[SCIP] conflict analysis

[Marc Pfetsch]

Dear James,

yes, encoding the reason into just an int is quite limiting. In most 
cases, this was enough in the past and this information has to be stored 
for every bound change - so it is better to keep storage small.

Your consdata might have significantly changed between the time you 
performed the propagation and the time the resolving happens, because 
the latter is called at the time infeasibility occurs - possibly much 
deeper in the tree than the propagation.

One solution that often helps is to just indicate the type of reduction 
and then rerunning the reduction steps. You can access the bound changes 
as they have been at the time of reduction - so it is "just" a question 
of running time. Of course, if your reduction is slow, then this might 
not be an option.

Hope this helps.

Best

Marc

On 01/08/2024 13:18, James Cussens wrote:
> Hi all,
> 
> I am considering adding conflict analysis to a constraint handler 
> (CONSRESPROP) for which I already have a propagator (CONSPROP). The 
> constraint handler involves only n X n binary variables which represent 
> the adjacency matrix for a directed graph. The "reason" for a 
> propagation is that a set of binary variables are locally fixed to 1 
> which means that another set of binary variables should all be fixed to 0.
> 
> I'm predicting that encoding the "reason" - the relevant set of binary 
> variables locally fixed to 1 as a single inferinfo int will be tricky, 
> particularly for large values of n. That's a bit frustrating since it's 
> easy enough to store this set in the consdata at propagation (CONSPROP) 
> time. Is there any reason to think that this information will not still 
> be in my consdata when CONSRESPROP is called?
> 
> James
> 
> James Cussens
> Room MVB 3.26
> School of Computer Science, University of Bristol
> Phone: +44 (0)117 455 8723
> https://jcussens.github.io/ <https://jcussens.github.io/>
> 
> _______________________________________________
> Scip mailing list
> Scip at zib.de
> https://listserv.zib.de/mailman/listinfo/scip