The current version of ipopt is 3.10.1, and the current version of pyipopt is 0.8. The author of openopt Dmitrey claims that recent versions of ipopt does not work well together with openopt, and he uses version 3.8.3. Anyway, the focus of this post is to build ipopt and pyipopt. I use ubuntu 11.10 within a vmware virtual machine. There is a 3.9 version ipopt in ubuntu repository, however that one does not work well when compiling pyipopt, which motivated me to build by myself.

The document of ipopt is fabulous. However, there are so many steps that you may easily miss some. This post is by no means a replacement of the original document. So before you have any questions, make sure you have read the original document carefully.

Tip 1, remove existing ipopt

I had ipopt 3.9.1 from the ubuntu repository installed, but I did not check that. So I had everything built successfully, but bunches of errors when trying to import pyipopt in python. Thus, remember to uninstall existing ipopt before you start.

Tip 2, BLAS

ubuntu has a pre-compiled BLAS library, or if you have MKL, you may have your own BLAS. So I thought I could supply these library without downloading the source of BLAS. Everything worked well at ipopt side. However, when it comes to pyipopt, you will find yourself missing all kinds of symbols. So, even you have BLAS installed, do not supply it use the –with-BLAS options. Remember to download BLAS and LAPACK source. Refer to the Thirdparty directory of the ipopt source package for detailed information. By the end, you will get libcoinblas.so, libcoinlapack.so, which are useful to build pyipopt. If you don’t do this, very possibly you will get an error when importing pyipopt stating certain symbols are not defined.

Tip 3, HSL

Well, this maybe not a tip, but a work around.  The files I downloaded from HSL website have different names as described in ipopt document. Thus, in order to use MA57 and MC19, I have to build them separately and supply them as pre-compiled library using the –with-hsl option. Read the help file came with MA57 to detailed information of compiling them.

Tip 4, Modify the configuration files

You actually do not need to modify anything for the ipopt configuration file. However, for pyipopt, make sure you modify the makefile and setup.py file to reflect the environment of your own machine.

My first optimization related project dates back 5 years ago, when I needed to improve some models built by some other senior students. Those models were written in AMPL, using cplex as a solver. My preferred computation tool was Matlab at that time, but cplex did not support Matlab yet. What I did was some thing really dirty. I wrote script to generate AMPL data files in Matlab,  parse the screen print out of AMPL and then send back to Matlab. Until recently, I got to know that TOMLAB provides an interface between AMPL and Matlab. Actually, if you use Matlab a lot, there are more interfaces provided by TOMLAB.

Later I need to use column generation to solve a problem, and by then I started to use the C++/Concert/Cplex tool chain. Honestly, it is more time-consuming to learn this, and less efficient to write models. However, it does offer more control to the solve engine. I would be satisfied with this solution, until Gurobi came out. Then I think, maybe I want to switch among solvers to see which one fits my problem best. However, if you have your model written with C++ and concert, basically it means to rewrite the model again with Gurobi API again if you really want to use Gurobi. At this time, I would really appreciate AMPL, which has interface to many mainstream solvers.

The story did not stop here. Later on, I started a new project, where optimization was just part of the problem, statistical modeling was another part. At that time, I suddenly got some interests on Python, and switched almost everything to Python and R. When cplex 12 came out, it started to provide a Python API. Same problem, I wanted to try different solvers,  but there is no interface between Python and AMPL. This time, I tried to have a better solution rather than parsing screen output. I heard about Pyomo in INFORMS southwest regional meeting. I liked this idea very much, but there was not much document available. After googling around, I found a similar tool, pulp, which gives similar functionality as AMPL, but within Python. I finished one project with pulp, and it did quite good job, except that it does not support 64 bit cplex, either in Windows or Linux. Pulp has relatively better documentation than Pyomo, but most of the time you need to reply on their examples to figure out how to use it.

Finally, I would suggest to put a middle layer between modeling and the actual solve engine for two reasons:

• A modeling language usually provide better readability of the models and easier to writer models.
• A modeling middle layer enables you to switch among different solvers much easier.

AMPL is good, but it is really hard to interact with other languages. IBM OPL is another choice, but it only supports solvers from IBM, i.e. cplex and cp. In python, you could choose pulp or Pyomo. There is also a package called optimJ, which provides same functionality as pulp, but within JAVA. I haven’t tried this out yet, but it looks promising.

As a student, pricing of the tools also concerns me. For your information, I made the following list

AMPL: usually not free, may be available from your department if you are in an IE or OR department

cplex studio: free for academia use. Includes OPL, cplex, cp

Gurobi: free for academia use

TOMLAB: commercial

optimJ: The company is close, the products are free~~~

pulp: MIT license

Pyomo: BSD license

There are some open source solvers I did not mention in this post just because I did not use them. For example, lp_solver, cvxopt. Non linear solvers are not covered as well. For a complete guide of solver choice, I suggest you go to Dr. Mittelmann’s website: http://plato.asu.edu/guide.html

STEP 1 SET UP THE OPERATING SYSTEM

STEP 2 INSTALL RAPACHE

apt-get install r-base-dev apache2-mpm-prefork apache2-prefork-dev
wget http://biostat.mc.vanderbilt.edu/rapache/files/rapache-latest.tar.gz
rapachedir=`tar tzf rapache-latest.tar.gz | head -1`
tar xzvf rapache-latest.tar.gz
cd $rapachedir
./configure
make
sudo make install

STEP 3 SET UP R

install.packages() 

 sudo R

.libPaths()
[1] "/usr/local/lib/R/site-library" "/usr/lib/R/site-library"
[3] "/usr/lib/R/library"

install.pacakages(‘qcc’,lib="/usr/local/lib/R/site-library")

STEP 4 CONFIGURE A RAPACHE ENABLED SITE

<VirtualHost *:80>
	ServerAdmin LJ@localhost
	LoadModule R_module /usr/lib/apache2/modules/mod_R.so
	# Place this in the config file to turn error messages into HTML which
	# are printed in the browser. Without this, all warning and error messages
	# are printed to the Apache error log file.
	ROutputErrors 

	DocumentRoot "/home/ilpil/www"
	REvalOnStartup "setwd('/home/ilpil/www/') "
	RSourceOnStartup "./R/startup.r"
	<Directory />
		Options FollowSymLinks
		AllowOverride None
		# Order allow,deny
		# Allow from all
	</Directory>
	<Location /RApacheInfo>
		SetHandler r-info
	</Location>

	<Location /hello>
		SetHandler r-handler
		RFileHandler /home/ilpil/www/helloworld.r
	</Location>

	<Location /sendBin>
		SetHandler r-handler
		RFileHandler /home/ilpil/www/sendBin.r
	</Location>

	<Directory /home/ilpil/www/brew>
		SetHandler r-script
		RHandler brew::brew
	</Directory>

	<Location /R>
		SetHandler r-handler
		RFileHandler /home/ilpil/www/test.R
	</Location>

	<Location /Rblah>
		SetHandler r-handler
		RFileHandler /home/ilpil/www/test.R::main
	</Location>

	# Prints out a nice report about R running within Apache

	<Directory />
		Options FollowSymLinks
		AllowOverride None
	</Directory>

	<Directory /home/ilpil/www/>
		Options Indexes FollowSymLinks MultiViews
		AllowOverride None
		Order allow,deny
		allow from all
	</Directory>

	ScriptAlias /cgi-bin/ /usr/lib/cgi-bin/
	<Directory "/usr/lib/cgi-bin">
		AllowOverride None
		Options +ExecCGI -MultiViews +SymLinksIfOwnerMatch
		Order allow,deny
		Allow from all
	</Directory>

	ErrorLog ${APACHE_LOG_DIR}/error.log
	# Possible values include: debug, info, notice, warn, error, crit,
	# alert, emerg.
	LogLevel warn 

	CustomLog ${APACHE_LOG_DIR}/access.log combined 

	Alias /doc/ "/usr/share/doc/"
	<Directory "/usr/share/doc/">
		Options Indexes MultiViews FollowSymLinks
		AllowOverride None
		Order deny,allow
		Deny from all
		Allow from 127.0.0.0/255.0.0.0 ::1/128
	</Directory>
</VirtualHost>

sudo a2dissite default
sudo a2ensite Rapache
sudo /etc/init.d/apache2 reload

STEP 5 THE USER2007 EXAMPLE

<div id="plot"><% brew('test/brew/useR2007plot.rhtml') %> </div>

<div id="plot"><% brew('/brew/useR2007plot.rhtml') %></div>

LAST CAUTION

Something Unrelated:

I used to use AMPL as my modeling language and solve the problem with CPLEX. However, later I realized that AMPL is just a modeling language, but not a generic programming language, which makes it incapable to handle my problem since I have to customize some parts of the algorithm. Then I started to use CONCERT from ILOG. At the same time, a “fancy friend” introduced me to use Python. I do not know for what reason I switched to LINUX, maybe because of the powerful command line. I got cplex set up, I got ampl set up and I got gcc set up, and then I need an IDE. I chose the popular one, eclipse. However, annoying things happened, I cannot compile my programs inside the IDE, while I can compile it with a self written make file. I need IDE because I want to make the debug easier. During that time, I need to reboot my computer many times to switch between windows and Linux, (I was thinking to use a virtual machine, but I did not do that). Fortunately, I got eclipse to work with cplex. It maybe not the best solution, but good enough to me, a newbie of Linux, programming, eclipse and cplex. There was another story of installing Linux in my thinkpad T60, and make everything run smoothly.

Something Related:

OK, first thing is to install eclipse and CDT. There are lots of tutorials on this, so I will not cover this. Suppose you have CPLEX installed in the default directory: /usr/ilog/ . The setting is really nothing but putting the correct path into the correct place. As an old Chinese saying goes, a picture is better than thousands of words, I will put the screen shots below.

@@update June 5, 2010:

I wrote this post two years ago, and I was using cplex 10.0. I don’t have a chance to try the new versions yet.

@@ update Nov 23,2010:

Thanks Yuan provided the setting for CPLEX 12

http://www4.ncsu.edu/~yzhang13/eclipse-cplex-linux2.html

BTW: If you happened to read this, would you mind to know that I am looking for a analytic job? Here is my linkedin profile: http://www.linkedin.com/in/ljxue 

update: April 1, 2009

I found a lot people googled and got here. The above mentioned method doesn’t work anymore since I upgrade to ubuntu 8.04. Also, I found changing your wireless card  may help. I did nothing to the software, but changed my wireless card, and it is working smoothly now.  BTW, I am using the driver from the repository now.

Matlab provides many GUIes to do many operations. If you do not have many file to import, just use those tools. Also those tools can generate code to do similar work with slightly modification.

Important: I only put the most common usage here, please refer to the matlab help file for detail usage.

1. write & read EXCEL file.read:
   num=xlsread(filename,sheetname,range)e.g.:
   num=xlsread('datafile.xls','sheet3','B2:F11')

   read from file datafile.xls sheet3, cell range from B2 to F11, and assign the value to num

   PS: this style only read numeric data, if you have text data, the left side should change to be [num,txtdata]

   PS2:only num=xlsread(filename)will read from the first sheet range from the up left cell

   Write:
   xlswrite(filename,data,sheetname,range)

   e.g.:
   xlswrite('outputfile.xls',data,'sheet1','B2')

   PS: range here only means the start cell. others are the same to read

2. wav file (Thanks to Starfish)read:
   [y,Fs,bits] = wavread('filename')
e.g.:
   [data,Fs,bits]=wavread('sample.wav');

   This command returns the sample rate (Fs) in Hertz and the number of bits per sample (bits) used to encode the data in the file

   Write:
   wavwrite(y,Fs,'filename')
e.g.:
   wavwrite(firmat,Fs,'samplefilted.wav');

   This command write the data of firmat into ‘samplefilted.wav’ with Fs frequency.
   My current work does not need to read wav file. I remember I used this when I was taking the DSP class.

3. txt files
4. 1. Read ASCII-delimited file of numeric data into matrixread:
      M = dlmread(filename, delimiter)
e.g.:
      M = dlmread('data.txt', ' ');

      This command read the numeric data from data.txt, the delimiter is space(‘ ‘). If delimiter is omitted, comma (,) is the default delimiter. Reading a specific range is also possible, refer to the matlab help file for detail.

      Write:
      dlmwrite(filename, M, 'D')
e.g.:
      dlmwrite('output.txt', M, ' ')

      Write the data M into file output.txt, delimitted by space(‘ ‘). A comma(,) is the default delimiter. Writing range is also available, refer to the matlab help file.

   2. ordinary txt fileto be continued.

To be continued. I will write others in the next few days. I am sleepy these days~~~

I will only write how to read txt file in the future. I won’t write how to read image file and sound file.

PS: A little update can be found here

http://planetmath.org/encyclopedia/FleurysAlgorithm.html

http://www.ctl.ua.edu/math103/euler/ifagraph.htm

I prefer the first one because there is an example.

Fleury’s Algorithm is used to find the Eulerian Cycle (if there is one) or the Eulerian Path (if there is no Eulerian Cycle).

The computation complexity is O(n^4). There is another O(n+e) algorithm which uses the cycles.