[Coin-discuss] Problems with Rounding Heuristic

[Esben Mose Hansen]

On Friday 02 February 2007 10:19, Mauro Scarioni wrote:
> I've a problem using CbcRounding heuristic.
> I've a problem where i need only to find a feasible integer solution, the
> obj function is a minor interest.
> In this sense i set
>
> (thisModel)->setMaximumNodes(1);

This won't work. The branch&bound might use quite a few nodes before finding 
any solution. I have personally seen serveral 100K nodes without a solution.

>
> in order to stop the B&B at the first integer solution

try is setAcceptableGap(1e40) or something like that. Or you could install a 
user event handler.

>
> In some instances i obtain the following log message:
>
> ...................................................
> Cbc0015I Node 8556 Obj 0 Unsat 7 depth 22
> Cbc0015I Node 8557 Obj 0 Unsat 7 depth 22
> Cbc0012I Integer solution of 1e+050 found by heuristic after 184896
> iterations and 8557 nodes (68.91 seconds)
> Cbc0019I Exiting on maximum solutions
> Cbc0005I Partial search - best objective 1e+050 (best possible 0), took
> 184901 iterations and 8558 nodes (68.91 seconds)
> Cbc0006I The LP relaxation is infeasible or too expensive.
>
> This solution is not tracked by newSolution method of CbcCompare and is not
> usable calling the last solution

This is coinspeak for no integer solution found. 

>
> But the number of solution find is incremented.

I have seen this too. I think it is a bug, sometimes a solution is reported by 
CbcRounding (I think) even though no solution is found.

>
> It is possible to define a skip of the heuristic solution finded in this
> particular case?

Probably, but that won't really help you.


> IBM Global Business Services

Heh, I used to work there, I think. (I never really found out where I worked 
in IBM, exactly).


-- 
kind regards, Esben