[Coin-ipopt] question on "overall NLP error"

[Andreas Waechter]

Tran,

Please send messages regarding Ipopt to the Ipopt mailing list.

On Tue, 6 Nov 2007, Tran Minh Tuan wrote:

> Thanks, it clarifies !
> But in your experience, what can we say about a result in which primal 
> infeasibility is around 1e-5, complementarity is around 1e-11 but dual 
> infeasibility is around 1e02 ?

Under usual circunstance, 1e02 is a very large error.  Everything of 
course depends on the scale of the problem, but usually I would not 
considered a problem with that large an error to be solved or close to be 
solved.

Andreas

> On 5 Nov 2007, at 9:33 PM, Andreas Waechter wrote:
>
>> Hi
>> 
>> You also asked about the meaning of the different error types.  Their 
>> definition can be found in the Ipopt implementation paper,
>> 
>> http://www.research.ibm.com/people/a/andreasw/papers/ipopt.pdf
>> 
>> on page 3, Eq. (5).
>> 
>> The terms in the max are:
>> 
>> 1. Dual infeasibility
>> 2. Primal feasibility
>> 3. Complementarity
>> 
>> And the quantity E_{\mu}(x,\lambda,z) is the overall NLP error (with \mu=0)
>> 
>> Hope this clarifies,
>> 
>> Andreas
>> 
>> On Sun, 4 Nov 2007, Peter Carbonetto wrote:
>> 
>>>> I suppose that the derivative check module is very useful, I used it to
>>>> fix gradient calculation and now my program runs quite well. But although
>>>> the obtained result is good, I found that "dual infeasibility" in the
>>>> output of the program is still high (around 1.0e02) (the number for
>>>> "overall NLP error" is the same).
>>>> So I wonder what is the role of "dual infeasibility" into the solution 
>>>> and
>>>> whether we can have a good result even this error is not perfect.
>>>> By the way, how do the "constraint violation" and "complementarity" are
>>>> calculated and their signification ? They are not cummulated to "overall
>>>> NLP error" ?
>>> 
>>> You might find this previous discussion on the IPOPT mailing list useful:
>>> http://list.coin-or.org/pipermail/coin-ipopt/2007-February/000700.html
>>> 
>>> Peter Carbonetto
>>> Ph.D. Candidate
>>> Dept. of Computer Science
>>> University of British Columbia
>>> _______________________________________________
>>> Coin-ipopt mailing list
>>> Coin-ipopt at list.coin-or.org
>>> http://list.coin-or.org/mailman/listinfo/coin-ipopt
>>> 
>
>