Question about the effect of coefficient size on the converging speed of a MINLP model

Archive of Gamsworld Google Group
Post Reply
Archiver
User
User
Posts: 7876
Joined: 7 years ago

Question about the effect of coefficient size on the converging speed of a MINLP model

Post by Archiver »


Hello everybody,

I'm currently working on a MINLP with 1366 variables. The objective function is in the form shown below.
Supposedly, gamma(0) was to be a value of 5e4 and gamma(1) to be 1e6.
But during the process I made a mistake and put in the value of gamma(1) as 696, and ran the solver(ANTIGONE).
It returned an optimal solution in reasonable time, but I realized the mistake and changed the value back to 1e6, and ran the solver again, but this time it wouldn't converge, even after many hours.
I scaled the equations and the variables as well as the gradients, but it won't solve the problem, retaining about 10% of difference btw the lower bound and upper bound.
I'm very confused by this phenomenon since the only difference between the two input models is the value of gamma(1). Is it possible that the value of one coefficient can alter the convergence of a model so drastically?

I'm having a hard time figuring this out and would be very glad if anyone can give any sort of advice with this problem. Please tell me if you need more specifications about the problem
Thank you in advance!

Sincerely,
Changsoo Kim

--
Attachments

[The extension has been deactivated and can no longer be displayed.]

Archiver
User
User
Posts: 7876
Joined: 7 years ago

Re: Question about the effect of coefficient size on the converging speed of a MINLP model

Post by Archiver »


The short answer is yes and why not? Google for performance variability in MILP and get enlighten.

Is it a quadratic MIP btw? It is a bit hard to see what is parameters and what is variables.

Den mandag den 7. marts 2016 kl. 17.09.34 UTC+1 skrev Changsoo Kim:

Hello everybody,

I'm currently working on a MINLP with 1366 variables. The objective function is in the form shown below.
Supposedly, gamma(0) was to be a value of 5e4 and gamma(1) to be 1e6.
But during the process I made a mistake and put in the value of gamma(1) as 696, and ran the solver(ANTIGONE).
It returned an optimal solution in reasonable time, but I realized the mistake and changed the value back to 1e6, and ran the solver again, but this time it wouldn't converge, even after many hours.
I scaled the equations and the variables as well as the gradients, but it won't solve the problem, retaining about 10% of difference btw the lower bound and upper bound.
I'm very confused by this phenomenon since the only difference between the two input models is the value of gamma(1). Is it possible that the value of one coefficient can alter the convergence of a model so drastically?

I'm having a hard time figuring this out and would be very glad if anyone can give any sort of advice with this problem. Please tell me if you need more specifications about the problem
Thank you in advance!

Sincerely,
Changsoo Kim

--
To unsubscribe from this group and stop receiving emails from it, send an email to gamsworld+unsubscribe@googlegroups.com.
To post to this group, send email to gamsworld@googlegroups.com.
Visit this group at https://groups.google.com/group/gamsworld.
For more options, visit https://groups.google.com/d/optout.
Archiver
User
User
Posts: 7876
Joined: 7 years ago

Re: Question about the effect of coefficient size on the converging speed of a MINLP model

Post by Archiver »


Thank you for the reply I'm looking up the subject.
Yes it is quadratic. g(i,t) is a binary variable, a(i,t) and d(p,t) continuous variables. There should be a y(p,t) multiplied to gamma 2 which is a binary variable.
Sorry for the confusion!

Do you happen to know a way through this problem then? other than scaling the variable d(p,t)

2016년 3월 8일 화요일 오후 7시 10분 27초 UTC+9, Erling D. Andersen 님의 말:

The short answer is yes and why not? Google for performance variability in MILP and get enlighten.

Is it a quadratic MIP btw? It is a bit hard to see what is parameters and what is variables.

Den mandag den 7. marts 2016 kl. 17.09.34 UTC+1 skrev Changsoo Kim:

Hello everybody,

I'm currently working on a MINLP with 1366 variables. The objective function is in the form shown below.
Supposedly, gamma(0) was to be a value of 5e4 and gamma(1) to be 1e6.
But during the process I made a mistake and put in the value of gamma(1) as 696, and ran the solver(ANTIGONE).
It returned an optimal solution in reasonable time, but I realized the mistake and changed the value back to 1e6, and ran the solver again, but this time it wouldn't converge, even after many hours.
I scaled the equations and the variables as well as the gradients, but it won't solve the problem, retaining about 10% of difference btw the lower bound and upper bound.
I'm very confused by this phenomenon since the only difference between the two input models is the value of gamma(1). Is it possible that the value of one coefficient can alter the convergence of a model so drastically?

I'm having a hard time figuring this out and would be very glad if anyone can give any sort of advice with this problem. Please tell me if you need more specifications about the problem
Thank you in advance!

Sincerely,
Changsoo Kim

--
To unsubscribe from this group and stop receiving emails from it, send an email to gamsworld+unsubscribe@googlegroups.com.
To post to this group, send email to gamsworld@googlegroups.com.
Visit this group at https://groups.google.com/group/gamsworld.
For more options, visit https://groups.google.com/d/optout.
Archiver
User
User
Posts: 7876
Joined: 7 years ago

Re: Question about the effect of coefficient size on the converging speed of a MINLP model

Post by Archiver »


If

L_p>=0, gamma_0>=0

then the objective is convex quadratic objective. Assuming the constraints are linear then you could use one optimizer for mixed integer convex quadratic optimization e.g. MOSEK.
There is a good chance that they would do better than general purpose MINLP code like ANTIGONE. No promises though.

Den onsdag den 9. marts 2016 kl. 08.21.11 UTC+1 skrev Changsoo Kim:

Thank you for the reply I'm looking up the subject.
Yes it is quadratic. g(i,t) is a binary variable, a(i,t) and d(p,t) continuous variables. There should be a y(p,t) multiplied to gamma 2 which is a binary variable.
Sorry for the confusion!

Do you happen to know a way through this problem then? other than scaling the variable d(p,t)

2016년 3월 8일 화요일 오후 7시 10분 27초 UTC+9, Erling D. Andersen 님의 말:

The short answer is yes and why not? Google for performance variability in MILP and get enlighten.

Is it a quadratic MIP btw? It is a bit hard to see what is parameters and what is variables.

Den mandag den 7. marts 2016 kl. 17.09.34 UTC+1 skrev Changsoo Kim:

Hello everybody,

I'm currently working on a MINLP with 1366 variables. The objective function is in the form shown below.
Supposedly, gamma(0) was to be a value of 5e4 and gamma(1) to be 1e6.
But during the process I made a mistake and put in the value of gamma(1) as 696, and ran the solver(ANTIGONE).
It returned an optimal solution in reasonable time, but I realized the mistake and changed the value back to 1e6, and ran the solver again, but this time it wouldn't converge, even after many hours.
I scaled the equations and the variables as well as the gradients, but it won't solve the problem, retaining about 10% of difference btw the lower bound and upper bound.
I'm very confused by this phenomenon since the only difference between the two input models is the value of gamma(1). Is it possible that the value of one coefficient can alter the convergence of a model so drastically?

I'm having a hard time figuring this out and would be very glad if anyone can give any sort of advice with this problem. Please tell me if you need more specifications about the problem
Thank you in advance!

Sincerely,
Changsoo Kim

--
To unsubscribe from this group and stop receiving emails from it, send an email to gamsworld+unsubscribe@googlegroups.com.
To post to this group, send email to gamsworld@googlegroups.com.
Visit this group at https://groups.google.com/group/gamsworld.
For more options, visit https://groups.google.com/d/optout.
Post Reply