariate calibration), and techniques for discovery and description of totally unexpected phenomena in data (multivariate residual analysis).
Watch the high-resolution version of this video.
