I’ve been trying to run a model that works more or less like this. I have a set of stress-strain curves, where the data for each curve is collected under a fixed strain rate and temperature. Each individual curve is supposed to fit a model like this: stress = A + B*strain^n A and B are supposed to…

A and B are supposed to vary with strain rate and temperature according to a very nonlinear relationship For each curveInd, it’s a different set of strain rates and temperatures. Is the notRealFunction a function that has to be fixed to one form? Or is there a chance to model notRealFunction non-…

[image] bbbales2: For each curveInd, it’s a different set of strain rates and temperatures. Is the notRealFunction a function that has to be fixed to one form? Or is there a chance to model notRealFunction non-parametrically? Non-parametric isn’t a real option here. The choice of notRealFunct…

I get seemingly sane results on simulated data That’s encouraging, at least. Just sidestepping the fixed model issue, probably just start with pairplots and try to nail down exactly what parameters are causing divergences. There’s a big section on doing it with Bayesplot: Visual MCMC diagnostics…

I tried mcmc_parcoord and nothing really stood out, even after normalizing the parameters to put them on the same scale. One catch is that I have about 70 parameters, mostly due to ABn_indiv and unit_ABn, so a lot of plots are not very useful, especially the pairs plots. I tried using the stan_check…

Hi, If you can get out the divergences then you could manually calculate the normalized values (e.g mean 0, sd 1 for all data) and calculate mean and sd for divergent values. See if there are values that stand out either with mean or sd. Wonder if normalizing data to 0, 1 with inverse cdf (to uni…

something that would stand out to someone as obviously flawed (edit: was gonna comment on this then didn’t then forgot to remove the quote, so non-sequitur to start) Does the LKJ prior on noise between A, B, n change the results much? I’m not sure that: A, B, n ~ multi_normal(f(strain rate of c…

[image] bbbales2: I’m not sure that: A, B, n ~ multi_normal(f(strain rate of curve, temperature of curve), Sigma); is preferable to A ~ normal(f(strain rate of curve, temperature of curve), sigma_A); B ~ normal(f(strain rate of curve, temperature of curve), sigma_B); Unless you know that th…

BTW, I checked and found that the parameters for the individual curves do fit quite well. Nice, can you just run the fits for all the curves and then make plots of posterior estimates of A and B vs. strain rate and temperature? Then maybe that’ll reveal what the issue might be with notRealFuncti…

Here are the plots I have so far: [A_fit_vs_A_calc] [B_fit_vs_B_calc] These are what I get from comparing the means of the fits for the individual curves versus what they should be according to notRealFunction(), The solid line is, of course, has a slope of one and intercept of zero. Looks like t…

Trying to do hierarchical nonlinear model in Stan; it's slow and probably wrong

Modeling

bbbales2 September 22, 2017, 11:03pm 5

I meant to mention also to make sure and try out the mcmc_parcoord stuff. It’s from this thread Concentration of divergences. It might be easier to find out what’s happening if you have a ton of parameters.

Topic		Replies	Views
Dealing with uncertain parameters that aren't purely fitting parameters Modeling techniques	45	4480	May 23, 2017
Help with hierarchically modeling parameters of a nonlinear model Modeling fitting-issues , specification	10	396	September 8, 2024
Unable to retrieve parameters from a dynamic model Modeling fitting-issues , specification , ecology	33	2162	March 11, 2020
Speeding up a hierarchical nonlinear model Modeling specification , performance	17	2097	February 28, 2018
NLME model in STAN for missing + unbalanced longitudinal data Modeling rstan , mixed-model , missing-data	20	777	October 21, 2024

Trying to do hierarchical nonlinear model in Stan; it's slow and probably wrong

Related topics