[ADMB Users] Are parameters scaled for optimization.

[dave fournier]

On 14-01-01 01:43 PM, Mark Maunder wrote:

for an unbounded parameter  (an init_number)
  the scalefactor is implmented OK.

238 void df1b2_init_number::set_value(const init_df1b2vector& _x,const 
int& _ii,
  239     const df1b2variable& pen)
  240 {
  241   ADUNCONST(init_df1b2vector,x)
  242   ADUNCONST(int,ii)
  243   if (scalefactor==0.0)
  244     operator = (x(ii++));
  245   else
  246     operator = (x(ii++)/scalefactor);
  247 }
  248

So this is OK.

However the code for an init_bounded_vector never wrote itself.

569 void df1b2_init_bounded_vector::set_value(const init_df1b2vector& _x,
  570   const int& ii,const df1b2variable& pen)
  571 {
  572   init_df1b2vector& x=(init_df1b2vector&) _x;
  573   if (initial_params::mc_phase)
  574   {
  575     ::set_value_mc(*this,x,ii,minb,maxb);
  576   }
  577   else
  578   {
  579     ::set_value(*this,x,ii,minb,maxb,pen);
  580   }

I was hoping that someone would take up the job after the 
init_bounded_number was done,
but there seems to be more interest in doxygen and unbelievably 
complicated script files.

If there is some interest I can extend this a bit.  One application 
would be to multivariate
random effects models.  It appears that scaling the std dev terms 
greatly improves the stability of
glmmadmb and I expect this is a fairly general result.

> I assume it is the same for standard parameters and re parameters.


I'm not sure what is meant here.


       Dave

















> Thanks Dave. What if it is unbounded?  I assume it is the same for standard parameters and re parameters.
>
> Regards,
>
> Mark
> IATTC
>
>> On Dec 31, 2013, at 1:03 PM, "dave fournier" <davef at otter-rsch.com> wrote:
>>
>>
>> For a random effects model
>> The relevant code appears to be
>>
>> if the scalefactor has not been set (i.e. =0)
>>
>> 409 df1b2variable boundp(const df1b2variable& x, double fmin, double fmax,
>> 410   const df1b2variable& _fpen)
>> 411 {
>> 412   ADUNCONST(df1b2variable,fpen)
>> 413   df1b2variable t;
>> 414   //df1b2variable y;
>> 415   //y=x;
>> 416   double diff=fmax-fmin;
>> 417   const double l4=log(4.0);
>> 418   df1b2variable ss=0.49999999999999999*sin(x*1.57079632679489661)+0.50;
>> 419   t=fmin + diff*ss;
>> 420
>>
>> So the parameter is effectively scaled by diff.
>> Better would be
>>
>> 418   df1b2variable ss=0.49999999999999999*sin(x/diff*1.57079632679489661)+0.50;
>>
>> The only difficulty would be that the function minimizer report does not deal with large numbers well.
>>
>> For parameters whose scalefactor has been set the code is
>>
>>
>> 460 df1b2variable boundp(const df1b2variable& _x, double fmin, double fmax,
>> 461   const df1b2variable& _fpen,double s)
>> 462 {
>> 463   ADUNCONST(df1b2variable,fpen)
>> 464   df1b2variable t;
>> 465   df1b2variable x=_x/s;
>> 466   //df1b2variable y;
>> 467   //y=x;
>> 468   double diff=fmax-fmin;
>> 469   const double l4=log(4.0);
>> 470
>> 471   // ss is underlying varialbe on [0,1] and t lives in [fmin,fmax]
>> 472   df1b2variable ss=0.49999999999999999*sin(x*1.57079632679489661)+0.50;
>> 473   t=fmin + diff*ss;
>> 474
>>
>> again
>>
>> 472   df1b2variable ss=0.49999999999999999*sin(x/diff*1.57079632679489661)+0.50;
>>
>> would be better.
>>
>> I wrote a note about this earlier, but it did not generate much interest.
>>
>>
>>
>>
>>
>>
>>
>>
>> _______________________________________________
>> Users mailing list
>> Users at admb-project.org
>> http://lists.admb-project.org/mailman/listinfo/users