topic Re: Problem in NLPQN in SAS/IML Software and Matrix Computations

Problem in NLPQN

ayaafarag — Thu, 16 Apr 2015 20:42:55 GMT

this my code and after I run it I got this error message:

ERROR: Overflow error in NLPQN.

operation : NLPQN at line 5279 column 1

operands : *LIT1222, k, optn, , , , , , *LIT1223

*LIT1222 1 row 1 col (character, size 3)

fun

k 1 row 2 cols (numeric)

0 0

optn 1 row 11 cols (numeric)

*LIT1223 1 row 1 col (character, size 3)

con

statement : CALL at line 5279 column 1

*****************************************

where l and g are vectors

the code is :

proc iml;

start fun(K) global
(l,e,g,p);

sumf=0;

do i = 1to p;

sumf = sumf + (((e*l)+ ( k**2
* g**2 )) /(l+ k)**2);

end;

return (sumf);

finish fun;

start con(k) global
(l,lg,e1,n,ny,p,g2,z,g);

c=j(5,1,0);

sumc1=0;

do i = 1to p;

sumc1 = sumc1 + (
(g2*l**2)
/(l+k)**2 );

end;

sumc3=0;

do i = 1to p;

sumc3 = sumc3 + (((e1*l)+ ( k**2
* g**2 )) /(l+ k)**2);

end;

sumc41=0;

do i = 1to n;

sumc41 = sumc41 +(
(ny-((z[i,1]*g[1]*l[1])/(la[1]+k[1]))-((z[i,2]*g[2]*l[2])/(l[2]+k[2]))
)**2 );

end;

sumc42=0;

do i=1to n;

sumc42=sumc42+( 1-((z[i,1]**2)/(l[1]+k[1]))-((z[i,2]**2)/(l[2]+k[2]))
);

end;

sumc4=(n*sumc41)/(sumc42)**2;

c[1]= lenghtg-sumc1;

c[2]=k[1];

c[3]=k[2];

c[4]=3-sumc3;

c[5]=4-sumc4;

return (c);

finish con;

k=j(1,2,0);

optn=j(1,11,.);
optn[11]=0;
optn[10]=5;

call nlpqn (rc, kres, "fun", k, optn) nlc="con";

Re: Problem in NLPQN

Rick_SAS — Fri, 17 Apr 2015 13:45:16 GMT

Since you haven't provided values for the global variables, we can't run your program, but a likely cause of your problem is that you are dividing by zero. Notice that your functions have expressions like

sumf = sumf + ((stuff) /(l+ k)**2);

which is undefined when l=k=0.

Notice also that you have set k={0 0} as an initial value.

An easy way to debug these problems is to evaluate the objective functions (FUN) and the constraint function (CON) at several points in the domain prior to trying to use them in an optimization. For an example of evaluating the objective function and for getting a good initial guess, see the article "How to find an initial guess for an optimization"

Re: Problem in NLPQN

ayaafarag — Sat, 18 Apr 2015 14:57:22 GMT

I tried another initial values and this error message disappeared but I repeate this code 50000 times in simulation study. and now the runs do not stop even along the whole day. So can I use call nlpnms instead of call nlpqn for this optimization model?

Re: Problem in NLPQN

Rick_SAS — Sun, 19 Apr 2015 01:17:12 GMT

Would you like to share what you are trying to accomplish? What is the objective function? Are you sure that there is an optimum value?

As to the speed, you would solve this problem much faster if you got rid of all those DO loops and vectorized the problem. For example, it looks like the objective function is just

return( sum((e*l+ (k##2 # g##2)) /(l+ k)##2) );

If you aren't familiar with vectorization, look at some of these examples: http://blogs.sas.com/content/iml/tag/vectorization

Re: Problem in NLPQN

ayaafarag — Sun, 19 Apr 2015 22:04:09 GMT

I tried this code :

return( sum((e*l+ (k##2 # g##2)) /(l+ k)##2) );

but I got this error message

ERROR: (execution) Matrices do not conform to the operation.

operation : # at line 4183 column 39

operands : _TEM1002, _TEM1003

_TEM1002 1 row 2 cols (numeric)

0 25

_TEM1003 2 rows 1 col (numeric)

0.1154035

0.3658517

statement : RETURN at line 4183 column 1

traceback : module FUN at line 4183 column 1

ERROR: (execution) Matrices do not conform to the operation.

operation : NLPQN at line 4235 column 1

operands : *LIT1218, k, optn, , , , , , *LIT1219

*LIT1218 1 row 1 col (character, size 3)

fun

k 1 row 2 cols (numeric)

0 5

optn 1 row 11 cols (numeric)

*LIT1219 1 row 1 col (character, size 3)

con

statement : CALL at line 4235 column 1

Re: Problem in NLPQN

Rick_SAS — Mon, 20 Apr 2015 12:51:39 GMT

The error says that k is 1x2 whereas g is 2x1, so you'll need to transpose one of them.

Re: Problem in NLPQN

ayaafarag — Thu, 23 Apr 2015 08:26:18 GMT

After I remove all the do loops and specify the initial values, I got the following error message using NLPQN.

How can I increase the number of iterations or number of function calls inside the optimization code?

ERROR: QUANEW Optimization cannot be completed.

ERROR: QUANEW needs more than 200 iterations or 500 function calls.

WARNING: The point x is feasible only at the LCEPSILON= 0.1 range.

Re: Problem in NLPQN

Rick_SAS — Thu, 23 Apr 2015 12:20:12 GMT

Sometimes in constrained optimization there is no solution inside the feasible region. A canonical example is when you try to optimize the function f(x)=1/x subject to the contraint that x>0. For your problem, I suspect that the variable K is trying to get close to L. You can watch the progress by printing the iteration history.

The documentation describes how to print the iteration history and how to increase the iterations:

Control printing: SAS/IML(R) 13.1 User's Guide

Control stopping criterion: SAS/IML(R) 13.1 User's Guide