Replies: 2 comments 2 replies
-
|
Hi @IVillalon - a quick look at the .py file shows you are setting observation weights to 1.0. What observation noise (if any) are you using? (Check the obs+noise file that pestpp-ies generates). Depending on the pestpp-ies version, noise will by default be randomly sampled from a normal distribution assuming a standard deviation as the inverse of the weight. If thats the case, too much noise would preclude the ability to improve the fit. You can also explicitly pass pestpp an ensemble of observation+noise. |
Beta Was this translation helpful? Give feedback.
-
|
Hi @rhugman Thanks a lot for the replay. The obs+noise.csv file is just a bunch of 0. with realization names from 0 to 18. So I suppose there is no noise by default and the observations are all at 0 as they should be. BTW I use the latest version of PESTPP-IES from GITHUB. |
Beta Was this translation helpful? Give feedback.
-
|
This is actually a pretty interesting lil problem. I did some digging... First, here is the result of using vanilla IES with 100 reals and 5iters (I ran a sweep over the par vals to get a colomap of the log of the phi surface for vis purposes):
Each of these options helps IES deal with increasingly nonlinear problems...I turned both on and I get this: But to your point, there is no free lunch with any of this. In some settings, especially low dimensional and fast running models, tools like pestpp-glm and pest_hp will be optimal in terms of computational demand. PS - just for fun, I changed the model function to remove the squared terms so that it is (a-2) + (b-3). If you do this, you a situation where IES excels because there is no longer a clearly defined minimum and instead there is a "trough of despair" where lots of combinations of 'a' and 'b' give a nearly equally good phi value: |
Beta Was this translation helpful? Give feedback.
-
|
Hello @jtwhite79, |
Beta Was this translation helpful? Give feedback.
Uh oh!
There was an error while loading. Please reload this page.
-
Hello everyone,
Im a geochemist and quite new to PEST. I want to use PESTPP to reverse model water-rock interaction test results and to calibrate reactive transport models. I decided to create a first simple test to check if i had installed everything right and to see if the results make sense. So my model is:
(a-2)^2+(b-3)^2=f(a,b)
the only obs i gave was 0, so the only optimal values for a and b are:
obs=0
a=2
b=3
This way, if im able to optimize this i can grow from that to more complex models.
I was able to run PESTPP-GLM to optimize this model with no issues, finding a good enough solution in a 3 to 4 iterations.
This was my Python script to use pyemu to generate the .PST file and other inputs:
https://github.com/IVillalon/PESTPP_tests/blob/main/create_pst_GLM.py
But, according to what I've read in the documentation, i should be using PESTPP-IES for complex multivariable, noisy models like inverse geochemical models. So i tried to replicate the results using PESTPP-IES to no avail. It runs several iterations with 100 runs per iteration, and it gets a little bit closer to the optimum but with a residual not far from the beginning. I've tried modifying the options for lambda_mults, num_reals, subset_size and subset_how, which made it converge a little bit better, but still far from the optimum for a and b.
This is my python script for the PESTPP-IES .PST input generation:
https://github.com/IVillalon/PESTPP_tests/blob/main/create_pst_GLM.py
Since im a bit new to PEST. I have plenty of questions, but, my first question would be, am I using PESTPP-IES wrong and this is not the kind of problem it should/could solve? Are there ways to make PESTPP-IES solve problems with unique deterministic minimums like this or it's only good for very complex, noisy and uncertain problems?
Please, any kind of insight into this would be great.
If anyone is interested, here are all the inputs for the problem: https://github.com/IVillalon/PESTPP_tests/blob/main/TEST-IES.7z
Beta Was this translation helpful? Give feedback.
All reactions