Feature requests and suggestions

RE: Inputting deviation vector for MG parameters in control file

2021-03-25T16:37:15Z

Thanks everyone for the thorough responses!

Peter

On Thu, Mar 25, 2021 at 8:56 AM Richard Methot <VLab.Notifications@noaa.gov> wrote:

One difference between the two approaches is that the deviation vector approach has the devs penalized by the dev se and there is option to use a rho also. For the block approach, the dev se would have to be used as a prior on each of the block parameters.

--
Richard Methot Stock Synthesis Virtual Lab Forum https://vlab.noaa.gov/web/stock-synthesis/public-forums/-/message_boards/view_message/14224700 VLab.Notifications@noaa.gov

RE: Inputting deviation vector for MG parameters in control file

2021-03-25T15:56:41Z

One difference between the two approaches is that the deviation vector approach has the devs penalized by the dev se and there is option to use a rho also. For the block approach, the dev se would have to be used as a prior on each of the block parameters.

RE: Inputting deviation vector for MG parameters in control file

2021-03-25T15:47:05Z

I use the approach Ian describes above, and can verify that it produces identical results to the deviation approach, with the added benefit that you can have 'holes' in the vector where you turn off uninformed deviations if there are no data for that year.

RE: Inputting deviation vector for MG parameters in control file

2021-03-24T23:52:17Z

I think Kathryn is right that there's no option for parameter devs (other than recdevs which are special in other ways too) to be input in the control file.

The .par file approach she proposes should work.

The other option that I can think of would be to define a block pattern with 1-year blocks spanning the 1982 to 2019 period you're looking at. That would then require new parameter lines for each period of the block, but you should be able to get identical results to the use of the devs (as long as the block functional form was chosen to match whatever you were using for the devs). This approach is clunkier but I think should give you all the control you would want, including the option for the parameters to replace the base parameter, rather than add to it or multiply by it.

RE: Inputting deviation vector for MG parameters in control file

2021-03-24T23:35:12Z

Hi Peter,

Thanks for the question! If your goal is to put in a vector of fixed parameter deviations, there isn't a way to do this in the control file to my knowledge. You can only have SS estimate deviations by setting up the parameters related to the parameter deviations in the desired manner in the control file.

However, the hack that I think would work would be to 1) run SS with all parameter values related to the devs not estimated (I think devs will stay at their starting values if this is the case), 2) input the parameter deviations into the ss.par file, 3)change the starter file to read from the par file, then 4) run SS again and see if this worked by seeing if the devs stayed the same or not. (Hopefully a more experience user will correct me if I missed a step here!)

Also, there may be some sort of alternative growth setup that accomplishes what you want to do, although I'm not sure...

Thanks,

Kathryn

Inputting deviation vector for MG parameters in control file

2021-03-24T23:18:56Z

Hi SS team,

What is the format for inputting a deviation vector for time-varying MG parameters (e.g. cohort growth) in the control file? The manual says to input them after the TG parameters but I think SS is reading them as variance adjustment parameters.

Say I've estimated cohort growth deviations in a model, and the control.ss_new file has commented descriptions of the deviation vectors.

From control.ss_new

0 # use 2D_AR1 selectivity(0/1)
#_no 2D_AR1 selex offset used
#
# Tag loss and Tag reporting parameters go next
0 # TG_custom: 0=no read and autogen if tag data exist; 1=read
#_Cond -6 6 1 1 2 0.01 -4 0 0 0 0 0 0 0 #_placeholder if no parameters
#
# deviation vectors for timevary parameters
# base base first block block env env dev dev dev dev dev
# type index parm trend pattern link var vectr link _mnyr mxyr phase dev_vector
# 1 3 1 0 0 0 0 1 1 1982 2019 4 -4.74739 -2.67466 -0.140121 0.472827 0.0923778 -2.56083 -1.85329 1.09853 -4.85296 1.68831 -0.603715 0.0574974 -1.2095 4.93564 -0.830913 0.15358 -0.363772 -1.36393 1.95603 -2.23447 4.07689 -4.96396 1.66844 0.929938 9.78155 4.38053 3.69499 -4.44708 8.47207 -1.20438 -3.44645 -4.31041 1.96447 -1.19382 0.147736 -2.70018 2.98459 3.20983
# 1 13 3 0 0 0 0 2 1 1982 2019 5 1.1639 -2.49719 1.95786 -2.18513 -0.252127 1.65013 -2.24579 -1.24304 -0.513772 0.0190251 -0.334822 0.969157 -1.57207 1.95141 -2.44934 2.88252 0.371885 0.574171 -0.637478 1.05763 0.789094 2.9881 -0.50236 3.67068 0.0724201 2.11733 0.241812 0.58569 -0.121206 -1.3105 1.72859 -0.0154649 -0.103282 -0.268945 -2.25246 -0.236905 1.78808 0.278959

I've tried a couple things, one of which was to try to uncomment the deviation vector lines above. I think SS is reading the lines I've tried to input as variance adjustment values (see below).

From echoinput.sso:

0 TG_custom (need to read even if no tag data ); tag_data?: 0 N_Fleet: 3
dev vector #: 1 setup: 1 3 1 0 0 0 0 1 1 1982 2019 4 3 0 phase: 4
dev vector #: 2 setup: 1 13 3 0 0 0 0 2 1 1982 2019 5 13 0 phase: 5
read var_adjust list until -9999
1 1 3
1 0 0
0 0 1
1 1982 2019
4 1.11173 0.0944435
0.135578 0.389412 0.226817
0.856733 0.103363 -0.359691
-1.48202 -0.278921 -0.366551
-0.570168 -0.22367 -1.25303
-0.931673 -0.157324 -1.99008
-0.269609 1.938 3.52326
3.08203 1.39447 2.41065
-0.7346 0.145877 -1.96864
-3.76639 -3.69407 -1.6164
0.0672479 -0.755547 -0.18352
0.556263 0.274177 0.996706
2.06624 0.492732 0.173065
-9999 1 0
number of variance adjustment records = 17

Thanks in advance for any suggestions

Peter Kuriyama

Dan,RE: recruitment forecast

2021-03-24T17:04:59Z

Dan,

That is the intention of the model code, but we have only been able to try it out in a limited number of model configurations, Please let us know if the model does not work as intended in your application. The code is designed so that recr_dist during the forecast is either a time-varying parameter, or it is set to the mean of a range of prior years. We'll get a beta of this updated exe posted soon. Note that SS3 changes often will require an update of your version of r4ss also.

RE: recruitment forecast

2021-03-24T16:54:04Z

Thanks Rick for the clear and comprehensive response. If I could just confirm that from the perspective of 3.30.17, if the area recruitment distribution is defined as a time vary parameter (e.g. simple annual deviations), we could simply use option 3 in the forecast file together with the year range to determine the average value to go into projection without invoking the additional time-block configuration. Is that correct?

Cheers,

Dan

RE: recruitment forecast

2021-03-24T16:33:29Z

Thanks Dan for the good questions and thanks Kelli for a good response. Let me recap here, from the perspective of the current code which we intend to release soon as 3.30.17.

The first line, for benchmark (aka reference points), does not consider recruitment deviations because these are equilibrium quantities. The quantities beg_recr_dist, end_recr_dist, on that line refer to the distribution of recruitment among biological entities, seasons, and areas.

The second line for forecast has seen some development recently, as well as some additional warnings to help users avoid incompatible combinations. Here the concepts of recruitment deviations and recruitment_distribution both apply. There has been augmentation of the controls for both of these.

This new control, shown below from forecast.ss, controls the mean recruitment in the forecast. Older versions of SS3 only had option 0:

1 #_Forecast recruitment: 0= spawn_recr; 1=value*spawn_recr_fxn; 2=value*VirginRecr; 3=recent mean from yr range above (need to set phase to -1 in control to get constant recruitment in MCMC).

With this new option 3, the year range comes from the 2nd line that you originally referred to using the values for: beg_mean recruits, end_recruits

The annual deviations around that mean remain under control of the control file input for the phase of the forecast recruitment deviations.

Recently we realized that the control of the distribution of each forecast recruitment needed to align with the control of the mean forecast recruitment.

So now if you select option 3 to get mean forecast recruitment from the mean of a range of years of historical recruitment, SS3 will also get the mean distribution of those recruitments among bio entities and areas from the same range of years. Here is where we have added some more warnings if incompatible conditions are detected.

I hope this helps. Keep the questions coming. Kelli already included a link into our github repo. Check it out.

Rick

RE: recruitment forecast

2021-03-24T12:53:52Z

Dan,

Unfortunately, I am not very well versed in multi-area models. It looks as though the forecast recruitment is unique to growth pattern, see here for the code.

I know that is not very helpful, but hopefully someone who is better at area-specific models will be able to answer your question with more detail.

Kind regards,

Kelli

RE: recruitment forecast

2021-03-24T12:21:17Z

Thanks Kelly. That is very helpfully.

So both lines refers to the year range for annual recruitment deviates - one for reference points calculation, the other for projection.

If I could clarify a bit more - Does this also apply to the regional recruitment distribution parameter which are defined as annual deviates in a multi-area model (not the annual recruitment deviates parameter)?

Cheers,

Dan

RE: recruitment forecast

2021-03-24T12:01:38Z

Dan,

Thank you for the well-formed question. The benchmark (Bmark) years are used for calculating reference points and the forecast (Fcast) years are used for the projection. Thus, if you want to use the new feature of supplying a range of years to define recruitment deviations to be used in creating an average forecast recruitment environment then you should use the Fcast lines. You can read about this in the manual on pages 29-30, available here https://github.com/nmfs-stock-synthesis/ss-documentation/blob/main/SS330_User_Manual.pdf, note that you must use option 3 to make the Fcast years translate into mean recruitment. Hopefully, others will chime in later if I have missed anything.

Kind regards,

Kelli

recruitment forecast

2021-03-24T10:23:38Z

Hi,

May I seek some clarification on the two lines relating to recruitment in the forecast files (highlighted):

#_Bmark_years: beg_bio, end_bio, beg_selex, end_selex, beg_relF, end_relF, beg_recr_dist, end_recr_dist, beg_SRparm, end_SRparm (enter actual year, or values of 0 or -integer to be rel. endyr)
296 296 289 296 289 296 13 296 13 296

and

#_Fcast_years: beg_selex, end_selex, beg_relF, end_relF, beg_mean recruits, end_recruits (enter actual year, or values of 0 or -integer to be rel. endyr)
296 296 296 296 248 288

If I understand correctly, the "beg_mean recruits, end_recruits" in the 2nd lines define the year range for average recruit to be used in projection.

What about "beg_recr_dist, end_recr_dis" in the first line? Do they define the year range for regional/seasonal recruitment distribution to be used in the projection? However, if the regional/seasonal recruitment distribution has been specified as time varying parameters, SS3 will use the initial mean in the projection any way, then the defined year range won't be relevant, is that correct?

I understand the beta version of SS3.16 allows flexibility in specifying year range of regional recruitment distribution for projection. Should it be specified through 1st or 2nd line?

Sorry if this sounds a bit messy.

Recruitment Allocation

2020-01-21T18:13:42Z

Add an area Recruitment Allocation parameter for Early recr dev with time varying functionality

Add an area Recruitment Allocation parameter for Rinit