|orchidas : Tutorials / ESA-DA / Part 3. Finding the optimal parameter values||Site Map|
Part 3. Finding the optimal parameter values
A) How to search for the optimal parameter values?
With the LSCE CCDAS we have two approaches for retrieving the optimal posterior parameters: gradient-based methods and "global search" methods. The illustrations below show very simply the differences between the two.
Very simple case: y = ax → Try to estimate parameter "a"
In the LSCE CCDAS we mostly use the gradient-descent approach, using the L-BFGS algorithm for the gradient-descent. To calculate the gradient of the cost function we have the tangent linear (TL) of the model (which provides the model's first derivatives). However for parameters that result in a threshold response of the model (i.e. highly non-linear behavior) we use the finite difference method. In certain site-based studies we've used the "Genetic Algorithm", which is one algorithm in a suite of global search methods.
B) Comparison of Methods: example at FluxNet site with ORCHIDEE
In this study we compared the ability of both the gradient-descent method and the genetic algorithm to find the minimum of the cost function, and hence the optimal posterior parameters. The data used are "pseudo-data", meaning they are not "real" data collected in the field, but generated from the model outputs (with known parameters), and perturbed with a certain level of uncertainty. We can therefore compare the posterior parameters retrieved from the optimisations with the "true" values that were used to create the "pseudo-data" from the model simulations.
Site: Beech forest in France
Description of the test
Results of the test
Using the Section A on "finding the optimal parameter sets" and your understanding of the ORCHIDEE model, could you answer the following questions about the results presented above (Figure 6)?
|Last modified: 12/02/2018 10:29:47|
|← Part 2. Sensitivity Analysis: which parameters to choose||Part 4. Single- versus multi-site optimization →|