Again, I'm no expert, but you could always separate the equality constraint
into two inequality constraints, yes? Maybe worth a try.
On Tue, Apr 26, 2016, 10:37 PM David Morris <otha...@othalan.net> wrote:
> I've been looking through the NLOpt code for AUGLAG. If I understand
> correctly, it seems that my problem is AUGLAG will not efficiently find a
> solution which fits my constraints and so never finds a valid solution (not
> even a bad one) within time limits I have defined. It usually finds a
> solution eventually, just not within an acceptable time for our project.
> The difficulty appears to be that a constraint of "sum(x) == 100" is not a
> type of constraint which works well with the algorithm.
>
> In particular, I am looking at auglag.c lines 251 - 268 where results are
> saved, but which potentially is not called if interrupted too early by a
> timeout.
>
> Does this make sense? Any suggestions for alternative ways to write this
> constraint?
>
> David
>
> On Tue, Apr 26, 2016 at 7:57 PM, David Morris <otha...@othalan.net> wrote:
>
>> I am using NLOpt in a python application and am having problems when
>> setting an equality constraint when using LN_AUGLAG.
>>
>> If I use LN_COBYLA, optimization works perfectly. However, if I use
>> LN_AUGLAG and set LN_COBYLA as the local optimizer, the result is the exact
>> same as my initial guess. My goal is to use experiment with other
>> optimization algorithms (for example, LN_SBPLX), and use AUGLAG to provide
>> the quality constraint.
>>
>> Can anyone help determine why the equality constraint causes AUGLAG to
>> fail?
>>
>> Below is sample code showing how I use NLOpt.
>>
>>
>>
>> ###################################################################################
>> # CODE Sample:
>>
>> args = ( ... )
>> kw = { ... }
>>
>> # Optimization Function for NLOpt
>> def optfunc(x,grad):
>> # Big complex function which calculates a single return value:
>> val = model_p_mixer(x,*args,**kw)
>> return val
>>
>> # Constraint Function for NLOpt
>> # x -> percentage of total content for each component
>> # sum(x) == 100 %
>> def opt_constraint(x, grad):
>> val = float(100.0 - x.sum())
>> return val
>>
>> gopt = nlopt.opt(nlopt.LN_AUGLAG, len(guess))
>> lopt = nlopt.opt(nlopt.LN_COBYLA, len(guess))
>>
>> gopt.set_min_objective(optfunc)
>>
>> gopt.set_lower_bounds([98.62, 0.0, 0.0])
>> gopt.set_upper_bounds([99.5, 1.0, 1.0])
>>
>> gopt.add_equality_constraint(opt_constraint, 0.001)
>>
>> # Set tolerances to determine when the optimizer stops looking
>> for solutions
>> gopt.set_xtol_abs(1E-6)
>> gopt.set_ftol_abs(0.001)
>> lopt.set_xtol_abs(1E-6)
>> lopt.set_ftol_abs(0.001)
>>
>> # Set initial step size
>> gopt.set_initial_step(0.01)
>>
>> gopt.set_maxtime(8.0)
>>
>> gopt.set_local_optimizer(lopt)
>>
>> # Run the optimizer
>> mix = gopt.optimize([99.0, 0.5, 0.5])
>>
>> # Initial Guess : [99.0, 0.5 , 0.5 ]
>> # Expected Result: [99.2, 0.35, 0.45]
>> # Actual Result: [99.0, 0.5 , 0.5 ]
>>
>> ###################################################################################
>>
>>
>> Thank you,
>>
>> David
>>
> _______________________________________________
> NLopt-discuss mailing list
> NLopt-discuss@ab-initio.mit.edu
> http://ab-initio.mit.edu/cgi-bin/mailman/listinfo/nlopt-discuss
>
_______________________________________________
NLopt-discuss mailing list
NLopt-discuss@ab-initio.mit.edu
http://ab-initio.mit.edu/cgi-bin/mailman/listinfo/nlopt-discuss
