You are here

mxTryHard attempts in umx

7 posts / 0 new
Last post
mirusem's picture
Offline
Joined: 11/19/2018 - 19:32
mxTryHard attempts in umx

Is there a way to modify/control the amount of tries in the tryHard option for umx for any given function being called?

I appreciate it!

tbates's picture
Offline
Joined: 07/31/2009 - 14:25
mxTryHard for more control

Hi mirusem,
tryHard = "yes" in the umx functions allows different forms of tryhard, but doesn't expose the guts of each of the mxTryHard family of functions: To get down into the nitty gritty, call model = mxTryHard*(model)

There are just too many parameters and umx is designed to make using this easy for 90% of cases, with an easy fallback to more complex options like extraTries, greenOK, scale = 0.25, initialGradientStepSize etc. etc.

mirusem's picture
Offline
Joined: 11/19/2018 - 19:32
Thanks!

I see, interesting. That does make sense. I am glad you are free to call directly OpenMx functions in the case of umx (right)? That's kind of cool actually.

So that would be just mxTryHard(model) no * (I get an error with the star)? Unless by * you mean whichever mxTryHard type we want to do. I think it works with respect to that, but I guess starting values is an issue for these things (since it ultimately leads to an error with one of the models for certain things I am looking into)--so for those cases in the interim I am going with Hermine's other script which is strictly OpenMx since I am still not yet super familiar as to how to alter the starting values in umx (without yet looking into it more deeply, though I figure umxSetParameters() with some specification is the trick).

Thanks again!

tbates's picture
Offline
Joined: 07/31/2009 - 14:25
start values, tryHard = "ordinal", mxAutoStart

hi - yes, '*' meant "choose from options"

The joy of tryHard, esp. tryHard = "ordinal" (which calls mxTryHardOrdinal is that it will explore new starts in a way that would be tiresome to emulate as a human.

PS: umxRAM and the umx twin models are good are picking viable start values. For occasions that doesn't work, you might also consider mxAutoStart That runs a WLS version to find start values that are in the ball park.

mirusem's picture
Offline
Joined: 11/19/2018 - 19:32
Sorry for the late response

Sorry for the late response (been quite hectic). That makes sense. I was wondering, if you have the case in which it doesn't find a solution but it does have valid attempts, would it be acceptable to go with that 'valid attempt' solution? My feeling is that I don't necessarily want to go with a forced converged solution if that is too forced? At least, that's my impression, but maybe it is better to just figure out how to brute-force converge the result?

tbates's picture
Offline
Joined: 07/31/2009 - 14:25
solutions

Forced isn't the right word for any solutions tryHard (or openmx in general) finds. And not sure what no solution but valid attempts means. It might be you have a wrongly specified model or a small dataset where widely differing solutions have similar fit.

Typically I would m1 = mxRun(m1) the model again from the last solution to confirm that the solution is reliable.

OpenMx has flexible criteria for giving up, and criteria for accepting a solution as not improvable.

You might want to explore altering (reducing) the value of mvnRelEps. e.g.

umx_set_optimization_options("mvnRelEps", .001)

mirusem's picture
Offline
Joined: 11/19/2018 - 19:32
So by no solution I mean when

So by no solution I mean when I get (for example) this result that is being returned:
"Retry limit reached; Best fit=-353.49615 (started at -352.57463) (11 attempt(s): 2 valid, 9 errors)" etc. as posted elsewhere, where it actually does return a set of p-values, etc. but it doesn't say "Solution reached!" or whatever is clearly the case for no errors. Does that change anything you might conclude? I am doing it on quite a few phenotypes, and ultimately, I would say that this only ends up being the case for a very small fraction of them.
Of that fraction subset, mxTryHard resolves all but even more a marginal few, which at that point I would have to really alter the starting values in a random way (which feels forced)--but ultimately even if those don't get the 'solution reached' but rather 'best fit=' there is some value returned, and am thinking since it's so marginal, to just note as such.

It's specifically with the scripts online, so I think it just has to do with those phenotypes, etc.

I might go ahead and try that optimization option, in any case, though!