[Ipopt] Some suggestions to speed up code sought

Jonathan Hogg jonathan.hogg at stfc.ac.uk
Thu Nov 12 05:48:21 EST 2015


The system solved by MA97 changes each iteration, with diagonal entries
in the (1,1) block becoming more extreme towards 0 and +infinity [give
or take the diagonal entries of the Hessian] depending on whether the
associated constraint is active or not. This is what I'd expect to be
causing the singular error messages at some iterations and not at others
- it is perfectly normal for some problems (and by singular here we mean
effectively numerically singular rather than structurally or in exact
arithmetic).

IIRC Ipopt just takes a singular error return as a signal to increase
regularization of the system in the (1,1) block and then tries again.
Hence it is probably just a case of needing a patch to tell MA97 not to
print error messages where the user can see them :P.

It may be worth playing with the ma97 scaling options if you are getting
this a lot, and see if they make any difference. There is also the
control.small value in the ma97 interface which defines what values are
interpreted to define a matrix as singular. However you probably need to
comprehend the Ipopt algorithm as a whole before determining if that is
a sensible thing to do.

If your Hessian isn't changing, does that imply your problem is linear
or quadratic? There may be better solvers for such problems than Ipopt -
dedicated LP/QP solvers are able to better exploit the structure and/or
deal with the degeneracy in those cases iirc.

Jonathan.

On 11/11/15 18:27, jmogali at andrew.cmu.edu wrote:
> Hi,
>      Yes, this is despite setting Hessian as constant. On reading the
> function description for ReOptimizeTNLP(), it turns out that all they
> are expecting is the non zero structure of the matrices to be the
> same and nothing about the actual values. So, logically they would
> need to recompute the Hessian.
> 
> Regarding the MA97 throwing error messages, I feel it is rather strange. I
> would like to hear your take on the matter.
> So the Hessian and Jacobian are constant throughout the lifespan of my
> program. When I call ReOptimizeTNLP() at each iteration of my program, it
> is only at some iterations that MA97 displays that error message. I find
> that strange, since either I would expect that every time ReOptimizeTNLP()
> is called, MA97 should have displayed that message or never display it at
> all. If you have any insights into the matter, can you please share your
> take on why it is happening.
> 
> PS-: I have thoroughly checked my implementation, if indeed the hessian
> and jacobian I provide is changing. Nothing seems to be wrong in my code.
> 
> Thanks,
> Jayanth
> 
>> Hi,
>>
>> On 11/11/2015 04:52 AM, jmogali at andrew.cmu.edu wrote:
>>> Hi,
>>>       @Stefan -: Thanks for letting me know the ReOptimizeTNLP() option.
>>> Also warm start has significantly improved the timings.
>>>
>>> However, I observe that Hessian is being recalculated when
>>> ReOptimizeTNLP() executes. I know that this has been included for
>>> generality. I would like to however exploit the fact that throughout the
>>> lifespan of my program, the Hessian I provide remains a constant. Is
>>> there
>>> any way I can make IPOPT exploit this fact ?
>>
>> Even if you set the mentioned hessian_constant option?
>>
>>> On a side note, the MA97 solver is being used to solve the KKT matrix
>>> right ?
>>
>> Yes, I would think so.
>> To see what exactly is solved there, see (13) in the implementation
>> paper (http://www.optimization-online.org/DB_HTML/2004/03/836.html).
>>
>> Stefan
>>
>>>
>>> Thanks,
>>> Jayanth
>>>
>>>> Hi,
>>>>
>>>> if the structure of the Jacobian and Hessian do not change, have you
>>>> considered using ReOptimizeTNLP() instead of OptimizeTNLP()?
>>>> I haven't checked, but likely that will avoid calls that get the
>>>> problem
>>>> structure (n, m, sparsity pattern).
>>>>
>>>> There are also option to tell Ipopt that the Jacobian and the Hessian
>>>> are constant (jac_c_constant, jac_d_constant, hessian_constant):
>>>> http://www.coin-or.org/Ipopt/documentation/node43.html#SECTION000114070000000000000
>>>>
>>>> The TNLP is mostly implemented by you, while Ipopt only defines the
>>>> interface. From the Ipopt point-of-view, it should be save to reuse the
>>>> TNLP object.
>>>>
>>>> MA97 error -7 means that the matrix passed to MA97 is singular. See
>>>> http://www.hsl.rl.ac.uk/specs/hsl_ma97.pdf for a full documentation of
>>>> MA97.
>>>>
>>>> Stefan
>>>>
>>>> On 11/09/2015 03:35 AM, jmogali at andrew.cmu.edu wrote:
>>>>> Hi,
>>>>>        I would like to have some suggestions to speed up my code. I
>>>>> run
>>>>> my
>>>>> code in a loop where at each iteration I call the function
>>>>> OptimizeTNLP(); however I create a single TNLP object that calls
>>>>> OptimizeTNLP at each iteration. After every iteration, I change the
>>>>> objective function I am optimizing by a little.
>>>>>
>>>>> 1. The structure of the hessian is constant across iterations, is
>>>>> there
>>>>> a
>>>>> way to make IPOPT avoid calling eval_h with values=null be passed to
>>>>> my
>>>>> program at every iteration ?
>>>>> On a similar note, my Jacobian is a constant
>>>>> for all iterations, is there a way to make IPOPT avoid calling
>>>>> eval_jac_g
>>>>> , get_bounds_info for every iteration ?
>>>>>
>>>>> Please note that in the above question, by iterations I am NOT
>>>>> referring
>>>>> to those when IPOPT executes OptimizeTNLP().
>>>>>
>>>>> 2. Is it safe to reuse the TNLP object between iterations?
>>>>>
>>>>> 3. I sometimes get the message "Error return from ma97_factor. Error
>>>>> flag
>>>>> =  -7" , does this mean the KKT matrix is singular ?
>>>>>
>>>>> Thanks,
>>>>> Jayanth
>>>>>
>>>>> _______________________________________________
>>>>> Ipopt mailing list
>>>>> Ipopt at list.coin-or.org
>>>>> http://list.coin-or.org/mailman/listinfo/ipopt
>>>>>
>>>>
>>>>
>>>> --
>>>> http://www.gams.com/~stefan
>>>>
>>>
>>>
>>
>>
>> --
>> http://www.gams.com/~stefan
>>
> 
> 
> _______________________________________________
> Ipopt mailing list
> Ipopt at list.coin-or.org
> http://list.coin-or.org/mailman/listinfo/ipopt
> 



More information about the Ipopt mailing list