How easy is it to do?
This is about balancing simplicity and accuracy of the reduced model.
Conservation analysis is an especially important parametric method as it
does not incur any error and therefore retains all properties of the
original model. It is often therefore recommended as a first step for
reduction of a large model. However, finding conserved moieties is not
straightforward for larger models and it seldom reduces a model by more
than 10-15% [33]. Most methods are automatable, meaning that user
input is not required for the algorithm to run. This makes complicated
mathematical methods, such as linearisation and lumping of nonlinear
models [18], more user friendly and accessible. The simplest method
that retains acceptable accuracy is using an Empirical Approximation
however it has limitations regarding generality.
Most of the model-order reduction methods require the use of a
specialised software, except for the Empirical Approximation method.
Many of the algorithms are available as pre-coded toolboxes in MATLAB.
For example, the Global Optimisation Toolbox for simulated annealing,
Deep Learning Toolbox for ANNs, and the SAFE Toolbox for global
sensitivity analysis [53].
In all parametric approaches there is a trade-off between model
simplicity (model-order) and its predictive performance. Hasegawa et al.
[54] proposed a composite criterion-based approach that couples the
two opposing factors together to find the optimal reduced model. For all
methods (except conservation analysis) it will be necessary for the user
to define the trade-off.