hello,
Is there a way to avoid using the Dube code for MPEC (since it requires knitro and tomlab)? would using the matlab command 'fmincon' work?
would appreciate any advice from whoever has tried to use the MPEC method
Using fmincon to program MPEC for BLP
-
specify the objective function as the GMM objective function, combined with nonlinear equality constraints (predicted market shares = observed market shares) with analytic gradients. The latter is specified in the nonlinear constraint option of fmincon.
beautiful, thanks, ill try this.
ill also try nlopt
-
You can basically use the dube code and replace the ktrlink call with fmincon. There are a few other tweaks you may need to do, but for that simple problem, you should get the same results. It's been several years since I did this, but I did get it to work and it wasn't too painful.
-
You can basically use the dube code and replace the ktrlink call with fmincon. There are a few other tweaks you may need to do, but for that simple problem, you should get the same results. It's been several years since I did this, but I did get it to work and it wasn't too painful.
i tried doing this and it was more complicated than i anticipated. i can check again
-
I haven't see this particular code, but my experience is that fmincon for bilevel programs requires regularization of the lower level problem by replacing exact optimality with instead epsilon optimality in the lower level problem. The complimentary slackness in particular needs to be revularized. In general the regularization leads to improved numerical performance because of some technical issues involving optimality conditions of the overall problem. There is a dedicated literature in optimization on this topic.
-
I haven't see this particular code, but my experience is that fmincon for bilevel programs requires regularization of the lower level problem by replacing exact optimality with instead epsilon optimality in the lower level problem. The complimentary slackness in particular needs to be revularized. In general the regularization leads to improved numerical performance because of some technical issues involving optimality conditions of the overall problem. There is a dedicated literature in optimization on this topic.
I haven't studied this but it sounds unnecessary to me.
mpec works perfectly with fmincon if you've got your derivatives and sparsity game right.