logkappa init and bounds

[kellijohnson-NOAA]

Starting with version VAST_v1_6_0 ...

VAST/R/make_parameters.R

Line 119 in 339db31

Return = list("ln_H_input"=c(0,0), "beta1_ct"=NA, "gamma1_j"=rep(0,ncol(DataList$X_xj)), "gamma1_ctp"=array(0,dim=c(DataList$n_c,DataList$n_t,DataList$n_p)), "lambda1_k"=rep(0,ncol(DataList$Q_ik)), "L1_z"=NA, "L_omega1_z"=NA, "L_epsilon1_z"=NA, "logkappa1"=log(0.9), "Beta_mean1"=0, "logsigmaB1"=log(1), "Beta_rho1"=0, "Epsilon_rho1"=0, "eta1_vf"=NA, "Omegainput1_sf"=NA, "Epsiloninput1_sft"=NA, "beta2_ct"=NA, "gamma2_j"=rep(0,ncol(DataList$X_xj)), "gamma2_ctp"=array(0,dim=c(DataList$n_c,DataList$n_t,DataList$n_p)), "lambda2_k"=rep(0,ncol(DataList$Q_ik)), "L2_z"=NA, "L_omega2_z"=NA, "L_epsilon2_z"=NA, "logkappa2"=log(0.9), "Beta_mean2"=0, "logsigmaB2"=log(1), "Beta_rho2"=0, "Epsilon_rho2"=0, "logSigmaM"=rep(1,DataList$n_c)%o%c(log(5),log(2),log(1)), "eta2_vf"=NA, "Omegainput2_sf"=NA, "Epsiloninput2_sft"=NA )

I see that the initial value for logkappa is log(0.9) where before it was 0.0.

I am trying to fit models to data from the US West Coast within the US EEZ and I am getting the following error

Check bounds for the following parameters:
       Param starting_value     Lower        MLE      Upper final_gradient
42 logkappa2     -0.1053605 -6.061506 -0.8317317 -0.8317317     -0.1577178

Can you please point me to documentation on why the bounds are being set in this way and why the init value is log(0.9) for recent versions? The model fits if I enter the function and increase the upper bound but I am not sure if this is something that I actually should be doing and I wanted to ask if we should think about the init value being outside of the default bounds because of the distances that are being calculated. Thank you for any guidance that you can provide.

[James-Thorson-NOAA]

yeah, logkappa is translated into Range ( = distance with 13% correlations for an isotropic model). If Range is greater than the maximum distance among knots, or less than minimum distance, then the model typically loses the gradient for logkappa. I therefore use those bounds and back-translate to logkappa.

kappa is technically the decorrelation rate (decrease in correlation per unit distance). So logkappa hitting upper bound is equivalent to a fast decorrelation. This can sometimes be solved by increasing the number of knots, so that knots are closer together, which automatically results in a higher upper bound.

happy to keep chatting on this thread and definitely curious to hear what you conclude! Feel free to email me directly if that's easier.

[kellijohnson-NOAA]

Thanks Jim, this was exactly what I needed. Changing the number of knots led to the bounds changing and the warning went away. I often start the number of knots at a small number to ensure that things are "working" prior to running a full model. Though, when I changed the upper bound here to Inf, the model did not converge. IMO it would be nice to allow users to change these bounds in a more simple way than entering the function via a browser call; not sure how easy that would be to code though.

[James-Thorson-NOAA]

hmm. you mean that you want to reference docs to include upper and lower as arguments in ?fit_model rather than digging through the `...` arguments? I'm not sure I'm following about the function and browser call piece.

…

On Fri, Apr 9, 2021 at 12:09 PM Kelli Johnson ***@***.***> wrote: Thanks Jim, this was exactly what I needed. Changing the number of knots led to the bounds changing and the warning went away. I often start the number of knots at a small number to ensure that things are "working" prior to running a full model. Though, when I changed the upper bound here to Inf, the model did not converge. IMO it would be nice to allow users to change these bounds in a more simple way than entering the function via a browser call; not sure how easy that would be to code though. — You are receiving this because you commented. Reply to this email directly, view it on GitHub <#300 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AL62VMQIDC6R55B7LYMUN23TH5GHRANCNFSM42VOFAEA> .

-- James Thorson Program leader, Habitat and Ecological Processes Research (HEPR) Alaska Fisheries Science Center, NMFS Affiliate Faculty, University of Washington and Oregon State University The contents of this message are mine personally and do not necessarily reflect any position of NOAA.

[kellijohnson-NOAA]

I don't have a handle on how would be best or if users should even be able to alter the bounds, but I used `debugonce(make_model)` and then worked through the code until I got the bound that I wanted, changed it to a higher value, and then entered `c` to continue through the rest of the code automatically. We can close this, as I don't think it is a huge deal, I got it to work and mainly I am just happy for your comments on the parameters and logic behind the bounds so that I could understand the direction and pattern to change them to get things to work. On Fri, Apr 9, 2021 at 12:12 PM Jim Thorson ***@***.***> wrote:

…

hmm. you mean that you want to reference docs to include upper and lower as arguments in ?fit_model rather than digging through the `...` arguments? I'm not sure I'm following about the function and browser call piece. On Fri, Apr 9, 2021 at 12:09 PM Kelli Johnson ***@***.***> wrote: > Thanks Jim, this was exactly what I needed. Changing the number of knots > led to the bounds changing and the warning went away. I often start the > number of knots at a small number to ensure that things are "working" prior > to running a full model. Though, when I changed the upper bound here to > Inf, the model did not converge. IMO it would be nice to allow users to > change these bounds in a more simple way than entering the function via a > browser call; not sure how easy that would be to code though. > > — > You are receiving this because you commented. > Reply to this email directly, view it on GitHub > < #300 (comment) >, > or unsubscribe > < https://github.com/notifications/unsubscribe-auth/AL62VMQIDC6R55B7LYMUN23TH5GHRANCNFSM42VOFAEA > > . > -- James Thorson Program leader, Habitat and Ecological Processes Research (HEPR) Alaska Fisheries Science Center, NMFS Affiliate Faculty, University of Washington and Oregon State University The contents of this message are mine personally and do not necessarily reflect any position of NOAA. — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#300 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AA7LCFB7VGYH6SHDR4ZKOPLTH5GSDANCNFSM42VOFAEA> .

-- Kelli Faye Johnson, PhD Research Fish Biologist Northwest Fisheries Science Center National Marine Fisheries Service 2725 Montlake Boulevard East Seattle, Washington 98112 (206) 860-3490 ***@***.***

[Jason-Conner-NOAA]

Hi all,
In Modsquad, we have run into bounds issues with kappa as well, for some crab runs and for Pacific cod numeric abundance. Is there a way to set the initial bounds for kappa specifically in fit_model(), rather than relaxing all bounds using optimize_args? I am working through optimizing P. cod numbers for the eastern and northern Bering Sea, with the cold pool covariate, at 750 knots and ran into the upper bound for logkappa1 at -1.819147.

Thanks!

[James-Thorson-NOAA]

its not super-easy to change individual bounds using the high-level interface, without running it twice. If you don't mind running it twice, you can do:

fit_orig = fit_model( ..., build_model=FALSE )
Lower = fit$tmb_list$Lower
# Change some bounds in Lower using grep(.) etc.
fit = fit_model( ..., Lower=Lower )

does that make sense?

[Jason-Conner-NOAA]

Yeah, that's great. Thanks Jim!

[HaikunXu]

Hi Jim,

I tried what you suggested here "fit = fit_model( ..., build_model=FALSE )" to change parameter bounds. But I did not find fit$tmb_list$Lower; fit$tmb_list only includes Map, Data, Parameters, and Random. FYI, I am using version 3.8.0. Thank you in advance.

[James-Thorson-NOAA]

Please try with fit = fit_model( ..., run_model=FALSE ) instead

[Blevy2]

Hi everyone,

As a quick follow up, I was also trying the above method to change the lower bound for logkappa and I think that it needs to be entered as lower case in the fit_model call. It took some time to figure that out so I thought I would mention it here.

So that would be

fit_orig = fit_model( ..., run_model=FALSE )
Lower = fit$tmb_list$Lower
# Change some bounds in Lower using grep(.) etc.
fit = fit_model( ..., lower=Lower )

lower=Lower is the small difference

Ben