5. So summing up : without SAS 9.22 or 9.3 and IMPVAR statement I won't be able to solve it in short time? Is there any reason why Excel solves it in 2-3 seconds? I found that Excel Solver uses  Generalized Reduced Gradient Algorithm http://support.microsoft.com/kb/82890 ( Solver Uses Generalized Reduced Gradient Algorithm) . Is there a possibility to use the same algorithm in optmodel and get results in seconds without running out of memory? ... View more

2. I was running codes that you posted https://communities.sas.com/servlet/JiveServlet/download/116717-11110/tom022012_heur.sas.zip and https://communities.sas.com/servlet/JiveServlet/download/116717-11111/tom022012.sas.zip but slightly modified - I just removed impvar statements that caused errors in sas 9.2 and included parts of that codes in objective function. Now I know that impvar statements should be replaced with var and con statements and properly formulated quadratic function with QP solver works effectively. 3. Ok, now I understand. You mean to use drop statement after QP solver and before MILP solver not removing constraints form the whole code. Right? 4. Good to know. 5. I seem to have still problem here. When I changed objective to non quadratic one I cannot solve problem even with 1000 observations in less than 10 minutes (300 obs - 2 seconds, 500 - 10 seconds, 1000 -10minutes). The time needed increases expotentially and since I need to solve problem with 10.000 or more observations it becomes unsolvable for me. Morover I run out of memory with 2000 observations. What are solving times on you computer when you change solver to eg. NLPC and change objective to : max max_shp = a / (sqrt(sum {i in VSET} (r - a)**2)/NOBS);    (AVERAGE/STDEV) 6. Is there a possibility to enforce SAS to use some disk space so that the solver wouldn't run out of memory? I need to make code that always works - even if the problem is too complicated and lasts long. ... View more

Thank you for help. Coming back to discussion: 1) Almost at the same time I started to try enforce different solvers and figured out also that QP is the quickest. I'ts good that you recommend QP - I wasn't sure If I can trust QP beacuse it has "experimental" label. Thank you 2) After adding proper constraints istead of impvar my calculation times dropped to 10 sec with 10.000 obs. and seem to grow linearly (in comparison to expotential growth without QP solver and proper constrainsts). Now the times are really ok to me - however further speed up is important to me as I may have to make such optimizations many times (100, 1000) depending on scenario defined. 3) Can you explain why I should make sure to drop rc and avgc constraints before solving MILP?. I thought that I should solve normal QP problem as quick as possible and then find solutions that meet discrete criteria most close to my optimal solution (solved in step one)? 4) You mentioned somewhere that OPTMODEL in SAS 9.3 is much quicker than in SAS 9.22 . Does it apply to SAS 9.22 too? Can you roughly estimate how quicker can it be with my problem on the same configuration? 5) One thing I did not realised could be an issue here too: Apart from minimization of standard deviation I need to maximize also Average/standard deviation ratio. I thought that I will use the same solver in this case too but QP solver says that It needs quadratic objective and can't solve it. How can I formulate such objective in order to use QP solver or other solver to get reasonably quick result? 6) When I try to solve Average/standard deviation ratio problem with just SOLVE then I constantly get out of memory error saying that SAS needs 295 MBytes of memory : (ERROR: Unable to allocate sufficient memory. At least 294925K bytes were requested. You must        either increase the amount of memory available, or approach the problem differently) First af all SAS has much memory available more for the problem. Secondly I tried to put various statements in SAS cfg file (-MEMSIZE 0, -MEMSIZE MAX , -LOADMEMSIZE 0 etc.) without success. Can you help me with setting sas not to run out of memory possibly ever - my scripts will be run hundreds of times so I prefer to get results even later to getting just an error. By the way - on SAS startu I get message in log:   Memory                        3908k   OS Memory                     4976k Does SAS mean megabytes instead of kilobytes or these are kilobytes and it is a real cause of my memory problems? I have 3GB of RAM. ... View more

1) I attached zipped files with over 67000 observations. I need to make optimization on around 20.000 observations. 2) I use SAS 9.2 so I can't use impvar declarations which, if I undestood well, significantly increase speed of optimization. I tried to substitute them with var and con declarations but I get error: Your original code: impvar r {i in VSET} = (sum {j in WSET} values[i,j] * weights ) - margins; impvar avg = (sum {k in VSET} r ) / NOBS; min min_std = sum {i in VSET} (r - avg)**2; num std = sqrt(min_std.sol / NOBS); My try to replace them: var r {i in VSET}; con rc {i in VSET} = sum {j in WSET} values[i,j] * weights - margins; var avg; con avgc = sum {k in VSET} r / NOBS; min min_std = sum {i in VSET} (r - avg)**2; num std = sqrt(min_std.sol / NOBS); How should I change them to work on SAS 9.2? ... View more

SAS 9.2. The problem is the same as one in answer 9. - the code with heuristic approach: In that example there were only 3 observations. I need to perform optimization on at least 20000 many times so I was expecting that SAS can make one optimisation much quicker than Excel's solver. When I try with 1000 observations SAS with heuristic approach needs 54 seconds to optimize while Excel's solver with 10.000 oservations (10x more) needs 5 second. I don't know why the difference is so huge. ... View more

hm.. variable 1 - 10 combinations variable 2 - 10 combinations ... variable n - 10 combinations then alltogether 10*10 * .... = 10^n without sum condition ? Am I wrong? ... View more

There's something wrong with the syntax. Which ones should be of macro style (%if) and which of data step style (without %)? As far as i know if you change "if sum (of v1-v&N)=100 then output;" to macro style then you can't do arithetics. Thanks for help anyway. ... View more

works thank, but is there a possibility to enter near linear some generic contraint with meaning: sum of all x variables like:  linear sum (x ) = 10 or similar? Then I wouldn't have to generate dynamically constraint with dynamic number of variables. I will have N variables - N can change and will be a input parameter of a macro. ... View more

It is helpful with 10 or 15 variables - then the full number of combinations would be 10^10 and 10^15 respectively and with given sum the number decreases dramatically. ... View more

The problem is that I need to calculate that with given number of variables. I need a macro that has N-variables parameter on entry. In that case depending on N I will have N loops ... View more

For example - I need all combinations of 3 variables v1, v2, v3 (each can be either of these values: 0,10,20,30,40.... 100) but only combinations in which v1+v2+v3 =100 v1 v2 v3   sum 10 80 10  100   - OK 10 90 10  110   - I don't want this combiantion 30 30 40  100   -OK This is an example. I need to have possibility to make combinations for N variables. The clue is not to make all combinations and afterwards eliminating those that don't meet conditions - the number of all combinations is much greater then the number of combinations with sum=100 - So during generation of combinations sum should be checked to optimize generation time. ... View more