[Ipopt] enforcing once a step size of alpha=1

Stefan Vigerske stefan at math.hu-berlin.de
Mon Apr 24 12:36:50 EDT 2017


there are a number of hidden options in Ipopt to control the line 
search, but nothing that seems to do exactly what you are requesting.
Anyhow, the GAMS/Ipopt documentation has a list of all options, maybe 
something is useful for your:


On 04/24/2017 03:40 PM, Martin Neuenhofen wrote:
> Hi all,
> my problem has linear equality conditions and my problem is badly scaled. I
> thought that ill scaling of linear equality constraints is not necessary to
> be considered because it once solves the residual to zero and then does not
> occur in the step size scaling any more, BUT however this only holds after
> the residual of this equality has once reached zero.
> Explanation:
> Solve F(x)=0 with F = [A*x1-b ; G(x2)] for x = [x1;x2].
> In this case the Newton Step is x = x - alpha * [invA, 0; 0, inv(DG)] *
> [A*x1-b ; G(x2)] , where alpha is the step size. So when A*x1-b is not zero
> yet the NLP is ill-scaled or the Hesse-approx bad, then the residual of
> these equalities will slow down convergence *until* one step with alpha=1
> is performed because afterwards A*x1-b=0 will always hold for all
> subsequent steps.
> Therefore, is there an option in ipopt such that I can once enforce in the
> beginning a stepsize of alpha=1? If not, did someone have a similar issue
> and found a work-around? I am afraid one cannot simply use an alpha=1 in a
> hand-build feasibility Newton solver in advance of the first step because
> of the inequality constraints.
> Kind regards,
> Martin
> _______________________________________________
> Ipopt mailing list
> Ipopt at list.coin-or.org
> https://list.coin-or.org/mailman/listinfo/ipopt


More information about the Ipopt mailing list