Call Bonmin.

[UniMilky]

Hi, Johan,

           I have a procedure attached below, and when i use the Bonmin solver, it will report this :

Error using bonmin

*** Error using Bonmin Matlab interface: ***

You have specified a nonexistent BONMIN/IPOPT option

("pardiso_redo_symbolic_fact_only_if_inertia_wrong").

Error in callbonmin (line 126)

[xout,info] = bonmin(model.x0,funcs,options);

Error in solvesdp (line 350)

    eval(['output = ' solver.call '(interfacedata);']);

Error in optimize (line 31)

[varargout{1:nargout}] = solvesdp(varargin{:});

Error in GBDnewceil (line 109)

    result = optimize(iCons,Objceil(x),options)

The procedure uses ceil and max function at first , and when it reports errors,i remove these constrains, but the erros still exits.

So i want to know where is wrong.

Besides, the yalmip will auto convert ceil and max function model to a mixed integer model, how can i look at the new model?
Any help would be appreciated.

[Johan Löfberg]

I cannot reproduce that, it hangs though (probably as the current problem is too simple). YALMIP does not manually create any options but simply creates them as  ops.bonmin= bonminset([],'noIpopt'), so if is specified in there, but bonmin doesn't like it, that sounds like a bug in opti toolbox.

Weirdly complicated code though. You can simply do Cons = [-x(:) <= 0, x(:) <= 1]  etc

[Johan Löfberg]

and Objceil = -sum(x(:))

[Johan Löfberg]

and using globals is really dangerous programming pattern. Simply use for i = 1 : size(x,1) etc, if you want to loop on variables

[Johan Löfberg]

and what is you really try to model? your commented use of ceil smells like some odd approach to implement some pretty integrality conditions

[UniMilky]

Many thanks. 

The skills is very useful, make the code greatly simplified.

[UniMilky]

My model is as follows.

I have transfered it to a MILP model , and now i want to directly use yalmip to solve this model to see if the result is less better than the MILP model. 

And i know yalmip will auto convert ceil and max function model to a mixed integer model, how can i look at the new model?

Besides , is there any solver could solve models that contains ceil and max function?

[Johan Löfberg]

This cannot be transfered to a MILP model.

And if it was a MILP, you would not use bonmin, as bonmin is for convex imixed-integer programs

[Johan Löfberg]

It is possible to linearize the model though, as you essentially have nonlinearities in the form of continuous*binary

To begin, note that ceil(alpha) can be handled easily by introdung a new binvar beta with beta >= alpha.

Then, alpha/(w/sum(ceil(alpha))) is alpha*sum(ceil(alpha))/w which is alpha*sum(beta)/w. alpha*beta is continuous*binary, and thus linearizable. Either you do that manually or you use binmodel.

Constraint involving max(beta)  is non-convex, but is MILP-represented automatically by YALMIP

[UniMilky]

Yeah, as you said , i introduce a binvar beta to reprent the ceil(x), but we need too make sure it satify following constrains:

        if alpha > 0 , then beta = 1;

        if beta = 1, then alpha > 0;

        if beta = 0, then alpha = 0;

        if alpha = 0, then beta = 0;

and the four conditions can be written as :

        alpha <= beta;

        beta < alpha + 1;

Then , the max can be rewritten as:

        beta <= sum(alpha) <= 1;

Last , i use the Mccormick Envelope to linerize the alpha*beta and use intlinprog to solve this , it was too slowly.

Does there any better choice?

Besides, i just googled 'binmode', it's really useful.

Thanks a lot for your help.

 

[Johan Löfberg]

No reason to outer bound w=x*y when y is binary.

x*y is simply implies(y, w == x) + implies(~y, w == x) which has big-M model [-M(1-y) <= w-x <= M*(1-y), -M*y <= x <= M*y]

[Johan Löfberg]

and that is precisely the model used to represent ceil in YALMIP, which you see if you open sdpvar/ceil, but better to do it manually though as you then have full control of the fact that the result is a binary variable, which you then can use to easily linearize expressions