[Cbc] reformulated s2 bug repro

[John Forrest]

James,

I can see the problem.  It is not just an SOS issue, although it is most 
likely to occur there.  In the Lp's solved in branch and bound 
feasibility is a tolerance issue while for IPs/SOS it is a yes-no issue.

Throwing all fixed variables out of basis before solving LP would fix 
problem, but has significant performance issues.  I am thinking (which 
hurts) and I think I will have to add an option so that if integer/sos 
variables are "just infeasible" after a solve then they will be made 
non-basic at bound and the problem resolved.  This option may add 1-2% 
to solution times.

May take a few days.

John Forrest
On 14/05/18 19:03, James Howey wrote:
>
> I sent in an s2 issue yesterday.
>
> Here is a slightly more reduced repro:
>
> Maximize
>
> b1y
>
> SUBJECT TO
>
> -agi + 0 m1 +32000 m2 + 1e8 m3 = 0
>
> m1 + m2 + m3 = 1
>
> -b1y +32000 m1 = 0
>
> BOUNDS
>
> agi = 1
>
> m1 <= 1E+15
>
> m2 <= 1E+15
>
> m3 <= 1E+15
>
> SOS
>
> X: S2:: m1:1 m2:2 m3:3
>
> END
>
> With latest trunk build:
>
> cbc b1.lp -preprocess off -solve -solu a –quit
>
> comes up
>
> Integer infeasible - objective value -31999.99968000
>
>       0 b1y                32000                       0
>
>       1 agi                    1                 0.00032
>
>       2 m1            0.99999999                       0
>
>       3 m2                     0                31989.76
>
>       4 m3                 1e-08                       0
>
> Where cplex says:
>
> MIP - Integer optimal solution:  Objective =  3.1999000000e+04
>
> Solution time = 0.00 sec.  Iterations = 0  Nodes = 0
>
> Deterministic time = 0.01 ticks  (6.71 ticks/sec)
>
> CPLEX> display solution variables *
>
> Incumbent solution
>
> Variable Name           Solution Value
>
> b1y             31999.000000
>
> agi 1.000000
>
> m1 0.999969
>
> m2 0.000031
>
> All other variables matching '*' are 0.
>
> If I change the 1e8 to 1e7, then both cbc and cplex reach correct 
> optimum of 31999.
>
> Thanks for any attention you give this.
>
> jkh
>
>
>
> _______________________________________________
> Cbc mailing list
> Cbc at list.coin-or.org
> https://list.coin-or.org/mailman/listinfo/cbc


-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://list.coin-or.org/pipermail/cbc/attachments/20180515/0edd8f17/attachment.html>