[Couenne] Rounding NLP heuristic

[Pietro Belotti]

Akshay,

Couenne runs Ipopt at the beginning to search for a feasible solution, 
both when solving NLPs and MINLPs. However, Ipopt might find a solution 
that is not "feasible enough" for Couenne, i.e., the constraints 
violations might be too large. All solvers, including Ipopt, allow slight 
violation and have a tolerance parameter. You may want to set that 
tolerance a little higher by adding, to the couenne.opt option file, a 
line

feas_tolerance 1e-5

or higher if necessary.

Regards,
Pietro

--
Pietro Belotti
Dept. of Mathematical Sciences
Clemson University
email: pbelott at clemson.edu
phone: 864-656-6765
web:   myweb.clemson.edu/~pbelott

On Thu, 31 Mar 2011, Akshay Gupte wrote:

> Pietro,
> I am testing some continuous NLP problems with Couenne. Although these 
> problems are large in size (~300 variables), I know some trivial 
> feasible solutions. However, Couenne is unable to find any feasible 
> solution after a long time (many hours). When I solved the NLP with 
> Ipopt, I do get a nontrivial feasible solution without providing any 
> starting point of my own.
>
> Does this mean that for NLPs, Couenne does not really try to solve the 
> NLP with a local optimizer at some nodes in the branch-and-bound tree ? 
> Is there any heuristic in-built specifically for NLPs ?
>
> Regards,
> Akshay Gupte
> PhD student, Operations Research
> Georgia Institute of Technology
> akshayg at gatech.edu
>
>
>
>
>
>
> On Mar 29, 2011, at 8:59 AM, Pietro Belotti wrote:
>
>> Akshay,
>>
>> we are working on it now, it should be available in a month or two.
>>
>> Pietro
>>
>> --
>> Pietro Belotti
>> Dept. of Mathematical Sciences
>> Clemson University
>> email: pbelott at clemson.edu
>> phone: 864-656-6765
>> web:   myweb.clemson.edu/~pbelott
>>
>> On Tue, 29 Mar 2011, Akshay Gupte wrote:
>>
>>> Pietro,
>>> Thanks for the info. Is there any estimate on when FP will become available, atleast in the trunk version ?
>>>
>>> Regards,
>>> Akshay Gupte
>>> PhD student, Operations Research
>>> Georgia Institute of Technology
>>> akshayg at gatech.edu
>>>
>>>
>>>
>>>
>>>
>>>
>>> On Mar 28, 2011, at 8:21 PM, Pietro Belotti wrote:
>>>
>>>> Akshay,
>>>>
>>>> just to add to what Giacomo said: FP is still under development and not usable at all.
>>>>
>>>> Pietro
>>>>
>>>> --
>>>> Pietro Belotti
>>>> Dept. of Mathematical Sciences
>>>> Clemson University
>>>> email: pbelott at clemson.edu
>>>> phone: 864-656-6765
>>>> web:   myweb.clemson.edu/~pbelott
>>>>
>>>> On Mon, 28 Mar 2011, Giacomo Nannicini wrote:
>>>>
>>>>> Both FP and IR are in Couenne trunk. I do not know if they are
>>>>> released in stable yet.
>>>>> IR is working in trunk, but it is *highly* recommended to compile with
>>>>> Cplex if you want to use it.
>>>>>
>>>>> Giacomo
>>>>>
>>>>> On Mon, Mar 28, 2011 at 5:32 PM, Akshay Gupte <akshayg at gatech.edu> wrote:
>>>>>> Giacomo,
>>>>>> Thanks for the tip. I will try the branching idea.
>>>>>>
>>>>>> Are FP and IR codes publicly available within Couenne or otherwise ?
>>>>>>
>>>>>> Regards,
>>>>>> Akshay Gupte
>>>>>> PhD student, Operations Research
>>>>>> Georgia Institute of Technology
>>>>>> akshayg at gatech.edu
>>>>>>
>>>>>>
>>>>>>
>>>>>>
>>>>>>
>>>>>>
>>>>>> On Mar 28, 2011, at 5:29 PM, Giacomo Nannicini wrote:
>>>>>>
>>>>>>> Akshay,
>>>>>>> typically what I do when I am only interested in heuristics, is to set
>>>>>>> the maximum number of nodes to 0 (the option is: node_limit 0).
>>>>>>> However, as far as I know the default heuristic Couenne Rounding NLP
>>>>>>> needs candidate points discovered during Branch-and-Bound as input. If
>>>>>>> you stop at the root node, you will not have many candidates. Try to
>>>>>>> branch for a large enough number of nodes with a fast branching
>>>>>>> strategy, to get many candidate points.
>>>>>>>
>>>>>>> You should also consider using some of the other available heuristics
>>>>>>> (FeasibilityPump and IterativeRounding). Those can do a lot of work
>>>>>>> just at the root node.
>>>>>>>
>>>>>>> Giacomo
>>>>>>>
>>>>>>> On Mon, Mar 28, 2011 at 5:09 PM, Akshay Gupte <akshayg at gatech.edu> wrote:
>>>>>>>> Hi all,
>>>>>>>> Is there an option to run only the rounding heuristic within Couenne for a specified amount of time ? I am particularly interested in finding feasible solutions to my model.
>>>>>>>>
>>>>>>>> Regards,
>>>>>>>> Akshay Gupte
>>>>>>>>
>>>>>>>>
>>>>>>>>
>>>>>>>>
>>>>>>>>
>>>>>>>>
>>>>>>>>
>>>>>>>> _______________________________________________
>>>>>>>> Couenne mailing list
>>>>>>>> Couenne at list.coin-or.org
>>>>>>>> http://list.coin-or.org/mailman/listinfo/couenne
>>>>>>>>
>>>>>>
>>>>>>
>>>>>
>>>>> _______________________________________________
>>>>> Couenne mailing list
>>>>> Couenne at list.coin-or.org
>>>>> http://list.coin-or.org/mailman/listinfo/couenne
>>>>>
>>>
>>>
>
>