[Dip] Extreme rays and artificial column bounds

[kim]

I have now updated to trunk and no longer get the CBC errors. I am 
sometimes getting warnings about slightly violated constraints but more 
seriously I am often getting an infeasibility error when I am nearing 
termination of MILPBlock. I get the following error message:

"Solve as IP found a solution.
COIN Exception [ Recomposed solution is not feasible ] at 
../../../Dip/src/DecompAlgoPC.cpp:L815 in DecompAlgoPC::solutionUpdateAsIp"

Inspecting the code it seems this happens when solving the restricted 
master as an IP. I have turned up the log levels, but don't know 
MILPBlock/DIP well enough to make sense of it. I didn't want to clutter 
up the mailing list with long output transcripts, but I can email you 
the last bit of output if you like.

It seems to me the easiest temporary fix for me would be to disable the 
upper bound heuristic. Is there a way to do this in the parameter file 
or would I need to modify my copy of the DIP code?

- Kim


Matthew Galati wrote:
>> I'm a bit puzzled by this. CBC 2nd status=7 means the LP-relaxation is
>> unbounded (according to "CbcModel.hpp"). But the pricing problems
>> generated in my modified block file are all 0-1 Knapsack Problems and
>> should never be unbounded, unless the upper bound is also relaxed along
>> with the integrality requirement for some reason.
>>     
>
>
> [MVG] Right. If it does return that for one of your knapsack problems -- then it is a CBC bug.
>
> Turn up the log levels to see when that happens:
> [DECOMP]
> LogDebugLevel    = 3
> LogLevel         = 3
> LogDumpModel     = 2
> LogLpLevel       = 1
>
>
>
>
>  
>   
>> Is CBC used for anything besides pricing problems? Your dissertation
>> mentions  solving the restricted master problem to optimality to obtain
>> valid upper bounds (assuming minimization), so I would assume CBC could
>> also be used here.
>>     
> [MVG] Yes - it is also used to solve the master problem as an IP. That is, it periodically changes the variables in the master problem (lambda) to integers and solves the master as an integer program. This is a heuristic to try and get incumbent solutions for the entire problem. This, in my thesis, is called Price-and-Branch.
>
> Let me know if it is failing in the subproblem or the master IP.
>
>
>
>   

-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://list.coin-or.org/pipermail/dip/attachments/20101012/d313fb15/attachment.html