Originally posted by: pradeepram80

I'm working on a column generation implementation using OPL scripting. I'm noticing that the original model (without decomposition) solves in 2 minutes for a test case but my decomposed model takes longer. Upon further investigation, I'm finding that the generate() function in master problem takes longer to execute with every successive iteration. Any ideas on how this can be sped up? I'm deleting and recreating the IIOoplModel object for the master problem every iteration. Hence I have to call the generate() function every iteration. Is there any other way I can add the new column without having to generate the model again?

#DecisionOptimization
#OPLusingCPLEXOptimizer

Hi,

Indeed you may regenerate or you may also use setcoef in order to change the matrix.

See

https://www.ibm.com/developerworks/community/forums/html/topic?id=77777777-0000-0000-0000-000014727055

regards

#DecisionOptimization
#OPLusingCPLEXOptimizer

Originally posted by: pradeepram80

Thanks a lot for the reply. Helps.

Here is an example of a constraint whose coefficient I'm changing: (There are many such constraints)

lamda[p] is my decision variable in the restricted master problem.  p.moves[] comes from my subproblem.

 I will have to calculate the sum of               cc*p.moves[<f,a,t,s1,s2,ws1,b1,ws2,b2,sw1,sw2,ow>]  for every constraint defined in WBIndex if my understanding is correct. In order to do that I'll have to write "for loops" to access the right elements for the various tuple sets like setSequence. My questions:

a)    Is there a easier way to calculate the new coefficients other than using for loops?

b)    Is OPL scripting suitable to handle decomposition problems of this complexity? Should I be looking at CPP or Java callable libraries instead?

 

Regards,

pr

 

#DecisionOptimization
#OPLusingCPLEXOptimizer

Hi,

instead of loops you may use find in scripting.

Some prefer OPL scripting, some prefer CPP or java. Depends.

regards

#DecisionOptimization
#OPLusingCPLEXOptimizer

Originally posted by: pradeepram80

Alex,

Think I was probably not clear.

Could you please point me to examples/documentation where OPL scripting recalculates the right set of coefficients and then uses "setcoeff()" to update the coefficients? Also, should I declare empty variables beforehand for the column generation and then update them?

 

Regards,

Pradeep

#DecisionOptimization
#OPLusingCPLEXOptimizer

Hi

see https://www.ibm.com/developerworks/community/forums/html/topic?id=77777777-0000-0000-0000-000014839106#77777777-0000-0000-0000-000014839171

and

https://www.ibm.com/developerworks/community/forums/html/topic?id=77777777-0000-0000-0000-000014727055&ps=25

regards

#DecisionOptimization
#OPLusingCPLEXOptimizer

Originally posted by: pradeepram80

Hi,

Examples you provided do not show any recalculation of the coefficients. Let me share more  (Hoping to be very clear):

I'm writing for loops in OPL scripting to iterate through the various tuples to calculate new coefficients in each iteration, as shown below. These for loops take forever. I dont seem to find an alternate way to implement this in OPL scripting. Reason why I've for loops is because I'm joining the different tuples until i get a match.

Can you please let me know if there is a better way?

 

#DecisionOptimization
#OPLusingCPLEXOptimizer

Hi,

have you tried find ?

http://www-01.ibm.com/support/knowledgecenter/#!/SSSA5P_12.4.0/ilog.odms.ide.help/refjsopl/html/IloTupleSet.html

Or have you tried to use an array in order to directly reference items ?

regards

#DecisionOptimization
#OPLusingCPLEXOptimizer

Originally posted by: pradeepram80

Hi,

Your idea worked. Thank you. Here is my new code. I was able to get everything done in a single for loop. Do you think there could be further improvements in runtimes possible?

 

Regards,

pr

for (var seq in masterOpl.setSequence) {

//<f,a,t,s1,s2,ws1,b1,ws2,b2,sw1,sw2,ow> in seq --> <str1,str2,str3,str4,str5,str6,str7,str8,str9,int1,int2,int3>

//<f,ws2,b2,sw2> in capConstr --> <str1,str2,str3,int1>

//<f,t,s2,ws2,sw2> in CCIndex --> <str1,str2,str3,str4,int1,int2,float1,float2>

masterOpl.SumCoeff[masterOpl.WSBldngIndex.find(seq.str1,seq.str8,seq.str9,seq.int2)]

= masterOpl.SumCoeff[masterOpl.WSBldngIndex.find(seq.str1,seq.str8,seq.str9,seq.int2)]

+ subOpl.moves[seq]*masterOpl.CC[masterOpl.CCIndex.find(seq.str1,seq.str3,seq.str5,seq.str8,seq.int2)];

 

//<f,a,t,s1,s2,ws1,b1,ws2,b2,sw1,sw2,ow> in seq --> <str1,str2,str3,str4,str5,str6,str7,str8,str9,int1,int2,int3>

//<f,b1,b2,sw2> in BldngIndex --> <str1,str2,str3,int1>

//<f,ws1,b1,ws2,b2,sw2> in VehConstIndex2 --> <str1,str2,str3,str4,str5,int1>

if(!masterOpl.VehConstIndex2.find(seq.str1,seq.str6,seq.str7,seq.str8,seq.str9,seq.int2))

continue;

 

masterOpl.SumCoeff2[masterOpl.BldngIndex.find(seq.str1,seq.str7,seq.str9,seq.int2)]

= masterOpl.SumCoeff2[masterOpl.BldngIndex.find(seq.str1,seq.str7,seq.str9,seq.int2)]

+ subOpl.moves[seq]*masterOpl.VehConst[masterOpl.VehConstIndex2.find(seq.str1,seq.str6,seq.str7,seq.str8,seq.str9,seq.int2)];

 

//update moves variable in master opl

masterOpl.moves[count+1][seq]=subOpl.moves[seq];

}

#DecisionOptimization
#OPLusingCPLEXOptimizer

Originally posted by: pradeepram80

Also one more question:

In your post : https://www.ibm.com/developerworks/community/forums/html/topic?id=77777777-0000-0000-0000-000014839106#77777777-0000-0000-0000-000014839171

You show how to change the UB.

However when I try to change the UB of a similar constraint, I get the following error:

Description Resource Path Location Type
Scripting runtime error: cannot add properties to this value, "[a IloConstraint]". 

Documentation for IloConstraint says:

Caution:
For technical reasons, the lower and upper bound properties LB and UB are currently accessible only during postprocessing and flow control in constraints assigned inside a forall declaration. The undefined value is returned instead. You can nevertheless modify the bounds.

So I defined the constraint with a forall statement but I still see the same error. Ideas on getting around this problem will be greatly appreciated.

Here is how I'm defining the constraint in my model:

{ti} N={<1>};

minimize something;

subject to {

 some constraints

forall(<n> in N) ctIter2[<n>]: lamda[1] == 1;

}

main {

....

for (var nn in masterOpl.N)

masterOpl.ctIter2[nn].UB=0.0;

 

}

Can you show how to make this work?
 

 

 

 

#DecisionOptimization
#OPLusingCPLEXOptimizer

Hi,

have you done the masterOpl.generate();

?

regards

#DecisionOptimization
#OPLusingCPLEXOptimizer

Originally posted by: pradeepram80

Yes I did. That was my first attempt. However the generate function takes about 60-90seconds per iteration. Using the above for loop reduces it to ~30 seconds. Its good but not enough. So wondering if there is any other way I could write the for loop.

Also need your thoughts on the previous post - how to change the UB of convexity constraint.

Thanks,

Pradeep

#DecisionOptimization
#OPLusingCPLEXOptimizer