What is lp_solve and what is it not?
The simple answer is, lp_solve is a Mixed Integer Linear Programming (MILP) solver.

It is a free (see LGPL for the GNU lesser general public license) linear (integer) programming solver
based on the revised simplex method and the Branch-and-bound method for the integers.
It contains full source, examples and manuals.
lp_solve solves pure linear, (mixed) integer/binary, semi-continuous and
special ordered sets (SOS) models.

See the reference guide for more information.

lp_solve 5.5

Why a jump from version numbers 5.1 to 5.5 ?
This is done to indicate that this is more than just another update.
The solver engine was revised and optimised in such a way that performance has improved considerably.
Numerical stability is also better resulting in more models that can be solved.
The LUSOL bfp is now also the default. In the past, the etaPFI bfp package was the default,
but for larger models this leads faster to numerical instabilities and performance problems.

Overall, the v5.5 code is faster and more robust than v5.1.
This robustness is for example proven by the fact that many more models can now be solved even without scaling.

The API hasn't changed very much.
There are a couple of new routines and one routine has an extra argument.
Some constants got new values.

    * Fundamental internal change to the solver engine resulting in better performance and numerical stability.
      Both the LP solver and the B&B solvers are enhanced.
    * Optimised MILP branch truncation, with reduced cost fixing.
    * LUSOL bfp is now the default.
    * Presolve is improved in functionality and performance.
    * Better handling of degeneracy, with more options.
    * Store and read options from a file make it easier to set options.
    * Partial pricing for the primal simplex now works.
    * Full support for xli_ZIMPL v2.0.3.
    * The objective function is no longer stored as part of the constraint matrix.
    * Dual-long step code is in place, but not fully activated yet.
    * General code cleanup.
    * Added OBJSENSE and OBJNAME headers in the free MPS format (See MPS file format).
    * The MathProg xli driver has now the ability to generate a model.
    * New API routines

Start by taking a look at 'Changes compared to version 4', 'Changes from version 5.1 to version 5.5'
and 'lp_solve usage'
This gives a good starting point.


BFP stands for Basis Factorization Package, which is a unique lp_solve feature.  Considerable
effort has been put in this new feature and we have big expectations for this. BFP is a generic
interface model and users can develop their own implementations based on the provided templates.
We are very interested in providing as many different BFPs as possible to the community.

lp_solve 5.5 has the LUSOL BFP built in as default engine.  In addition two other
BFPs are included for both Windows and Linux: bfp_etaPFI.dll, bfp_GLPK.dll for Windows and, for Linux.  The bfp_etaPFI includes
advanced column ordering using the COLAMD library, as well as better pivot management for
stability.  For complex models, however, the LU factorization approach is much better, and
lp_solve now includes LUSOL as one of the most stable engines available anywhere.  LUSOL was
originally developed by Prof. Saunders at Stanford, and it has now been ported to C
and enhanced by Kjell.

If you compile BFPs yourself, make sure that under Windows, you use __stdcall convention and
use 8 byte alignments.  This is needed for the BFPs to work correctly with the general
distribution of lp_solve and also to make sharing BFPs as uncomplicated as possible.

See the reference guide for more information.


XLI stands for eXternal Language Interface, also a unique lp_solve feature. XLI's are stand-alone
libraries used as add-on to lp_solve to make it possible to read and write lp models in a format
not natively supported by lp_solve. Examples are CPLEX lp format, LINDO lp format, MathProg format,
XML format...

See the reference guide for more information.

lpsolve API

Don't forget that the API has changed compared to previous versions of lpsolve and that you just
can't use the version 5 lpsolve library with your version 4 or older code.  That is also the
reason why the library is now called lpsolve55.dll/lpsolve55.a.  lpsolve55.dll or lpsolve55.a are
only needed when you call the lpsolve library via the API interface from your program.
The lp_solve program is still a stand-alone executable program.

There are examples interfaces for different language like C, VB, C#, VB.NET, Java,
Delphi, and there is now also even a COM object to access the lpsolve library.  This means that
the door is wide-open for using lp_solve in many different situations.  Thus everything that is
available in version 4 is now also available in version 5 and already much more!

See the reference guide for more information.

Conversion between lp modeling formats

Note that lp2mps and mps2lp don't exist anymore. However this functionality is now implemented
in lp_solve:

lp2mps can be simulated as following:
lp_solve -parse_only -lp infile -wmps outfile

mps2lp can be simulated as following:
lp_solve -parse_only -mps infile -wlp outfile

via the -rxli option, a model can be read via an XLI library and via the -wxli option, a model
can be written via an XLI library.

How to build the executables yourself.

At this time, there are no Makefiles yet. However for the time being, there are batch files/scripts
to build. For the Microsoft compiler under Windows, use cvc6.bat, for the gnu compiler under Windows,
use cgcc.bat and for Unix/Linux, use the ccc shell script (sh ccc).

See the reference guide for more information.


Under Windows, there is now also a very user friendly lpsolve IDE. Check out LPSolveIDE

See the reference guide for more information.

Documentation (reference guide)

See lp_solve55.chm for a Windows HTML help documentation file.
The html files are also in lp_solve_5.5_doc.tar.gz. Start with index.htm
Also see for a on-line documentation

Change history:

17/05/05 version
- Beta release of version 5.5

??/??/05 version
- ?

26/06/05 version
- ?

29/06/05 version
- ?

16/08/05 version
- There are no API changes
- The LUSOL message routine could generate a crash under some cicumstances. Fixed
- A crash could occur when building the model in add_row_mode. Fixed.
- write_params didn't write the PRESOLVE and PRESOLVELOOPS correctly. Fixed.
- write_params didn't write constants with value 0. Fixed.
- The library did not compile under msdev 2002 (VC 7.0 _MSC_VER 1300). Fixed.
- There were some problems with printing long long variables which could generate a crash. Fixed.
- An overflow error could occur because memory was sometimes overwritten. Fixed.
- Presolve routines are revised. They are again improved and made faster.
  Also some problems with it are fixed (possible crashes).
- Solver revised. Again made faster and more stable.
- get_row/get_column returned FALSE if the row/column is empty. Fixed.
- get_rowex/get_columnex now returns -1 if and error is detected. This instead of 0.
  This to know the distinction between an empty row/column and an error.
- set_bounds had a possible problem when min and max are equal. Fixed.
- A crash/damage error could occur when rows/columns are added after a solve. Fixed.
- The my_chsign macro in lp_types.h gave warnings with some compilers. Fixed.
- The lp_solve program now returns 255 if an unexpected error occurs. Before this was 1
  But this interferes with the lpsolve library return codes.
- With the lp_solve program, debug and print modes were not written correctly in a
  specified parameter file. Fixed.
- With the lp_solve program, presolveloops was not set correctly. Fixed.

17/09/05 version
- In some cases, SOS restrictions were not optimized till optimality. Fixed.
- Presolve sometimes generated 'Column -xxx out of range during presolve' with a possible crash.
- Presolve sometimes removed integer and/or semi-cont variables that should not be deleted. Fixed.
- B&B sometimes didn't find the most optimal solution. Fixed.
- Internal constant COMP_EQUAL renamed to COMP_PREFERNONE because this could interfere with a define
  in the standard header files.
- The lp parser had problems with variables starting with INF and there is a + or - sign just before it.
- Added options -presolvem, -presolvefd, -presolvebnd, -presolved, -presolveslk
- Updated documentation. put_bb_branchfunc, put_bb_nodefunc, set_epslevel, dualize_lp, set_basisvar

16/11/05 version
- set_add_rowmode should not be called after a solve. There is now a test in this routine when this is
  done and it then returns FALSE.
- When an empty string ("") as filename is provided to set_outputfile, then output is completely
- In some cases, SOS models did not solve to their most optimal solution.
- There was as problem with get_sensitivity_objex. Calling it (sometimes only after multiple times)
  resulted in protection errors/core dumps.
- When a model has no constraints, it did not solve most of the times.
- column_in_lp didn't work anymore.
- Large upper bounds could make the model unstable. A change was made to fix this.
- set_improve could let crash the model.
- lp_params.c used the non-ANSI function unlink(). Changed to ANSI function remove().
- Presolve is again revised considerably.
- SOS handling is improved when there are a lot of SOS constraints.
- Limited constraint-to-SOS1 presolve to constraints with at least 4 variables.
- Limited bound tightening presolve loops.

12/02/06 version
- When SOS restrictions are added after a previous solve, a crash could occur.
- Optimized renaming a variable when the new name is equal to the old name.
- A possible crash could occur when presolve was enabled
- The constant ANTIDEGEN_DEFAULT is changed. ANTIDEGEN_INFEASIBLE is removed from it.
  This constant should not be used unless you have some very tight and hard to solve
  models where the optimal solution numerically straddles infeasibility.
- There was a possible problem with set_row(ex). It sometimes wrongfully changed the row.
- When integer variables were scaled, it could happen that because of rounding errors,
  a loop was created.
- Sometimes integer models kept on looping in the B&B algorithm.
- A memory overrun could occur when an initial basis is set. This when variable names
  are in Rnnn format and constraint names in Cnnn format.
- Some fixes are made in presolve.
- On 64-bit systems, compiler warnings were given and some code worked wrong resulting in
  wrong results.
- lp_solve.c didn't compile with some compilers because if a very deep nested if statement.
- The distributed files now have the version number include in the filename.
  For example
  This for a possible move to SourceForge in the (near?) future.
- When illegal bounds are specified in the MPS format (lower bound larger than upper bound)
  then a warning was given but the illegal bound was just ignored and the model was solved.
  This resulted in a solution that did not comply to the original model. Now the message is
  seen as an error and solving is aborted.

06/09/06 version
- When presolve is active and columns are removed and there are SOS constraints, then presolve
  had an error which could result in hanging while solve or maybe wrong solutions.
- set_row(ex) set wrong values when used after a previous solve and scaling is active.
- disabled PRESOLVE_REDUCEMIP since it is very rare that this is effective, and also that it adds
  code complications and delayed presolve effects that are not captured properly.
- made routine guess_basis available for all languages (now exported by the dll).
  The routine is now also documented.
- some bug corrections in guess_basis.
- Corrected a problem with add_column(ex) when add_rowmode is enabled.
- write_lp now wraps long lines over multiple lines to make it more readable.
- A compilation warning/error sometimes occured on is_fixedvar in lp_lusol.c with some compilers.
- Added options -wxlisol and -wxlisolopt to lp_solve program to write a solution file for those
  XLIs that support it.
- Updated CPLEX XLI to support constants in objective.
- Added documentation on infeasible models, guess_basis, DIMACS models, CPLEX format, Zimpl, GNU Mathprog.
  Corrected/updated documentation on get_col_name, get_row_name, get_nameindex, write_xli,
  External Language Interfaces.
- The mps reader was improved for the rarely cases where the same elements are provided multiple
  times. These are now added.
- Revised the java unittest example because it gave some errors that it shouldn't give.

07/10/06 version
- set_row(ex) could sometimes set wrong values in the model.
- Sometimes models with semi-cont variables which are also integer and scaling is active, a solution
  was returned that is not most optimal or it returns infeasible.
- write_mps didn't write semi-cont variables with an infinite upper bound.
- When presolve can solve the model on its own and objective direction is maximize then a wrong sign
  was shown in the reported price on screen.
- write_lp writes constraint and bounds in the same way if a constraint is not named. If a constraint
  only has one variable then it looks like a bound. This can give problems because when a constraint
  is interpreted as bound and it is negative then the problem definition changes.
  Therefore a constraint which is not named and having only one variable in it is getting a name to
  make sure it is interpreted as a constraint.
- The lp_solve program didn't interprete the PRESOLVED solve return code very well. Fixed.
- bfp_GLPK and xli_MathProg are now compiled against GLPK 4.11
- When an integer model is aborted before the most optimal solution is found (timeout or
  break at first, ...) solve returned OPTIMAL (0) instead of SUBOPTIMAL (1). This is now corrected.

14/01/07 version
- If a model has integer variables, but the model is already integer optimal in the simplex fase,
  then it was reported as suboptimal feasible.
- get_objective, get_variables, get_ptr_variables, get_constraints, get_ptr_constraints, get_primal_solution
  reported 'not a valid basis' when presolve is active and the model is entirely solved by presolve.
- presolve on a model with SOS variables sometimes went wrong.
- presolve on a model with SOS variables where infeasibility is detected crashed.
- read_bas could fail when not all constraints had names or had names like default variable names.
- A crash could occur with set_row(ex) in rowmode.
- The lp format has been extended with a free section to define free variables.
- bfp_GLPK and xli_MathProg are now compiled against GLPK 4.13
- fixed bug in the pseudocost logic that can blow up subsequent pseudocost values in that
  branch and make them almost random.
- In some rare cases a memory overrun could occur when constraints are added after a previous solve.
- Made the copy_lp routine work. Note that information about the optimisation of the original model
  is not copied (at this time). Only the model is.
- Fixed a bug in the hashing routines that had only influence in some rare cases in the
  MATLAB, O-Matrix, Scilab, Octave, Python drivers.
- coldual sometimes worked on a uninitialised variable value with unpredictable results.

27/12/07 version
- Fixed a problem in presolve. Sometimes an array-index-out-of-bounds error occured.
- Added a makefile for Linux.
- When adding constraints, in some rare cases a memory overrun could occur resulting in a crash.
- add_constraintex with count=0 and row=colno=NULL gave a protection error.
  several XLIs didn't work anymore because of this.
- set_constr_type sometimes set wrong signs for the coefficient matrix when add_rowmode is on.
- presolve did an unnecessary extra loop. Also a test is added to stop presolve when very few
  changes are done.
- for very large models, a request of much more memory than is reasonable could occur. Fixed.
- Modified LINDO XLI to read keywords also not at column 1 and to accept an empty objective function.
  Previously this wat not possible.
- In some rare cases, numbers very close to their integer values (for example 11276.999999999998)
  were truncated to their ceiling value (for example 11276) instead of rounded
  (for example 11277).
- Solved a problem with presolve with an all-int constraint.
- Solved a problem with presolve coldominate
- Added stronger checking of the MPS format.
  Fields that must be blank are now also tested accordingly so that if data is there that it is
  not ignored as before.
- FREE MPS format was not read ok if row/column name were all numbers
  or a FR, MI, PL, BV bound was defined. Fixed.
- The lp-format now also supports a bin(ary) section to define binary variables.
- When an integer model is aborted before the most optimal solution is found
  via break at first or break at value, solve returned OPTIMAL (0) instead of SUBOPTIMAL (1).
  This is now corrected. Problem occured in version
- Fixed a problem with del_constraint. Constraints names were not shifted and reported variable result was incorrect.
- read_XLI failed with MathProg if datamodel was provided with "" when there is no datamodel.
  NULL was expected in the past. "" is now also accepted.
- Added an XLI to read Xpress lp files.
- Added routines MPS_writefileex, write_lpex.
- Added options -o0, -o1 to lp_solve command driven program to specify if objective is in basis or not.
- Added new information in the reference guide:
   - Linear programming basics
   - Presolve
   - Xpress lp files

04/01/08 version
- There was an error in the lp-parser resulting is misreading the model in many cases. This was fixed in the existing release and an update of following files is posted:

13/03/08 version
- When NODE_RCOSTFIXING (bound tightening during B&B) (enabled by default) was enabled some models were not solved
  to optimality.
- In following situation a (sub-optimal) solution was returned while no integer
  solution isn't found yet at this time.
   - The model contains integers
   - A break at first is set
   - A timeout is set
   - The timeout occurs before a first integer solution is found
   - When the timeout occurs, the simplex algorithm is in phase 2 and has a feasible (but non-integer) solution, but not optimal yet.
- If an integer solution is found with the same objective value as the relaxed solution then
  B&B is stopped. However this resulted in a report of a SUBOPTIMAL solution while this is not true.
  The found solution is then optimal.
- On 64-bit Linux systems crashes occured when information is printed via variable argument routines.
- There was a warning about redefinition of MAXINT32, MAXUINT32, MAXINT64, MAXUINT64 when compiled
  with MS Visual C 2008.
- The reference guide now also contains the Java documentation.

03/08/08 version
- The code should now completely be reentrant and thread safe.
  Especially the lp-parser, but also the solver code itself used static variables which made it
  not reentrant.
  The lp parser code has changed considerably for this. Also the lex and yacc code had to be made
  reentrant. This is only tested with GNU flex and bison. Not with lex and yacc.
  This should be no problem since the corresponding c files are included with the source so people
  don't have to build these themselves.
- commonlib.c/h were also under BFP/BFP_LUSOL/LUSOL while a more advanced version is also
  under the shared directory. The one from LUSOL is removed.
- If objective coefficients are all integer and the columns are integer or objective coefficients
  are zero, then B&B can stop when the objective function equals ceil/floor(real objective)
  which results in a faster solving process. This did not work anymore and is now fixed.
- The lpsolve IDE is now also on sourceforge (previously it was a link to another site).
  Also the sources are included.

02/02/09 version
- The lp format now allows spaces between the row label and the colon.
- Added better error handling in the copy_lp API
- Timeout in a MIP model sometimes resulted in an endless loop
- Revised isnan call for gnu compilers
- Removed some extra static variables for better reentrance support.
- REPORT_Objective now reports large objective values (>1e-5) with better precision
- On Windows the binaries are again compiled with compiler optimization /O2 for better performance.
- Compiled and tested on Windows and Linux 64-bit. 64-bit binaries are now provided.
  Changed project files and compile scripts for 64-bit compilation.
  Binaries are now located in a subdirectory bin\platform
- Added PHP driver so that lpsolve can be called from PHP
- When an MPS file is created, upper bounds were written before lower bounds.
  Now it is the other way around. First lower bounds and then upper bounds.
- The MPS reader could not handle negative lower bounds where the lower bound was specified after the upper bound.
- The MPS write now writes the implicit lower bound of zero if the variable has no lower bound and it is integer.
  This to make sure that other solvers (like CPLEX) interprete the variable as integer and not binary.
- The LINDO XLI parser interpreted negative lower and upper bounds as positive
- Added the option -stat to the lp_solve driver program. It prints statistics of the model like number of
  rows/columns used and smalled and largest value in objective, RHS and coefficient matrix.
- Added the option -plp to the lp_solve driver program. It prints the model via the print_lp API function.

09/09/09 version
- Improved guess_basis
- set_row(ex) corrections. Actually completely revised the routine.
  Sometimes the matrix was corrupted after the call.
- When in set_add_row mode, all API calls can now be used. For example printing or writing the model,
  retrieving data from the model and so on are now all possible in that mode.
  Will help greatly when debugging your models.
- Sometimes an integer model is reported infeasible while it isn't because the integer solution
  is much different from the relaxed (real) model. Then API call set_bb_depthlimit can be used
  to increase the default B&B depth. However its new value was not used for the depth limit of an
  individual variabled which could result in still infeasible results.
- modified the demo program not to use the str_ versions of API calls, because they are not
  performant and not intended to be used in real programs.
- Added an lpsolve driver to Sysquake.
- Added an lpsolve driver to FreeMat.
- Added an lpsolve driver to Euler.
- Added an Access example.
- Added documentation to use lpsolve from Sage via the existing Python driver.
- Changed the second parameter to read_mps, read_MPS, read_freemps and read_freeMPS from verbose to options.
  These routines now supports via this options parameter the native IBM interpretation of
  integer variables and allows to negate the objective constant.
  This is also supported by the lp_solve command line program via the -mps_ibm and -mps_negobjconst options.
  This is also supported by the IDE in the options tab.
- Removed read_LPhandle and read_MPShandle from the lprec structure.
  On their place there is now read_LP and read_MPS.
  The handle functions are not usable anyway under windows as dll because the FILE structure
  is searched in the dll space and therefore not working.
  read_LP and read_MPS work on the file name and always work.
  Since these handle routines were not documented anyway and the lp structure should not be
  accessed by applications (except XLI and BFP drivers) this change should not give any (compatibility) problem at all.
- write_lp/write_LP write just a + or - if the factor to a variable is +1 or -1
  However when the number was not exactly one, for example 1.0000000000001, then the test
  on equal to one gave false, but the writing was a 1 because only 12 significant digits are written.
  The result was that the 1 was written in that case. Not an error, but when lp files are compared,
  this is enoying. Now this is solved. When a +1 or -1 would be written then alyways only + and -
  is written and +1 or -1 will never occur again.
- When a message function is set via put_msgfunc,
  in message MSG_LPOPTIMAL it was not possible to get the solution. Now it is.
  Also when messages MSG_MILPFEASIBLE, MSG_MILPBETTER, MSG_MILPEQUAL were catched the
  returned solution was the one from the previous MIP solution. This is now also fixed.
  With this fix, the IDE now also shows the relaxed solution (column LP Optimal).
- Extended the MATLAB, O-Matrix, Scilab, Octave, Python, PHP and of course the new
  Sysquake, FreeMat and Euler drivers to support string constants. See the reference guide.
- Compiled and tested lpsolve on MAC OSX 10.4.8 (Intel). Compilation scripts for this platform
  were added and revised and binaries are provided to download.
- Revised the c# and lpsolve code to pass arrays back and forth to the dll to make it also work on 64 bit systems.
- IDE enhancements:
      - Can talk to the lpsolve55.dll large address aware version so access to 3.4G 32 bit memory.
      - Can write and read lp_solve parameter files ( type .lpi) via Options, Solver params, so
        it is easy to save a readable file of parameters for a particular problem class when it is found.
        The file can then be collected by the -rpar exe param or the read_params() API.
      - Added Options, Reset to solver defaults to go back to all lp_solve defaults.
      - Added in Help a link to the online help files.
      - Resized the help, main and statistics windows for Vista.
      - Added the Xpress XLI to the setup.
      - Use the latest dlls, especially lpsolve55.dll, in the setup.
      - Added the two new MPS options for IBM integer variables and negate objective constant.
      - Added the option to ignore integer restrictions.
17/09/09 version
- objfrom values where not correct if scaling is active.
  Routine to calculate these values and the objective from-till values is revised
  and made a bit more performant.
- write_lp did not write all column data. For objective function and constraints.

28/10/09 version
- In some cases with a model with integer variables lp_solve did not find the most optimal solution. Fixed.
- The reported relative gap for the optimal solution was not always ok.
- Sensitivity on a zero upper bounded variable was not calculated. Now it is.

08/03/10 version
- When there are integer variables in the model then sensitivity is not correct.

12/08/10 version
- add_SOS did a wrong test when SOS type is bigger than 2 if this SOS is allowed in lp_solve. Fixed
- new and improved MIP_stepOF function to find integer solutions.
- fixes in set_row.
- replaces code
   w *= 2*lp->is_lower[i] - 1;
   w = my_chsign(!lp->is_lower[i], w);
  because the old failed on AIX
  Also changed the definition of my_chsign in lp_types.h in the same way.
- Added the option -BR in the lp_solve driver program to select branch rule NODE_PSEUDORATIOSELECT
- lpsolve.pas updated for Delphi 2009.
- Python driver updated to allow also to provide numpy arrays to the lpsolve driver.
- xli_MathProg is now compiled against glpk 4.44 meaning that it now supports the MathProg Tables feature
  as described in
  csv, ODBC and MySQL are supported on all platforms.
- Added MSF support. MSF (Microsoft Solver Foundation) is a microsoft .NET library to access solvers via an OO way
  MSF has a default lpsolve driver, but the one on the sourceforge site is enhanced in functionality and performance.
  There is also documentation in the lpsolve reference guide.

We are thrilled to hear from you and your experiences with this new version. The good and the bad.
Also we would be pleased to hear about your experiences with the different BFPs on your models.

Please send reactions to:
Peter Notebaert:
Kjell Eikland: