[R-sig-hpc] Optimizer Question

Discussion:

Bromaghin, Jeffrey

2017-08-04 20:21:47 UTC

Dear r-sig-hpc members,

I have developed a new method of interest to ecologists that involves
solving a large optimization problem. One example I used in a recent paper
had over 5,700 parameters. I have been using Matlab and a optimization
library called Tomlab, which works quite well. However, I would like to
incorporate this new method into an R package, but question whether the
optimizers available in R can handle a problem of this size efficiently.
Does anyone have experience solving such large problems with any of the
optimizers available in R and, if so, what optimizer(s) would you recommend
I try?

Thank you,
Jeff
-----------------------------------------------
Jeffrey F. Bromaghin, PhD
Research Statistician
USGS Alaska Science Center
Marine Ecosystems Office
4210 University Drive
Anchorage, AK 99508
907-786-7086
***@usgs.gov
*http://alaska.usgs.gov/science/biology/quantitative_ecology/index.php
<http://alaska.usgs.gov/science/biology/quantitative_ecology/index.php>*

[[alternative HTML version deleted]]

Cristian Bologa

2017-08-04 20:31:44 UTC

Permalink

Hi Jeff,

For non-linear optimizations, I have found nloptr to be pretty good.

Cristian

Cristian Bologa, Ph.D.
Research Professor,
Div. of Translational Informatics,
Dept. of Internal Medicine,
Univ. of New Mexico, School of Medicine,
Innovation Discovery&Training Center, MSC09 5025,
700 Camino de Salud NE, Albuquerque, NM 87131
Telephone: +1 (505) 925-7534
Fax:+1 (505) 925-7625
--------------------------
"People don't want more information, they want insights"

-----Original Message-----
From: R-sig-hpc [mailto:r-sig-hpc-***@r-project.org] On Behalf Of Bromaghin, Jeffrey
Sent: Friday, August 04, 2017 2:22 PM
To: r-sig-***@r-project.org
Subject: [R-sig-hpc] Optimizer Question

Dear r-sig-hpc members,

I have developed a new method of interest to ecologists that involves solving a large optimization problem. One example I used in a recent paper had over 5,700 parameters. I have been using Matlab and a optimization library called Tomlab, which works quite well. However, I would like to incorporate this new method into an R package, but question whether the optimizers available in R can handle a problem of this size efficiently.
Does anyone have experience solving such large problems with any of the optimizers available in R and, if so, what optimizer(s) would you recommend I try?

Thank you,
Jeff
-----------------------------------------------
Jeffrey F. Bromaghin, PhD
Research Statistician
USGS Alaska Science Center
Marine Ecosystems Office
4210 University Drive
Anchorage, AK 99508
907-786-7086
***@usgs.gov
*http://alaska.usgs.gov/science/biology/quantitative_ecology/index.php
<http://alaska.usgs.gov/science/biology/quantitative_ecology/index.php>*

[[alternative HTML version deleted]]

Roger Koenker

2017-08-04 20:35:04 UTC

Permalink

You don’t say anything about the nature of your problem domain beyond its size,
but for convex problems Mosek is a good option, and there is an R interface
called Rmosek that is quite convenient.

url: www.econ.uiuc.edu/~roger Roger Koenker
email ***@uiuc.edu Department of Economics
vox: 217-333-4558 University of Illinois
fax: 217-244-6678 Urbana, IL 61801

Post by Bromaghin, Jeffrey
Dear r-sig-hpc members,
I have developed a new method of interest to ecologists that involves
solving a large optimization problem. One example I used in a recent paper
had over 5,700 parameters. I have been using Matlab and a optimization
library called Tomlab, which works quite well. However, I would like to
incorporate this new method into an R package, but question whether the
optimizers available in R can handle a problem of this size efficiently.
Does anyone have experience solving such large problems with any of the
optimizers available in R and, if so, what optimizer(s) would you recommend
I try?
Thank you,
Jeff
-----------------------------------------------
Jeffrey F. Bromaghin, PhD
Research Statistician
USGS Alaska Science Center
Marine Ecosystems Office
4210 University Drive
Anchorage, AK 99508
907-786-7086
*http://alaska.usgs.gov/science/biology/quantitative_ecology/index.php
<http://alaska.usgs.gov/science/biology/quantitative_ecology/index.php>*
[[alternative HTML version deleted]]
_______________________________________________
R-sig-hpc mailing list
https://stat.ethz.ch/mailman/listinfo/r-sig-hpc

Bromaghin, Jeffrey

2017-08-04 20:45:58 UTC

Permalink

My apologies! The structure is close to linear, but there is a scaling so
that modeled proportions sum to 1.0, so the problem is nonlinear. There
are quite a few linear constraints as all parameters are non-negative and
subsets of the
parameters must sum to 1.

Best regards,
Jeff

-----------------------------------------------
Jeffrey F. Bromaghin, PhD
Research Statistician
USGS Alaska Science Center
Marine Ecosystems Office
4210 University Drive
Anchorage, AK 99508
907-786-7086
***@usgs.gov
*http://alaska.usgs.gov/science/biology/quantitative_ecology/index.php
<http://alaska.usgs.gov/science/biology/quantitative_ecology/index.php>*

Post by Roger Koenker
You don’t say anything about the nature of your problem domain beyond its size,
but for convex problems Mosek is a good option, and there is an R interface
called Rmosek that is quite convenient.
url: www.econ.uiuc.edu/~roger Roger Koenker
vox: 217-333-4558 University of Illinois
fax: 217-244-6678 Urbana, IL 61801

Post by Bromaghin, Jeffrey
Dear r-sig-hpc members,
I have developed a new method of interest to ecologists that involves
solving a large optimization problem. One example I used in a recent

paper

Post by Bromaghin, Jeffrey
had over 5,700 parameters. I have been using Matlab and a optimization
library called Tomlab, which works quite well. However, I would like to
incorporate this new method into an R package, but question whether the
optimizers available in R can handle a problem of this size efficiently.
Does anyone have experience solving such large problems with any of the
optimizers available in R and, if so, what optimizer(s) would you

recommend

Post by Bromaghin, Jeffrey
I try?
Thank you,
Jeff
-----------------------------------------------
Jeffrey F. Bromaghin, PhD
Research Statistician
USGS Alaska Science Center
Marine Ecosystems Office
4210 University Drive
Anchorage, AK 99508
907-786-7086
*http://alaska.usgs.gov/science/biology/quantitative_ecology/index.php
<http://alaska.usgs.gov/science/biology/quantitative_ecology/index.php>*
[[alternative HTML version deleted]]
_______________________________________________
R-sig-hpc mailing list
https://stat.ethz.ch/mailman/listinfo/r-sig-hpc

[[alternative HTML version deleted]]

Hesen Peng

2017-08-05 21:12:20 UTC

Permalink

I apologize since this answer is kind of off topic. If you look beyond the
boundary, TensorFlow and Theano have been very popular these days in the
deep learning field and they can be used to solve optimization problems.
They are being used to solve problems involving millions of parameters. And
I do thing ecosystem-wise they are better than using R packages.

Post by Bromaghin, Jeffrey
My apologies! The structure is close to linear, but there is a scaling so
that modeled proportions sum to 1.0, so the problem is nonlinear. There
are quite a few linear constraints as all parameters are non-negative and
subsets of the
parameters must sum to 1.
Best regards,
Jeff
-----------------------------------------------
Jeffrey F. Bromaghin, PhD
Research Statistician
USGS Alaska Science Center
Marine Ecosystems Office
4210 University Drive
Anchorage, AK 99508
907-786-7086
*http://alaska.usgs.gov/science/biology/quantitative_ecology/index.php
<http://alaska.usgs.gov/science/biology/quantitative_ecology/index.php>*

Post by Roger Koenker
You don’t say anything about the nature of your problem domain beyond its size,
but for convex problems Mosek is a good option, and there is an R

interface

Post by Roger Koenker
called Rmosek that is quite convenient.
url: www.econ.uiuc.edu/~roger Roger Koenker
vox: 217-333-4558 University of Illinois
fax: 217-244-6678 Urbana, IL 61801

Post by Bromaghin, Jeffrey
Dear r-sig-hpc members,
I have developed a new method of interest to ecologists that involves
solving a large optimization problem. One example I used in a recent

paper

Post by Bromaghin, Jeffrey
had over 5,700 parameters. I have been using Matlab and a optimization
library called Tomlab, which works quite well. However, I would like

Post by Roger Koenker

Post by Bromaghin, Jeffrey
incorporate this new method into an R package, but question whether the
optimizers available in R can handle a problem of this size

efficiently.

Post by Roger Koenker

Post by Bromaghin, Jeffrey
Does anyone have experience solving such large problems with any of the
optimizers available in R and, if so, what optimizer(s) would you

recommend

Post by Bromaghin, Jeffrey
[[alternative HTML version deleted]]
_______________________________________________
R-sig-hpc mailing list
https://stat.ethz.ch/mailman/listinfo/r-sig-hpc

[[alternative HTML version deleted]]
_______________________________________________
R-sig-hpc mailing list
https://stat.ethz.ch/mailman/listinfo/r-sig-hpc

--
Hesen Peng 彭河森

[[alternative HTML version deleted]]

Suzen, Mehmet

2017-08-05 21:42:19 UTC

Permalink

Hi Hesen,

I am not sure if the premise "..ecosystem-wise they are better than
using R packages..". First of all almost all of the optimisation
techniques as far as I know in (deep) machine learning are the
gradient base. R provides superb ecosystem for gradient free
optimisation and other constraint based optimisations where Keras or
tensorflow solves. Yes, they scale better in Gradient decent base
algorithms but they don't cover the spectrum R ecosystem provides, at
least not as of today. Have a look at the spectrum of optimisations R
ecosystem provides:
https://cran.r-project.org/web/views/Optimization.html
It is pretty diverse.

Cheers,
-m

Post by Hesen Peng
I apologize since this answer is kind of off topic. If you look beyond the
boundary, TensorFlow and Theano have been very popular these days in the
deep learning field and they can be used to solve optimization problems.
They are being used to solve problems involving millions of parameters. And
I do thing ecosystem-wise they are better than using R packages.

Post by Roger Koenker
You don’t say anything about the nature of your problem domain beyond its size,
but for convex problems Mosek is a good option, and there is an R

interface

Post by Roger Koenker
called Rmosek that is quite convenient.
url: www.econ.uiuc.edu/~roger Roger Koenker
vox: 217-333-4558 University of Illinois
fax: 217-244-6678 Urbana, IL 61801

Post by Bromaghin, Jeffrey
Dear r-sig-hpc members,
I have developed a new method of interest to ecologists that involves
solving a large optimization problem. One example I used in a recent

paper

Post by Bromaghin, Jeffrey
had over 5,700 parameters. I have been using Matlab and a optimization
library called Tomlab, which works quite well. However, I would like

Post by Roger Koenker

Post by Bromaghin, Jeffrey
incorporate this new method into an R package, but question whether the
optimizers available in R can handle a problem of this size

efficiently.

Post by Roger Koenker

Post by Bromaghin, Jeffrey
Does anyone have experience solving such large problems with any of the
optimizers available in R and, if so, what optimizer(s) would you

recommend

Post by Bromaghin, Jeffrey
[[alternative HTML version deleted]]
_______________________________________________
R-sig-hpc mailing list
https://stat.ethz.ch/mailman/listinfo/r-sig-hpc

[[alternative HTML version deleted]]
_______________________________________________
R-sig-hpc mailing list
https://stat.ethz.ch/mailman/listinfo/r-sig-hpc

--
Hesen Peng 彭河森
[[alternative HTML version deleted]]
_______________________________________________
R-sig-hpc mailing list
https://stat.ethz.ch/mailman/listinfo/r-sig-hpc