IES phi factors #255
-
Beta Was this translation helpful? Give feedback.
Replies: 5 comments 12 replies
-
Ok, think I've answered my own question - the issue is that I had left "center on base" = true, which I should have turned off anyway for my multimodal update case, per @jtwhite79's suggestion here...if that option is true, then internal reweighting is done on the basis of the base realisation. For multimodal upgrades, this is not relevant, so should not be used for internal reweighting (I would have thought anyway). It's also worth noting for the record that when "center on base" option was true, and multimodal updgrades are used, I was seeing thread errors on the multimodal upgrade process. When I turn "center on base" == false, these no longer occur. Eg errors (only dumped to console, not logged): --- starting solve for iteration: 1 --- This was what my log file looked like: Cheers. |
Beta Was this translation helpful? Give feedback.
-
The phi factors stuff by default works on the base realization residuals if you passed "center_on_base" or on the mean prior residuals, so essentially it replicates the workflow we are all used to WRT balancing the objective function. If you want to explore the effect of differing weighting strategies, you have to use a weight ensemble in combination with the multimodal upgrade - this causes each realization to use a unique weight vector (which also effects which realizations that are found in its neighborhood). This can drive each realization to very different minima... There is also an even crazier option of supplying the phi factors file as a csv file where the index is the same realization names as the noise ensemble and the columns of the "tags" for the different obs (this requires a weight ensemble also) Then each realization should get rebalanced by its residuals using the values in the phi factor file for that realizations - WAT?! I setup this up and tried on a project but it didnt work like I expected (its a lot of fresh code in there and its undocumented so user beware if you try this! haha) |
Beta Was this translation helpful? Give feedback.
-
Keen to hear more! I too am plagued by "some more outlandish phi components for a few realz" that I suspect are inhibiting my model from attaining a fit commencerate with its level of perfection. (x-axis is log(Tr), so only off by several orders of magnitude, meaning a few thousand years!) |
Beta Was this translation helpful? Give feedback.
-
setup a weight ensemble that balances each realization so that those
outlier realz get the weights they deserve (haha!)
The idea being each real will have approx the same phi, but the outliers in
the relevant reals will be weighted really low?
The multimodal solve is not meant to be used with localization tho!
Yep, been there!!!
…--
"Perfect spheres are pointless."
|
Beta Was this translation helpful? Give feedback.
-
I have not tried it yet. I need to let it rattle around in my brain a bit
more.
…On Thu, Jun 15, 2023 at 6:23 PM J Dub ***@***.***> wrote:
yeah thats the idea - since each realization is being weight balanced
individually, those reals with "less than desirable" residuals for some obs
will naturally have weights that accommodate those special residuals...at
least in theory thats the idea. have you tried a weight ensemble yet?
—
Reply to this email directly, view it on GitHub
<#255 (reply in thread)>,
or unsubscribe
<https://github.com/notifications/unsubscribe-auth/ADSJXRGQDFPRGOGMG6UQNC3XLKS5NANCNFSM6AAAAAAZD6P3TA>
.
You are receiving this because you commented.Message ID:
***@***.***>
--
"Perfect spheres are pointless."
|
Beta Was this translation helpful? Give feedback.
Ok, think I've answered my own question - the issue is that I had left "center on base" = true, which I should have turned off anyway for my multimodal update case, per @jtwhite79's suggestion here...if that option is true, then internal reweighting is done on the basis of the base realisation. For multimodal upgrades, this is not relevant, so should not be used for internal reweighting (I would have thought anyway).
It's also worth noting for the record that when "center on base" option was true, and multimodal updgrades are used, I was seeing thread errors on the multimodal upgrade process. When I turn "center on base" == false, these no longer occur. Eg errors (only dumped to console, …