Requirements
This documentation for installing the code assumes a unix-like environment. Specifically, it assumes that the following programs exist (which they do for most unix systems, even if you have never heard of them before.)
- uudecode (if you got the codevia email)
- gunzip
- bash
- tar
- ar
- gmake (regular 'make' will not work without mods to Makefile)
- gcc (any ANSI C compiler will probably work though)
- flex ('lex' can be substituted with some minor mods)
- bison ('yacc' can be substituted with some minor mods)
Also, you can probably get this to compile with any ANSI C compiler, though you might have to muck with the Makefile a bit. There should be enough comments in the Makefile to help you with this process if you need it.
Installing
There are two ways that you might have received this code:- via email
- via web
The code will be contained in an email message in uuencoded form at the very end of the email. The uuencoded part looks something like this:
begin 644 pomdp-solve-4.0.tgz
M'XL(`-A9;S8``^Q;:WOB1I;.U]:OJ,3)8F8`W[KMB2?)+HUQ-QD,+.!T>F?R
M.$(4H%BH6)5D-_FPOWW?<ZI*$K8[B0-/YDO(D[:DJCIUZES?NJW4<KJJ:Q7=
MR8-6?_!^V'GS=OS);G]'AX>G+U^*3X009Z?F[Y%]I]\QBH4X??7J].STY.SL
...
<More gibberish omitted
to save space>
...
end
Cut out everthing from 'begin' to 'end' and put it into its own file and call it 'pomdp-solve-4.0.uu'. Then type the following command:
% uudecode -p pomdp-solve-4.0.uu > pomdp-solve-4.0.tgz
At this point you should have the decoded file in the file pomdp-solve-4.0.tgz. This file is a compressed version of a "tar" archive file. It was compressed with the program 'gzip' after all the relevant files were joined into one file with the 'tar' program. The next thing is to undo these two steps.
Now, move that file into the directory where you want to put the program directory and "cd" into that directory. e.g., Suppose you wanted to put the program in ~/src, then type the commands:
% mv pomdp-solve-4.0.tgz
~/src
% cd ~/src
Now, type the following command:
% tar zxvf pomdp-solve-4.0.tgz
This both uncompresses the file and unpacks all the files that it contains, e.g., the source files, documentation, examples, etc. Some versions of "tar" do not support the "z" flag in which case you will need to "gunzip" it and the do the "tar" without the "z" flag.
That should do it as far as installing the code. You should first look at the files in the 'docs' subdirectory and also the files in the "examples" subdirectory. Before compiling, you might need to change some things in the Makefile and after that, you can then try typing "make dep" and then "make pomdp-solve" and see what happens.
Compiling
There are a few points I have to make concerning compiling the code:1) CPLEX issues: (ignore if these you don't use CPLEX)
a) For the CPLEX license stuff, you must definitely have the following environment variable set:
CPLEXLICENSE=/pro/cplex/license/cplexlicfile
This is how it was for a fixed-CPU license of CPLEX version 3.0, but a floating license our CPLEX future changes could make this differ. Essentially, a good check is to first make sure you can run the stand-alone CPLEX program by typing 'cplex' at the command line. If this doesn't work, then there is no hope of the pomdp-solve code working. Either way, the Makefile uses the presence or absence of the environment variable CPLEXLICENSE to decide whether to compile CPLEX into the pomdp-solve code or not.
b) You will need to make sure the location of the CPLEX libraries and include files are known to the pomdp-solve code as well making sure the CPLEX version (3.0 or 4.0) is known. You do this by defining the environment variable CPLEX_HOME to be the locations of the top-level CPLEX directory where CPLEX was installed. Also you can define the CPLEX_VERSION variable to be either '3' or '4'. Alternatively, you can change the Makefile to have the home directory and version of your CPLEX stuff hardwirded.
Also, note that recent versions of CPLEX (4.0 and above) seem to have a different include file and different library names. Worse yet is that they have changed all their function names. They now prefix all routines with 'CPLX', however they were kind enough to provide the file "coldcplex.c" that helps get around this. The environment variable CPLEX_VERSION will dictate whether this file should be compiled in or not.
Here are examples of the environment settings you need if you have CPLEX. Note that the actual values depend upon your particular installation, license and version of CPLEX.
For csh/tcsh shell:
setenv CPLEXLICENSE /usr/prop/cplex/cpxlicptr
setenv CPLEX_HOME /usr/prop/cplex
setenv CPLEX_VERSION 4
For bash shell:
export CPLEXLICENSE=/usr/prop/cplex/cpxlicptr
export CPLEX_HOME=/usr/prop/cplex
export CPLEX_VERSION=4
b) At this writing, I believe CPLEX was up to version 6.0. I have no experience with anything but 3.0 and 4.0, so you are on your own if you have a more recent version. However, I would like to hear what, if any, changes had to be made.
2) The Makefile attempts to automatically adjust the compilation depending upon the operating system type. Current types I have tried include Linux and Solaris (UltraSparc and Intel variations). The Makefile detects which is being used through the environment variable OSTYPE, so make sure this variable exists and is set to either 'linux' or 'solaris'. For most systems, these will probably exist without you having to do anything.
3) To build the thing, do this:
make clean
make dep
make pomdp-solve
Makefiles are still a bit mysterious to me, and I haven't had the time to do the IMakefile or "autoconfig" thing, so no guarantees that I actually have this set up to work properly on your system. It works on the few systems I have tried (Linux, Solaris, Solaris with CPLEX 3.0 and 4.0) , and I have tried to make it so it would work other places, but that's all I know.
Also, you can do:
make test
to build and execute a program which automatically runs the code through some simple tests This is used to ensure that installation went well and for regression testing on subsequent versions (if you or I change the code around).
If you want, you can also do:
make clean
to clean up all the object files, libraries, etc. or
make clean-local
to leave the libraries in tact, but clean up the main pomdp-solve code. This leaves the libraries in-tact and helps speed up development and compilation time.
Executing
You can run the pomdp-solve program with:
% pomdp-solve <args>
and
% pomdp-solve -h
gives you the complete list of valid arguments. See the cmd-line.html page for more information.
Bugs
- The lp_solve LP solver is no where need as stable as the CPLEX solver. Thus, on some problems weird things can happen not to mention that you could get slightly different answers from CPLEX. These will likely be related to the numerical stability of the solver code, though I wouldn't completely rule out a bug on my part.
- The linear support algorithm only works if you have CPLEX. I have not yet ported the vertex enumeration to the lp_solve code yet (and may never).
- There are still some lingering mysteries about the differences between "lex" and "flex". I have caught a few cases in the Makefile for lp_solve, but maybe not all of them, and it still gives me warnings for some versions.
- The code does not handle POMDPs with the utilities given as costs. The file lets you do it and some parts of the code can deal with it, but some parts no longer do. Much easier to multiply all the utilities by -1 and change to rewards in the POMDP parameter file.
- All the bugs I haven't discovered yet.
Compilation Transcript
Not sure if this will be helpful or not, but here is the transcript of what it looks like when I compile the code on linux without CPLEX.arc on marie % make dep
make -C ./mdp dep
make[1]: Entering directory
`/u/arc/proj/pomdp-solve-4.0/mdp'
gcc -E -M *.[c]
> Makefile.depend
make[1]: Leaving directory
`/u/arc/proj/pomdp-solve-4.0/mdp'
gcc -E -M -I./mdp -I./lp_solve
*.[c] > Makefile.depend
arc on marie % make
make -C ./mdp libmdp.a
make[1]: Entering directory
`/u/arc/proj/pomdp-solve-4.0/mdp'
bison -v -d parser.y
gcc -g
-c parser.tab.c
flex scanner.l
gcc -g
-c lex.yy.c
gcc -g
-c imm-reward.c
gcc -g
-c parse_err.c
gcc -g
-c parse_hash.c
gcc -g
-c sparse-matrix.c
gcc -g
-c mdp.c
ar -crv libmdp.a parser.tab.o
lex.yy.o imm-reward.o parse_err.o parse_hash.o sparse-matrix.o mdp.o
a - parser.tab.o
a - lex.yy.o
a - imm-reward.o
a - parse_err.o
a - parse_hash.o
a - sparse-matrix.o
a - mdp.o
make[1]: Leaving directory
`/u/arc/proj/pomdp-solve-4.0/mdp'
make -C ./lp_solve liblpk.a
make[1]: Entering directory
`/u/arc/proj/pomdp-solve-4.0/lp_solve_2.2'
gcc -g -Wall -pedantic
-ansi -DSYSLINUX -c lpkit.c -o lpkit.o
gcc -g -Wall -pedantic
-ansi -DSYSLINUX -c solve.c -o solve.o
gcc -g -Wall -pedantic
-ansi -DSYSLINUX -c debug.c -o debug.o
gcc -g -Wall -pedantic
-ansi -DSYSLINUX -c read.c -o read.o
read.c: In function
`read_lp_file':
read.c:572: warning:
implicit declaration of function `yyparse'
gcc -g -Wall -pedantic
-ansi -DSYSLINUX -c readmps.c -o readmps.o
gcc -g -Wall -pedantic
-ansi -DSYSLINUX -c hash.c -o hash.o
gcc -g -Wall -pedantic
-ansi -DSYSLINUX -c presolve.c -o presolve.o
bison -y lp.y
mv y.tab.c lp.c
flex lex.l
mv lex.yy.c lex.c
gcc -g -Wall -pedantic
-ansi -DSYSLINUX -c lp.c -o lp.o
/usr/lib/bison.simple:
In function `yyparse':
/usr/lib/bison.simple:327:
warning: implicit declaration of function `yyerror'
/usr/lib/bison.simple:387:
warning: implicit declaration of function `yylex'
lp.y:99: warning: implicit
declaration of function `check_decl'
lex.yy.c: In function
`yy_init_buffer':
In file included from
lp.y:204:
lex.yy.c:1362: warning:
implicit declaration of function `fileno'
lex.yy.c: At top level:
lex.yy.c:1116: warning:
`yyunput' defined but not used
ar rv liblpk.a lpkit.o
solve.o debug.o read.o readmps.o hash.o presolve.o lp.o
a - lpkit.o
a - solve.o
a - debug.o
a - read.o
a - readmps.o
a - hash.o
a - presolve.o
a - lp.o
ranlib liblpk.a
make[1]: Leaving directory
`/u/arc/proj/pomdp-solve-4.0/lp_solve_2.2'
gcc
-DSYSLINUX -g -pedantic -I./mdp -I./lp_solve -c global.c
gcc
-DSYSLINUX -g -pedantic -I./mdp -I./lp_solve -c timing.c
gcc
-DSYSLINUX -g -pedantic -I./mdp -I./lp_solve -c random.c
gcc
-DSYSLINUX -g -pedantic -I./mdp -I./lp_solve -c pomdp.c
gcc
-DSYSLINUX -g -pedantic -I./mdp -I./lp_solve -c alpha.c
gcc
-DSYSLINUX -g -pedantic -I./mdp -I./lp_solve -c stats.c
gcc
-DSYSLINUX -g -pedantic -I./mdp -I./lp_solve -c params.c
gcc
-DSYSLINUX -g -pedantic -I./mdp -I./lp_solve -c projection.c
gcc
-DSYSLINUX -g -pedantic -I./mdp -I./lp_solve -c lp-interface.c
gcc
-DSYSLINUX -g -pedantic -I./mdp -I./lp_solve -c region.c
gcc
-DSYSLINUX -g -pedantic -I./mdp -I./lp_solve -c parsimonious.c
gcc
-DSYSLINUX -g -pedantic -I./mdp -I./lp_solve -c neighbor.c
gcc
-DSYSLINUX -g -pedantic -I./mdp -I./lp_solve -c common.c
gcc
-DSYSLINUX -g -pedantic -I./mdp -I./lp_solve -c witness.c
gcc
-DSYSLINUX -g -pedantic -I./mdp -I./lp_solve -c cross-sum.c
gcc
-DSYSLINUX -g -pedantic -I./mdp -I./lp_solve -c enumeration.c
gcc
-DSYSLINUX -g -pedantic -I./mdp -I./lp_solve -c inc-prune.c
gcc
-DSYSLINUX -g -pedantic -I./mdp -I./lp_solve -c two-pass.c
gcc
-DSYSLINUX -g -pedantic -I./mdp -I./lp_solve -c vertex-enum.c
gcc
-DSYSLINUX -g -pedantic -I./mdp -I./lp_solve -c linear-support.c
gcc
-DSYSLINUX -g -pedantic -I./mdp -I./lp_solve -c policy-graph.c
gcc
-DSYSLINUX -g -pedantic -I./mdp -I./lp_solve -c cmd-line.c
gcc
-DSYSLINUX -g -pedantic -I./mdp -I./lp_solve -c signal-handler.c
gcc
-DSYSLINUX -g -pedantic -I./mdp -I./lp_solve -c pomdp-solve.c
gcc
-DSYSLINUX -g -pedantic -I./mdp -I./lp_solve -c main.c
gcc -o pomdp-solve
-L./mdp -L./lp_solve global.o timing.o random.o pomdp.o alpha.o stats.o
params.o projection.o lp-interface.o region.o parsimonious.o neighbor.o
common.o witness.o cross-sum.o enumeration.o inc-prune.o two-pass.o vertex-enum.o
linear-support.o policy-graph.o cmd-line.o signal-handler.o pomdp-solve.o
main.o -lmdp -llpk -lm