Statistical programming, matrix languages, and more

termination criteria

Accepted Solution Solved
Reply
Contributor
Posts: 25
Accepted Solution

termination criteria

Hello,

 

I am doing a quasi newton optimization method but I have a problem. The optimization stop before I find the global maximum.

 

I wonder if I could use a termination criterion of 1e-4 in order to stop the optimization after getting an objective function as small as I want and if it is possible, how to do it?

 

Here is my code

proc iml;

         

     use a;

     read all var _NUM_ into data;

     close a;

 

  

     reset storage = &lib..myfunctions;

     load module=(function1 function2);  

    

     parameters = j(3, 1, .) ;

     parameters[1] = 1.1 ;    

           

    parameters[3]=0.5;

      

 

     con = {-10  1E-9, 10 1}

     opt = {1 4} ;       

    

    

     *Quasi-Newton Method ;

     call NLPQN(rc, parameters, "function1", theta, opt, con);

    

 

quit ;

 

 

Thank you for helping


Accepted Solutions
Solution
‎11-28-2017 07:49 PM
SAS Super FREQ
Posts: 4,019

Re: termination criteria

There are multiple termination criteria that you can use. The documentation provides details. Common criteria include absolute function convergence criterion (ABSTOL, controlled by tc[3]) and absolute gradient convergence criterion (ABSGTOL, controlled by tc[6]). There are also "relative" versions of these criteria.

 

To use the termination criterion, allocate a vector of missing values (which means "use default values") and override the element that you want to control. For example,

 

tc = repeat(.,1,12);
tc[3] = 1.e-4;

call NLPQN(rc, parameters, "function1", theta, opt, con) tc=tc;

View solution in original post


All Replies
Solution
‎11-28-2017 07:49 PM
SAS Super FREQ
Posts: 4,019

Re: termination criteria

There are multiple termination criteria that you can use. The documentation provides details. Common criteria include absolute function convergence criterion (ABSTOL, controlled by tc[3]) and absolute gradient convergence criterion (ABSGTOL, controlled by tc[6]). There are also "relative" versions of these criteria.

 

To use the termination criterion, allocate a vector of missing values (which means "use default values") and override the element that you want to control. For example,

 

tc = repeat(.,1,12);
tc[3] = 1.e-4;

call NLPQN(rc, parameters, "function1", theta, opt, con) tc=tc;

☑ This topic is solved.

Need further help from the community? Please ask a new question.

Discussion stats
  • 1 reply
  • 169 views
  • 0 likes
  • 2 in conversation