[SCIP] Initial primal bounds when solving an NLP globally with SCIP

[Stefan Vigerske]

Hi,

whether Ipopt finds a local optimum and which one that is depends on 
many aspects:
- the problem, which is modified by SCIP, e.g., presolve may do 
reformulations and domain propagation can change variable bounds
- the starting point, which within SCIP is probably taken as the 
solution of the LP relaxation, which is very likely different than what 
you used to run Ipopt alone
- the Ipopt parameters, which are changed by SCIP's interface to Ipopt
- different code for the evaluation of functions and first and second 
derivatives, which can lead to small differences

Stefan

On 7/22/21 9:54 PM, Nelson Uhan wrote:
> Dear SCIP team,
> 
> I'm trying to globally solve an NLP with SCIP (SCIP 7.0.2, compiled with
> Ipopt 3.13.5). I've found that sometimes, the local optimal value found by
> Ipopt alone is better than the initial primal bounds found by SCIP. I see
> this behavior with and without presolving. Is this behavior expected?
> 
> This behavior seems counterintuitive to me: after solving the NLP at the
> root node with Ipopt, it seems that SCIP should be able to establish a
> primal bound equal to the local optimal value that Ipopt alone finds,
> unless SCIP runs Ipopt with different settings?
> 
> Any guidance you could provide would be really helpful. Thank you very much!
> 
> Best regards,
> Nelson
> 
> 
> _______________________________________________
> Scip mailing list
> Scip at zib.de
> https://listserv.zib.de/mailman/listinfo/scip
>