Originally posted by: davidoff

Hello

It is possible to run cp.solve() once the model has been notified through x.setLB() or x.setUB() without extracting the whole model.
However, this seems impossible while setting new coefficients for the objective function, which is often the case in a column generation scheme.

Exception from IBM ILOG Concert: You can not change the coefficient of a nonlinear expression

One workaround would be to define dual infos as additional variables and updating these variables with setBounds. This is also impossible since only discrete variables are authorized while dual values are float.

The only acceptable workaround I found is to remove the objective from the model and add a modified one. Is that the best way ? Is it possible also with OPL ?

David
#DecisionOptimization
#OPLusingCPOptimizer

Originally posted by: rdumeur

Dear David,

Thank you for reporting this problem. Could you please tell me what API do you use to set new objective coefficients?
Is it the IloObjective::setLinearCoeff(...) method used in the documented example about column generation?
Regarding OPL, we are asking the question to the relevant team and will be back to you as soon as we know more.

Cheers,

Renaud
#DecisionOptimization
#OPLusingCPOptimizer

Originally posted by: davidoff

Hello

i tried first in OPL to modify the objective this way :

main 
{   var m = thisOplModel; ... m.generate(); 

if (cp.solve()) 
{ 

for(var itv in m.couvertes)
{ m.lock[itv].UB = 0;
//OK : change bounds of a variable m.getObjective().setCoef(m.presenceCover[itv],2);
//ERROR Exception from IBM ILOG Concert: You can not change the coefficient of a nonlinear expression   
} 
}

However, adapting the C++ documentation sample facility.cpp (provided in cpoptimizer samples), it is there possible to update the objective function with

objective.setLinearCoef( myvar, mycoeff)

I haven't tried so far to test this approach to my original OPL model

Thanks

David
#DecisionOptimization
#OPLusingCPOptimizer

Hello,

what you could do is add artificial decision variables in order to modify the objective function as you like.

let me give you an example:

using CP;   dvar int+ x in 0..10; dvar int+ y in 0..10; dvar 

int c1; dvar 

int c2;     dvar 

int z;   maximize z ;   subject to 
{   z==c1*x+c2*y; x+y==10; 
}   main 
{ thisOplModel.generate(); thisOplModel.c1.UB=2; thisOplModel.c1.LB=2; thisOplModel.c2.UB=1; thisOplModel.c2.LB=1; cp.solve(); writeln(
"x=",thisOplModel.x); writeln(
"y=",thisOplModel.y); writeln(cp.getObjValue()); thisOplModel.c1.UB=1; thisOplModel.c1.LB=1; thisOplModel.c2.UB=3; thisOplModel.c2.LB=3;   cp.solve(); writeln(
"x=",thisOplModel.x); writeln(
"y=",thisOplModel.y);   writeln(cp.getObjValue()); 
}

which gives

x=10 y=0 20 x=0 y=10 30

NB:

if you need c1 and c2 to be decimal you may replace c1 and c2 by c1/100000 and c2/100000

Regards
#DecisionOptimization
#OPLusingCPOptimizer

Originally posted by: davidoff

Excellent !
Thanks for the workaround

In order to avoid loosing precision with rounding float values, I consider multiplying and dividing by 100000 the fractional part of the coefficients (see code below)

using CP;   

int dig = 6;
//how many digits do we keep range r = 1..2; 

float dual[r] = [1567.2389,  3.000678]; 

int ent[i in r] = ftoi(round(dual[i])); 

float res[i in r] = dual[i]- round(dual[i]); 

int resd[i in r] = ftoi(round(pow(10,dig)*res[i]));   dvar 

int coeffEnt[i in r] in ent[i]..ent[i]; dvar 

int coeffRest[i in r] in resd[i]..resd[i]; dexpr 

float coeff[i in r] = coeffEnt[i] + pow(10,-dig)*coeffRest[i];   dvar int+ x[r];   minimize sum(i in r) coeff[i]* x[i];   subject to 
{ sum(i in r) x[i] >=1; 
}   execute
{
//postprocess writeln(x); 
}   main
{ thisOplModel.generate(); 

if (cp.solve()) 
{ writeln(
"RUN 1 objvalue = ",cp.getObjValue()); thisOplModel.postProcess(); 
} writeln(
"update bounds"); thisOplModel.coeffEnt[1].LB= 2;
//change dual[1] to 2.2389 thisOplModel.coeffEnt[1].UB= 2;
//change dual[1] to 2.2389 

if (cp.solve())
{ writeln(
"RUN 1 objvalue = ",cp.getObjValue()); thisOplModel.postProcess(); 
} 
}

This code simply picks up the item with the smallest coefficient. The first coefficients are 1567.2389, 3.000678 and later I change them in 2.2389, 3.000678 . The result is correct :

RUN 1 objvalue = 3.000678 [0 1] update bounds RUN 1 objvalue = 2.2389 [1 0]

#DecisionOptimization
#OPLusingCPOptimizer