Re: [TANGO Project - ALGENCAN] Abridged summary of tango-project@googlegroups.com - 3 Messages in 1 Topic

[Andrea Cassioli]

Hi Mario,
I send you part of the log of the failing local search. In this
execution there are 80 variables and roughly half million of terms in
the objective function. The checkder routine reports no errors. For
what I understand the backtracking continues up to a null
step......arent' any safeguards?

--
Andrea Cassioli

[Jose Mario Martinez]

  Andrea: I read your output file. As I understand, at the current point the objective function value is approx 120, but, in the
line search, even with alpha = 1.e-70, the function value at the trial point is approx 630.
Of course this is impossible, because with such a step the trial point and the current
point are identical.
  As I remember, yes, there are safeguards, but this is not the main problem. The function
that is being considered at the current point is not the same as the one at the trial point, if
I undestand correctly the output. Ernesto may check it.
  Either there is a very simple mistake in the evaluation or a very stupid bug in Gencan. We
will find out.
  At this point we need Ernesto's help. I am not sure whether you are using the fortran version,
or a C interface (about which I don't know anything).
  Cheers,
  Mario

--
Andrea Cassioli

--
You received this message because you are subscribed to the Google Groups "TANGO Project - ALGENCAN" group.
To post to this group, send email to tango-...@googlegroups.com.
To unsubscribe from this group, send email to tango-projec...@googlegroups.com.
For more options, visit this group at http://groups.google.com/group/tango-project?hl=en.

[Jose Mario Martinez]

  Another impossibility is that, if we take seriously the output, the function could not be convex. Observe that when alfa is
reduced the objective function grows, and then decreases. This concavity cannot occur in this problem.
  Mario