GSA plot aesthetics:
Daniel Bending
on 25 Sep 2023
Latest activity Edit by Jeremy Huard
on 26 Sep 2023
Hi All,
I'm attempting to put a set of simbiology global sensitivity analysis plots into my thesis and I'm running into some issues with the GSA plots. Firstly, the figures are very large, it would be quite beneficial to grab a set of the plots and arrange them myself, is there any documentation on how to mess around with the '1x1 Sobol' produced by sbiosobol? Or just GSA plots in general.
The second problem is that the results appear to be relative to the most sensitive parameter in that run. Is it recommended to have a resonably sensitive 'baseline' parameter in each run? I find it difficult to compare plots when a not so sensitive parameter is being recorded as near '1' for the whole run because it's being stacked against a set of very insensitive parameters. I.e. if i have multiple sets of GSAs due to a large model, how can I easily compare results? If I could do some single run through with every parameter that would be the ideal, I imagine, but then the default plot would be half a mile off the bottom of my screen, haha! Perhaps there is a solution to the first question that might help there?
Thank you for your help,
Dan
2 Comments
Time Descending
Hi all,
Found the documentation for the plot function (... as it turns out googling 'Plot function Sobol Matlab is quite effective there..!) https://uk.mathworks.com/help/simbio/ref/simbiology.gsa.sobol.plot.html which helped in both regards, letting me specify a parameter vector in the plot function, which means I can run the GSA with all parameters.
Cheers,
Dan
that's perfect!
Indeed, you can specify the list of parameters and observables programmatically with the plot method and you can also select them manually in the Model Analyzer.
Alternatively, you can access the raw data in the results object and create your own customized plots.
I would also recommend to run the GSA with all parameters at once. The first order sensitivities represent the fraction of the response variance that can be attributed to variations of single parameters. They sum up to (1 - fraction of unexplained variance). The observed variance will depend on the set of inputs being varied. So, the same parameter might have different first-order indices in different runs when these runs include different parameters. See the sbiosobol ref page and Saltelli’s Global Sensitivity Analysis book for more details.
If the number of parameters is too large, you could perform an initial analysis with a parameter scan + scatter plot, possibly with both simulation results and local sensitivities. This way you may be able to eliminate insensitive parameters to make a GSA feasible.
Best,
Jérémy
Sign in to participate
