[Coin-ipopt] treatment of uncomputable values in eval_f, eval_grad_f

[Andreas Waechter]

Hi James,

> We are using Ipopt to do parameter estimation of ODE systems.
> For some parameter values, either the forward or the adjoint solution fails.
> It seems that in those cases, one should do a "return false" in the functions
> eval_f and eval_grad_f respectively and Ipopt would do the necessary
> backtracking to computable function values.

That's the idea.

> However, it seems that if eval_f is computed successfully, but eval_grad_f
> fails, Ipopt does not backtrack, but instead returns the message "EXIT:
> Invalid number in NLP function or derivative detected".
> Are we doing something wrong?

Well, that doesn't sound like Ipopt is doing what I would hope it to do...

Could you open a ticket at the Ipopt trac page and include some 
description of what is going on?  In particular, could you please attach 
(to the ticket, please don't send that to the mailing list) the output of 
Ipopt if you set the "print_level" option to 8, and "print_info_string" to 
"yes"?  Also, please include which of your eval_* methods has been called 
most recently before Ipopt quits.

Thanks

Andreas