Loading...
Searching...
No Matches
CHANGELOG April 2024 - Release 7.0.1
==================
code quality:
- update version of fmt from 6.1.2 to 7.1.3 due to deprecation
- add compile option -ffp-contract=off or /fp:precise to enhance reproducibility across different systems
fixed bugs:
- set the c++ standard consistently to c++14 for all build systems
- fix wrong detection of unboundedness for infeasible instance in primal simplex
- remove unnamed boolean setting STORE_BASIS_BEFORE_SIMPLEX_PIVOT
- continue parsing settings after errors and print them consistently
- find PaPILO installation by default
February 2024 - Release 7.0.0
==================
features:
- add option to print primal/dual solutions to a file using `-x=<filename>` or `-y=<filename>`, respectively
exact rational solutions can be written with `-X=<filename>` or `-Y=<filename>`. If `<filename>` is the
same for multiple options the different solutions are appended to the same file
- new option to use incremental precision boosting for rational LP solving. If MPFR is linked (requirement for boosting)
the new default configuration is precision boosting an outer iterative refinement loop. Can be changed with `bool:precision_boosting`
interface & parameters:
- allow reading of files with arbitrarily long lines by reallocating buffers dynamically
- add new functions to soplex_interface.cpp: SoPlex_readInstanceFile(), SoPlex_readBasisFile(),
SoPlex_readSettingsFile(), SoPlex_setBoolParam(), SoPlex_setRealParam(), SoPlex_removeColReal(),
SoPlex_removeRowReal(), SoPlex_getRedCostReal(), SoPlex_getStatus(), SoPlex_getSolvingTime(),
SoPlex_getNumIterations(), SoPlex_changeRowLhsReal(), SoPlex_changeRowRhsReal(), SoPlex_changeRangeReal(),
SoPlex_changeRowRangeReal(), SoPlex_changeLowerReal(), SoPlex_changeVarLowerReal(), SoPlex_getLowerReal(),
SoPlex_getObjReal(), SoPlex_changeUpperReal()
- removed parameters related to removed legacy code:
`bool:computedegen`, `bool:decompositiondualsimplex`,
`bool:explicitviol`, `bool:usecompdual`, `int:decomp_displayfreq`, `int:decomp_iterlimit`,
`int:decomp_maxaddedrows`, `int:decomp_verbosity`
- added parameters related to exact solving and precision boosting:
`bool:adapt_tols_to_multiprecision` -> should tolerances be scaled down when precision is increased
`bool:boosted_warm_start` -> should high-precision solve start from advanced basis
`bool:iterative_refinement` -> should iterative refinement be used
`bool:precision_boosting` -> should precision boosting be used
`bool:recovery_mechanism` -> should the old recovery mechanism be used when iterative refinement fails
`int:multiprecision_limit` -> maximal precision in decimal digits
`int:storeBasisSimplexFreq` -> frequency to store basis in high precision solve
`real:precision_boosting_factor` -> factor by which precision is increased in each boosting iteration
- changed minimum values for parameters `real:fpfeastol` and `real:fpopttol` from 1e-12 to 0
- new parameters to enable/disable individual presolving techniques:
`bool:simplifier_enable_domcol`, `bool:simplifier_enable_dualfix`, `bool:simplifier_enable_fixcontinuous`,
`bool:simplifier_enable_parallelcols`, `bool:simplifier_enable_parallelrows`, `bool:simplifier_enable_propagation`,
`bool:simplifier_enable_singletoncols`, `bool:simplifier_enable_stuffing`
performance:
code quality:
- rework the way tolerances are handled in SoPlex and remove many hard-coded parameters:
there now exists a shared_ptr<Tolerances> in soplex.h that is handed down to all the sub-classes
this is especially necessary for higher-precision solves with precision-boosting
- rewrite all preprocessor-defines to be unambiguous by prepending a `SPX`, e.g., `MSG_ERROR` is now `SPX_MSG_ERROR`.
- rework soplex debug messages to use SFINAE, since they no longer worked properly after switching to header-only
- add .lp instances to ctest
- remove unmaintained legacy code for decomposition based dual simplex
fixed bugs:
- fix unset tolerances for simplifier and ratio tester plugins in soplex' copy constructor
- fix unset outstream pointer in soplex' copy constructor
- fix bugs in function definitions in soplex_interface.cpp: SoPlex_changeVarUpperReal(), SoPlex_getUpperReal()
- fix bugs in LP file reader, concerning rational numbers as well as dynamic line length
- fix bug printing solutions and statistics when running from the command line
- fix bug in cmake build system when MPFR is available but boost is not
August 2023 Bugfix Release 6.0.4
=============================
features:
performance:
- avoid memory operations in SLUFactorRational::clear() if status is UNLOADED
interface & parameters:
fixed bugs:
- reoptimize without preprocessing and scaling if SoPlex terminates with ABORT_VALUE,
but objective value does not exceed the limit for the original problem
- make sure fast presolvers are finished before starting dual presolving
- fix numerical issues in solution polishing by reoptimizing afterwards,
if polishing affected the objective value
- make sure fast presolvers are finished before starting dual presolving
miscellaneous:
December 2022 Release 6.0.3
=============================
miscellaneous:
- changed license to Apache 2.0, see https://www.apache.org/licenses/LICENSE-2.0.
October 2022 - Release 6.0.2
=============================
- version bump
June 2022 - Release 6.0.1
=============================
fixed bugs:
- fix bug in varStatusToBasisStatusCol when setting Basis for implicitly free variables
January 2022 - Release 6.0.0
=============================
features:
- integrated PaPILO as new simplifier option
- improved the buildsystem by adding a cmake-configurable `config.h.in` header in the same fashion
as it is handled in SCIP
- replaced SoPlex internal `Rational` class by the boost multiprecision library for rational numbers;
this enables the possibility of performing a rational LP solve, even without linking GMP
- restored classic command line interface to remove dependency on the Boost program options library
- changed the command-line argument for running soplex with higher precision;
formerly the multiprecision mode was triggered by setting `solvemode=3` (multiprecision),
or `solvemode=4` (quadprecision); Now higher precision arithmetic has its own
argument `arithmetic` (0 - double, 1 - quadprecision, 2 - multiprecision);
This makes it possible to also use rational solving mode in combination with multiprecision
interface & parameters:
- added interrupt flag to optimize() (default NULL)
- added new simplifier enum SIMPLIFIER_PAPILO and SIMPLFIER_INTERNAL (replacing the deprecated SIMPLIFIER_AUTO)
performance:
code quality:
- fixed deprecated copy compiler warnings that occur with newer compilers
- added a ClassSet<T> class that functions in the same way as DataSet<T> for types that are not data-types
fixed bugs:
December 2020 - Release 5.0.2
===============================
performance:
- stricter limitation of refinement rounds with failed or slow progress in accuracy
fixed bugs:
- add missing invalidations of nonBasicValue (these led to incorrect cutoffs due to objective limit)
- fix comparison in postsolving of row singleton simplifier
- fixed overly conservative assert in postsolving of row objective simplifier
- always clear rational LU factorization when changing the problem to avoid inconsistent basis data
- ensure consistent basis data when floating-point solves fail during iterative refinement
- fixed bug regarding objective sense when starting SYNCMODE_MANUAL
- handle infinity values in IdxCompare of pricers
- fixed incorrect assert in devex and steepest edge pricer
23 June 2020 - Release 5.0.1
===============================
fixed bugs:
- add missing case of `scale = true` in spxlpbase.h change functions
- corrected feasibility checks for rational solution in debug mode
- Makefiles: fix a typo in SANITIZERFLAGS, if SHARED=true, add SANITZERFLAGS also to LIBBUILDFLAGS
and always add SANITZERFLAGS to LDFLAGS
- Makefiles: fix BOOST=false SHARED=true case
- CMake: add minimum gcc requirement
- CMake: fix for older cmake versions and for building without boost
- iterative refinement: fixed possible infinite loop and consistency issue
30 March 2020 - Release 5.0.0
=============================
interface & parameters:
- show number of violations in log to improve understanding of solving progress
- new parameter 'int:printbasismetric' to replace 'int:printcondition'
(-1: off, 0: condition estimate, 1: trace, 2: determinant, 3: condition)
- new parameter 'real:min_markowitz' to limit the Markowitz stability threshold
(range 0.0001 to 0.9999, high values sacrifice performance for stability)
- new parameter 'int:stattimer' to control time measurement for statistics during solving
(0: off, 1: user time, 2: wallclock time)
- new parameter 'bool:forcebasic', determines whether optimal solution has to be basic when performing
rational solve with iterative refinement, triggers rational factorization if reconstruction optimal
but not basic
- SoPlex can now do a multiprecision solve, i.e., `SoPlex<R>` with `R =
boost::multiprecision::number<T>` will work. In particular, for Linux and Mac,
SoPlex can do a multiprecision solve with `mpfr_float` and on windows with
`cpp_float`.
- extended parameter 'solvemode' with value (3 - multiprecision solve, 4 - solve with Quadruple precision (128bit))
- new parameter 'precision' (only active in solvemode 3) to set the precision
- Templated SoPlex, SPxSolver, and SPxBasis.
- Renamed SoPlex<R>, SPxSolver<R>, and SPxBasis<R> to SoPlexBase<R>,
SPxSolverBase<R>, SPxBasisBase<R> respectively for backwards compatibility
(and for SCIP.)
- Following renames:
| From | To |
|------------------------|-----------------------|
| `getDualFarkas` | `getDualFarkasSol` |
| `getDualFarkasT` | `getDualFarkas` |
| `getBoundViolationT` | `getBoundViolation` |
| `numColsT` | `numCols` |
| `numRowsT` | `numRows` |
| `getRedCost` | `getRedCostSol` |
| `getRedCostT` | `getRedCost` |
| `writeFile` | `writeFileLPBase` |
| `writeFileT` | `writeFile` |
| `getRowViolationT` | `getRowViolation` |
| `getDualViolationT` | `getDualViolation` |
| `getRedCostViolationT` | `getRedCostViolation` |
| `getDual` | `getDualSol` |
| `getDualT` | `getDual` |
| `getPrimal` | `getPrimalSol` |
| `getPrimalT` | `getPrimal` |
| `numNonZerosT` | `numNonZeros` |
| `_optimizeT` | `_optimize` |
| `getPrimalRay` | `getPrimalRaySol` |
| `getPrimalRayT` | `getPrimalRay` |
- Wrapped the following functions:
1. `getDualFarkasReal`
2. `numColsReal`
3. `getRedCostReal`
4. `getDualReal`
5. `getPrimalReal`
- Renamed `SPxSolverBase<R>::getDualfarkas` to `SPxSolverBase<R>::getDualFarkas`
(for consistency.) Note that there is a method with the same name in `SoPlexBase<R>`.
- Rewrote the code for command line argument parsing. Current code uses Boost Program Options.
Note that you do not need Program Options if you just want to build the SoPlex library.
- Some of the memory allocation from heap (using spx_alloc) changed to static memory.
- Improvements in exception handling in `src/soplexmain.cpp`
- Renamed `int numNonzeros` to `nNonzeros`
- Moved definitions of generic functions from a `.cpp` file to corresponding
`.hpp` file
- New features for Rational class:
1. Constructor from `boost::multiprecision::number<T>`.
2. Conversion from Rational to `boost::multiprecision::number<T>`, provided
that boost can handle `mpq_t` to `T` conversion.
- Changes to `VectorBase<T>` and `DVectorBase<T>`:
1. `VectorBase<T>` is now a wrapper of `std::vector<T>`, it no longer uses
pre-allocated memory.
2. The class can handle the additional functionality `DVectorBase<T>`, e.g.,
`memSize()`, `reSize()`, `scaleAssign(,,).`
3. internal member `int dimen` *removed*. `val.size()` or `VectorBase<R>::dim()`
is a replacement.
4. Replaced calls to memcpy and memmove with `std::copy().` Former functions
are not compatible with Types that are not trivially copyable.
5. Defined move constructor and move assignment.
6. Added begin() and end() iterators that return the corresponding iterator of
the internal std::vector.
7. Added vec() function that returns a const reference to the underlying vector.
8. *Reason for changes*: VectorBase previously used to manage a pointer and
DVectorBase managed memory allocation by doing malloc, realloc and free
calls. But this is only guaranteed to work with trivially copyable data types,
the new g++ compilers were already issuing warnings about this for some time.
The multiprecision vectors were not working. The same is also true for changes
inside CLUFactor classes.
8. *Removed* `isConsistent()`, which used to check a vector was empty and had no
memory allocated at the same time. This never happens with the std::vector
wrapper, hence the removal.
9. *Removed* DVectorBase. Replaced by VectorBase. The alias DVector is replaced by
the alias Vector. (This has the same problem as VectorBase<T>)
- DataArray now checks if template type is trivially copyable via `static_assert`.
Changed all instances of DataArray where this is not the case to Array.
- Changed Array to wrap std::vector to be more efficient and easier to use.
- Changed spxmainsm to use smart pointers inside of Array
- The following internal objects of the CLUFactor class was changed from a raw
pointer to a std::vector
1. `u.col.val`, `u.row.val`
2. `l.val`, `l.rval`
3. `diag`
4. `s_max`
- Replaced several instances of "`spx_alloc` an object, then `spx_realloc` the
object" with using `std::vector`. Because `realloc` is not guaranteed to work
on objects with non-trivial copy constructors, in particular, with
`boost::multiprecision::number.`
fixed bugs:
- use center value for very close bounds in singleton postsolver
- enable initial/crash basis methods again
- adjust bounds if they are equal within epsilon
- fix untriggered bug in copy constructor of UnitVector
10 July 2019 - Release 4.0.2
============================
fixed bugs:
- resolve compiler warning about data loss during implicit type conversion for timers
- return correct basis status when aborting a solve early
- ensure basis status is set correctly when aborting a solve early
- fix memory leak when copying SoPlex
- correct getBasisInverseCol for row representation
code quality:
- merge SoPlex::hasDual() and SoPlex::hasPrimal() into SoPlex::hasSol() and mark old methods deprecated
09 January 2019 - Release 4.0.1
===============================
fixed bugs:
- correct objective function update in aggregation simplifier
- use center value when fixing variables with slightly different bounds
- MPS reader: print error message and terminate in case of duplicate column names or not column-wise ordering
- LP reader: fix reading of infinite bounds and right-hand sides
- return clean solution values according to optimal basis (e.g. reduced cost 0 for basic variables)
- avoid unsuccessful termination when cycling and rather return sub-optimal solution
code quality:
- hide symbols in library to avoid symbol collisions and reduce library size
13 June 2018 - Release 4.0.0
============================
features:
- new simplifier step to perform variable aggregation for equations with two variables
interface & parameters:
- new option bool:ensureray to re-solve the original problem in case of
infeasibility/unboundedness to get a valid proof/ray (default: false)
- new status OPTIMAL_UNSCALED_VIOLATIONS to signal unresolved violations in original space
(use method SoPlex::ignoreUnscaledViolations() to reset the status to OPTIMAL)
- problem statistics distinguish ranged rows from equalities and show intuitive type counters
for variables and constraints
- remove LEGACY mode (C++11 standard is mandatory)
fixed bugs:
- allow only one clean up round to avoid cycling on numerically difficult infeasible/unbounded instances
- improved numeric stability in fixColumn and aggregation presolving steps
code quality:
- install all header files except for soplex.h into a subdirectory soplex/
05 February 2018 - Release 3.1.1
================================
fixed bugs:
- correct accumulation of shifted or relaxed bounds
- allow termination despite violations to prevent cycling on numerically difficult models
- add missing forward declarations of template specializations to spxlpbase.h
- changed all string copy methods to use spxSnprintf which is both safe and more efficient
than strncpy.
- set correct basis status for removed free variables (e.g. column singletons without objective)
- use consistent generic constraint names when writing .mps files
21 December 2017 - Release 3.1.0
================================
features:
- new scaling method combining geometric and equilibrium scaling (command line option '-g6')
interface & parameters:
- also print reduced cost values when passing '-y' on the command line
fixed bugs:
- correct API method for multiplication with basis matrix for sparse vectors
- fixed bug when setting and getting dual pricing norms
- don't print infeasible primal/dual solutions
- properly clear update vectors in initialization step
- fix bug in geometric scaling
- respect the time limit during solution polishing
- automatically remove zeros in sparse vector class SVector
01 September 2017 - Release 3.0.1
=================================
features:
- new condition number estimate based on the quotient of max and min diagonal element of LU factorization
(final condition number is also printed in statistics)
- new CMake build system alongside existing Makefiles
interface & parameters:
- allow arbitrary order of inputs/parameters, esp. filename is not required to be last
- new bool parameter 'printcondition' to print condition number in iteration log
- new option 'extsol' to validate objective value using an external solution
- new option 'l' to adjust the tolerance used in the external validation process
fixed bugs:
- fixed compilation error with WITH_LONG_DOUBLE
- fixed wrong objective value when using an obj offset
- use correct row and column names when writing the dual problem
08 March 2017 - Release 3.0.0
=============================
features:
- new scaler: Least Squares, aka Curtis-Reid-Scaling
- solution polishing: perform additional primal pivots at optimal basis
turn on with parameter 'int:solution_polishing'
0 - off
1 - try to increase number of basic slack vars
2 - try to decrease number of basic slack vars
- persistent scaling:
- keep scaled LP for multiple reoptimizations
- modifications to LP are correctly transformed
- access to (inverse of) basis matrix is correctly handled
- enabled by default
- decomposition based dual simplex:
- an experimental version of the a decomposition based approach to avoid degeneracy in the dual simplex method
- activated by setting the parameter bool:decompositiondualsimplex to true.
- this sets the basis representation to 'row', the solve mode to 'real', and the algorithm to 'dual'.
- additional parameters include:
usecompdual: setting to true using the dual formulation of the complementary problem, default is false.
explicitviol: setting to true explicitly computes the violation of the reduced problem solution instead of using the
complementary problem to identify violated rows, default is false.
decomp_iterlimit: the number of algorithm iterations for the decomposition based simplex.
decomp_maxaddedrows: the maximum number of rows added to the reduced problem in each algorithm iteration.
decomp_displayfreq: the frequency of output for the decomposition based simplex.
decomp_verbosity: the verbosity level of the decomposition based simplex.
- computing the degeneracy of a basis in each iteration:
- the parameter computedegen, when set to true will compute the degeneracy of each basis.
- write dual formulation to file: additional command line option "--writedual" is provided to write the dual formulation
to a file. When providing this option only the file is output, the problem is not solved.
- new condition number estimate based on the quotient of max and min diagonal element of LU factorization
performance:
- more efficient SoPlex class with better usability:
- avoid unnecessary resolves of problem after unsimplifying
- access to solution data also when suboptimal
- use exact weight initialization when switching to steepest edge in auto pricing
interface & parameters:
- new parameters to control refactorization triggers
- random seed for perturbation can be set from outside and is not reset in between
- set default representation switch to 1.2
- setTerminationTime() in SPxSolver will convert a negative limit to 0 (was set to infinity before)
- new interface methods to access rational factorization of basis matrix
- allow parsing of "inf" bounds in MPS reader
- new method to reset parameters to default values
- new method to change the objective function offset
- setParam() methods use default parameter 'init=true' to force updates in solver settings
- print primal ray and dual farkas proof with '-x' and '-y'
- submakefile for CrayXC systems added
- converted many error messages to info messages that can be turned off using low verbosity
- correct output of primal and dual solution values depending on set precision
- solve() is now called optimize() to distinguish it from solves with the basis matrix; solve() is deprecated
- LPs can be loaded into the solver w/o setting up a slack basis (new default param in loadLP() method)
- persistent scaling can be disabled by setting bool:persistentscaling to 0
- new bool parameter 'fullperturbation' to perturb the entire problem instead of only the bounds relevant
for the current pivot
- new bool parameter 'isOptimal' for unsimplify() to disable some debug checks that are only valid for optimal solutions
- remove static members Rational::{ZERO,POSONE,NEGONE} when compiling without GMP; make private when compiling with GMP
fixed bugs:
- fix inconsistent basis status in Basis and Solver
- correct definition of numeric constants when using LONG_DOUBLE mode
- fix bug in class SemiSparseVector
- correctly define WIN flags for Cygwin
- fix bug in left solve with rational LU factorization
- shared library knows its dependencies (GMP, ZLIB)
- when compiling with SHARED=true, the binary will link to libsoplex.so, resulting in a smaller file
- fixed memory bug in class DSVectorBase when resizing zero length vectors
- restore initial basis when given basis has too many basic variables
- improve behaviour for LPs with zero constraints
- cleaned up and corrected steepest edge pricing implementation
- check for very small weights in devex pricing (just like for steepest edge)
- forbid setting termination tolerances to zero when not compiled with rational GMP support
- fixed and simplified updating of range types for the rational LP
code quality:
- SoPlex is thread-safe by using the storage class specifier thread-local for static variables
- disabling verbosity when using PARASCIP is not necessary anymore
- use inline function to compute pricing values
- enable the user to compile without thread_local by setting USRCXXFLAGS=-DTHREADLOCAL=""
29 February 2016 - Release 2.2.1
================================
performance:
- faster updateFtest() by exploiting sparsity in solveVector3
- new method noViols() for checking violations, that terminates faster than maxInfeas()
- new refactorization trigger that takes basis size into account
interface & parameters:
- show current dual (primal) violation of dual (primal) simplex in log
- options -X and -Y to print primal and dual solution as rationals
- add new compiling flag OPENSOURCE to allow/forbid the usage of third party software
- improve Makefiles for better usage of MSVC on MinGW
fixed bugs:
- correct initialization of objective offset in simplifier
- fix for compilation with Visual Studio 2010
- allow to extend also empty SVectors
- better internal handling of singular bases
- check whether the solver is available when setting dual norms
- fix removing rows/cols in SoPlex interface
- check objective limit also when presolving is activated
- fixed compilation problems under Windows with MSVC
code quality:
- removed some hard float comparisons (use 'EQ()' instead of '==' )
- check scripts are now python3 compliant
30 June 2015 - Release 2.2.0
============================
features:
- support for accessing objective coefficients of implicit slack variables in column representation ("row objectives")
- rational solves with basis matrix based on a rational LU factorization
- rational reconstruction routines during iterative refinement
performance:
- iterative refinement for ranged rows without introducing explicit slack variables
- more efficient rational arithmetic through EGlib's memory allocator (available when building with EG=true)
- use sparse data structures (SSVector instead of Vector) for LU solves with 2 or 3 right hand sides
(necessary for steepest edge weights and/or bound flipping ratio test updates)
- do not remove bounds of boxed variables or sides of ranged rows if bound flipping is used
- faster file reader due to better hashing (increase hash table size earlier and use more prime numbers)
- more efficient arithmetic in class Rational by checking special cases
- faster iterative refinement by using powers of two as scaling factors
- reuse factorization during iterative refinement
- no transformation of objective sense for rational solve
- simplifier makes fewer constructor/destructor calls to DSVector class by reusing the data structures
- update nonbasic part of objective function value instead of completely recomputing it every time
- activate hyper sparse pricing per default
- faster solve methods by clearing temporary vectors more efficiently (esp. useful for steepest edge pricing)
- steepest edge and devex pricer share the same norm/weight array to avoid the initialization phase after a pricer switch
interface & parameters:
- new parameter bool:eqtrans to control transformation to equality form before rational solving
- new make option EG={true,false} to build with EGlib
- bound flipping is now enabled by default when using column representation
- new command line option --writefile to save LP file as read by SoPlex
- new parameter bool:testdualinf to decide dual feasibility even if LP is primal infeasible and farkas ray is available
- improved printing of rationals in limited precision
- added possibility to write basis file in original Cplex definition
- new parameter bool:ratfac to control whether rational factorization is attempted during iterative refinement
- new parameter int:ratfac_minstalls for minimum number of stalling refinements since last pivot to trigger rational
factorization
- new parameter bool:ratrec to control whether rational reconstruction is attempted during iterative refinement
- new statistics section on rational reconstruction
- new statistics section on size of rational solution in base 2 and base 10
- new parameter real:ratreq_freq for geometric frequency at which to apply rational reconstruction
- new parameter bool:powerscaling for rounding scaling factors during iterative refinement to powers of two
- new parameter bool:ratfacjump whether to continue iterative refinement with exact basic solution if not optimal
- new methods to store and set dual norms for pricing
- new parameter real:minred to control the minimal reduction factor in the simplifier to continue presolving (default 1e-4)
- new parameter bool:rowboundlfips to enable bound flipping also for row representation
fixed bugs:
- do not categorically remove nonzeros with small absolute value in simplifier
- allow empty lines in MPS files
- fixed default bounds of integer variables in MPS reader
- adjust basis status of nonbasic rows automatically to row type when reading basis file
- fix compiler error from missing operators between Rational and ints for GMP=false
- fix memory leak with static rational numbers Rational::{ZERO,POSONE,NEGONE} and list memory
- fix uninitialized value in class SPxWeightST
- fix wrong basis status in postsolving of aggregated variables
- eliminate incorrect switching of primal and dual feasibility tolerance when changing representation
- force recomputation of nonbasic objective value when reloading SPxSolver
- correct counting of LU solves (when solving with more than one rhs this is counted accordingly)
code quality:
- change flag DEBUGGING to SOPLEX_DEBUG
- deactivate SOPLEX_DEBUG flag in header files if defined externally
- SPxOut output stream is no longer a global variable
- hyper pricing parameters (i.e. candidate list size and activation threshold) are not linked to maxUpdates of factorization anymore
18 Feb 2015 - Release 2.1.0
===========================
SoPlex 2.1 is an internal release that was packed for installing an 80bit extended-precision version on the NEOS Server
for Optimization. It contains many of the features of SoPlex 2.2, in particular for solving LPs with rational input
data exactly.
18 Dec 2014 - Release 2.0.1
===========================
features:
- restore basis of original LP after performing iterative refinement on feasibility LP
performance:
- more efficient memory allocation of elements of class Rational (can be deactivated at compile time by defining
SOPLEX_NOLISTMEM or by Rational::disableListMem() at run time)
- more efficient assignment operators for class Rational
- return computed primal and dual activities by reference
- Real and Rational arguments passed by reference in changexyz methods
interface:
- new method Rational::freeListMem() to free unused GMP memory manually (usage optional)
- more statistics on type of refinement rounds
- type of timing can now be switched between CPU time and wallclock time
fixed bugs:
- fix Rational::isAdjacentTo() when GMP=false
- fix MPS reader for files without blank line at EOF
- fix pointer of activepricer in copy constructor and assignment operator of auto pricer SPxAutoPR
- fix bug from uninitialized Farkas vector when simplifier detected infeasibility during iterative refinement
- fix bug when perturbing fixed variables
- fix bug with incorrect index incrementation in sparse inner vector product
- fix compiler error with class Rational::Private and GMP=false
27 Feb 2014 - Release 2.0.0
===========================
new interface class SoPlex:
- (optional) handling of rational LPs: exact parsing, extended-precision solutions, exact computation of violations,
parallel handling of rational and floating-point versions of LP and solution data
- improved iterative refinement algorithm, handling also unbounded and infeasible LPs
- new command line interface
- customization via parameter system
- compatibility mode: old class SoPlex available as SoPlexLegacy; compile with make LEGACY=true to typedef
SoPlexLegacy to SoPlex and effectively use the 1.x interface
further interface changes:
- renamed class MpqReal to Rational
- new simplifier method to return result of the last simplification
- new SPxSolver status INForUNBD
- own objective sense enumerator in MPS parser
- improved iteration log
new features & performance:
- faster loops in vector classes
- implemented hyper sparse pricing (reduces the length of the 'for'-loops in pricing by only checking a
short list of good pricing candidates and all updates/changes introduced with the previous iteration)
- implemented bound flipping ratio test for row representation
- simplifier keeps relevant bounds if bound flipping ratio test is used
- new method to calculate the condition number of the current basis matrix (2-norm, estimate)
- new auto pricing switching from devex to steep after a set number of iterations
fixed bugs:
- refinement step to improve feasibility/optimality also for steep pricer (was devex only)
- allow instable enter pivot when no other is available; only after a fresh factorization
- ensure that problem is not perturbed when concluding optimality
- fixed numerical bug in unsimplification of doubleton equations
- (co)solve calls in SPxBasis do not trigger factorization() of SPxSolver but of SPxBasis
- fixed bug in unShift(): shifted bounds were not set to their original value even if the corresponding value was within
the original bounds (feasibility tolerance was checked in wrong direction)
- fixed behaviour in instable leave/enter: pricing values are now restored, before the instable pivot is performed
- do a better check in precisionReached() by first recomputing correct test/coTest/fTest values
- fixed MSG system to prevent unnecessary executions/operations in non-printed calls
- throw exception if immediately after a restart the basis matrix is infeasible
- correct update of infeasibilities after dual bound flips
code quality:
- new/delete calls are modified to channel memory (de)allocation through spx_alloc calls (-> "placement new")
- fixed make install for older Mac systems where install command does not have option -t
- fixed installation of headers by make install
- replaced preprocessor define WITH_ZLIB by SOPLEX_WITH_ZLIB
- new python scripts for evaluating and comparing test runs
16 Oct 2013 - Release 1.7.2
===========================
fixed bugs:
- fixed minor bug in SVSet::add()
- restore slack basis if a basis with too many or too few basic entries is loaded
- higher precison when writing .lp or .mps files, similar to SCIP standards
04 Jan 2013 - Release 1.7.1
===========================
code quality & new features:
- unfied members/methods of CLUFactor in new file clufactor.cpp (contains factor.cpp, forest.cpp, update.cpp, (v)solve.cpp)
- moved code from message.h,cpp to spxdefines.h,cpp to avoid name clashes with SCIP
fixed bugs:
- fixed inconsistent generic column and row names
- fixed bug in adjusting basis statuses to sides/bounds when loading or reloading the basis
- fixed bug in refactorization in the case when update detects singularity: invalidate only the factorization, not the
matrix since this would change the order of colums/rows in the basis
- catch possible exception when finally performing basis update
- fixed inconsistency in "Fast" ratiotester when "short" leave is possible
- fixed bugs in unsimplification of zero objective singleton columns and row singletons
- relaxed wrong assert in SPxSolver::unShift()
- fixed bug in ratio test: consider only nonbasic entering variables in leaving algorithm, also in row representation
- reverted wrong bugfix (99f913d7) in computeLeaveCoPrhs4Col() for column status D_UNDEFINED
interface:
- new methods in SPxSolver to set and get maximum number of updates before next refactorization
- implemented writing of basis file for row representation
- implemented Farkas ray (for infeasible LPs) and primal ray (for unbounded LPs) for row representation
31 Jul 2012 - Release 1.7.0
===========================
code quality & new features:
- partial pricing can be switched on for dantzig-, devex- and steep-pricing by defining PARTIAL_PRICING (only LEAVING mode)
- current basis matrix can be written to a file in MatrixMarket format using SPxBasis::printMatrixMTX
- exact initialization of steepest edge weights possible by calling setupWeights of SPxSteepPR
- option -bw writes the basis of the original problem to a file; works now with simplifier as well
- allowed primal and dual feasibility tolerance to be set independently in simplifier and simplex routines
- added GMP support and data type MpqReal, a wrapper for mpq_class if compiled with GMP=true, soplex::Real otherwise
- implemented iterative refinement for feasible and bounded LPs, applied when a feasibility or optimality tolerance below
the new parameter SPxSolver::irthreshold() is requested
- reduced and commented code by adding "void deleteVec(DLPSV* ps);" method in svset.{h,cpp}
fixed bugs:
- variables that do not appear in the objective function or any constraint are not printed anymore in the BOUNDS section
of the mps file
- allow instable leave iterations, i.e., leave iterations with pivot element below default stability threshold, only
after a fresh factorization
- check redundant variable bounds implied by rows more carefully: do not apply if max/min activity exceeds sides by a
factor greater than 1.0/epsilon to avoid roundoff errors
- initialize basis descriptor to statuses D_UNDEFINED in order to avoid access of uninitialized values at time out
- fixed bug in stalling detection: SoPlex incorrectly assumed that pricing had been finished and hence claimed optimality
- handled setting FIXED basis status when bounds are only approximately equal
- using a starting basis from file disables simplifier
- fixed bug in computeLeaveCoPrhs4Col() for column status D_UNDEFINED
- fixed bug in unsimplifying singleton rows
- fixed cycling caused by refactorization after pricer finds no more candidates
- fixed bug that bound changes were not performed if new bound deviates by only epsilon; this can lead to slightly
inconsistent bounds which trigger asserts later
- sparse pricing stores only infeasibilities > entertol() in leaving or > leavetol() in entering (was epsilon() before)
- fixed false infeasibility detection in ratiotest after fresh factorization
- fixed bug in the svset data array, when removing the first vector, the memory was not 'freed' and therefore the memory
consumption could have increased to a very huge value (bug 319)
- also fixed wrong ensureMem() call in SVSet create(), which lead to a to big data array and change an unnecessary
insert() call to an reSize() call
- corrected a memcpy() to a memmove() call due to possible overlapping memory
- fixed bug in counting unsuccessful pivots
interface:
- info message in case a fresh factorization is not stable enough and has to be done again with higher Markowitz threshold
- new command line option -p6 to use steepest edge pricer with exact initialization
- verb levels v4 and higher display current pricing mode (sparse or conventional)
- additional simplify() method with independent primal and dual feasibility tolerance as parameters
- new exception SPxInterfaceException to be thrown if inconsistent parameters are passed to interface methods
- new methods feastol(), opttol(), setFeastol(), setOpttol() in classes SoPlex and SPxSolver to get/set primal and dual
feasibility tolerance; delta() and setDelta() still exist and work as before
- new command line options -f and -o to set primal feasibility and optimality tolerance, respectively; -d still exists and
works as before
- exact steepest edge is the new default pricing strategy and is implemented as separate class (inherited from SPxSteepPR)
- correct dual statuses when loading a basis with SPxBasis::loadDesc(); they might have become inconsistent when bounds
or sides had been changed
- new methods SPxLP::computePrimalActivity() and SPxLP::computeDualActivity() for given primal and dual solution vectors
using exact arithmetic
- new methods irthreshold() and setIrthreshold() for threshold parameter below which iterative refinement is applied
- new command line option -R to set parameter irthreshold()
performance:
- sparse pricing for leaving and entering simplex is activated depending on ratio of basis dimension and number of infeasibilities
- fixed performance issue removing (all) vectors from a SVSet where it is better to delete the elements from behind to
the front
28 Oct 2011 - Release 1.6.0
===========================
code quality & new features:
- time for basis updates can now be measured by defining flag MEASUREUPDATETIME in spxbasis.h (off by default)
- removed "-static" option from Makefile due to linking troubles on newer systems without static libraries; can still be
added using the make option "USRLDFLAGS=-static"
- added -fp:precise and -Oi to make.mingw.x86.{cl,intel}.opt
- added define SOPLEX_SUBVERSION to track internal development version (src/spxdefines.h)
- added SHARED variable in Makefile for creating shared libraries and binaries
- added functionality to solve systems with three right hand sides
- added partial quicksort to sorter.h
- increased priority of debugging output
- added "install" target to Makefile; this target can be used to install SoPlex into certain directory on the system
(works for Linux and Mac); for this, you have to specify INSTALLDIR (make install INSTALLDIR=/urs/local/)
- removed compiler flag NO_ADDITIONAL_CHECKS in code and makefile options: now, additional checks are only performed
when ENABLE_ADDITIONAL_CHECKS or ENABLE_CONSISTENCY_CHECKS is explicitly defined (globally or in a specific file)
- githash of current commit is displayed (also in check mode)
interface:
- LP reader: print warning when encountering a floating point number with empty exponent (hints at variable name with
initial 'e' or 'E')
- methods selectEnter() and selectLeave() in ratio testers now receive index/id of leaving/entering variable as input
parameter
- added methods to solve the linear systems of equations for 3 right hand sides simultaneously
- added method getPrimalray() to classes SoPlex and SPxSolver for primal unbounded LPs
- added command line option '-L' to set iteration limit
- allowed free basis status (SPxSolver::ZERO, SPxBasis::Desc::P_FREE) for variables and constraints with finite bounds
- adapted method to read/write basis for free nonbasic variables
- added command line option '-t3' to use the bound flipping ratio test (long step dual)
- number of performed bound flips during ratio test are displayed as 'flips=' in line 'Finished solving...'
- added method to clear starting basis in class SoPlex (SoPlex::clearBasis())
performance:
- increased default Markowitz threshold to 0.04 to increase numerical stability of the factorization
- implemented bound flipping ratio test ("long step dual") for column representation
fixed bugs:
- added missing -m64 in LDFLAGS of make.{darwin,sunos}.x86_64.gnu.* and make.{linux,freebsd}.x86_64.gnu.opt*
- added missing -DNDEBUG in make.mingw.x86.{cl,intel}.opt
- fixed incorrect access to member variables in constructors of class DataHashTable
- fixed bug in SPxFastRT: also in the first phase consider only nonbasic entering variables in COLUMN/LEAVING mode
- fixed bug in COLUMN/ENTERING algorithm yielding incorrect detection of unboundedness or inverted primal ray: when
deciding whether a free variable entering the basis should increase or decrease from zero, we have to look at the
reduced cost, not the objective function
- fixed too early check of iteration limit (now not checked in SPxSolver::terminate() anymore, but at beginning of each
iteration)
- fixed bug in unsimplifcation of duplicate rows presolving
- fixed bug when setting basis through class SoPlex
- fixed bug occuring with large column/row bounds: if we refactorize, we also have to recompute the right hand side of
Bx=b, because roundoff errors might have occured during its update
- fixed segmentation fault in simplifier when postsolving is interrupted with an exception
- fixed wrong error output in statistics when an instance is solved during presolve
- pricer SPxDevex now performs one additional pivot if only slightly feasible/optimal; this helps to prevent suboptimal
results on some unsimplified instances
- fixed potential NULL pointer dereference in nameset.h, when name is not found in hashtable
- fixed potential segmentation fault when calling sorting routines on empty arrays (sorter.h)
- send debugging output to standard out instead of error stream
- fixed bug in unsimplification of duplicate columns (do not replace two linearly dependent columns by one if free basis
status can not be resolved in unsimplification)
- fixed several bugs when unsimplifying infeasible bases
- fixed counting of removed rows during simplifier
30 Sep 2010 - Release 1.5.0
===========================
code quality & new features:
- changed memory handling in class DataKey: replacing bitfield implementation by two int's
- improved handling of singular bases and cycling: automatically restart from initial basis with tighter tolerances
- major rework of simplifier SPxMainSM: "localized" postsolving steps; several bugfixes concerning basis statuses, dual
values and reduced costs
- performance improved in simplifier SPxMainSM: always processing singleton rows before finding duplicate rows; ignoring
singleton rows and columns when checking for linear dependency
interface:
- submakefile for FreeBSD added
- added command line option '-y' to output dual row multipliers
- additional interface method SoPlex::setBasis to set starting basis if no simplifer present
fixed bugs:
- bug in class SoPlex appearing when LP is solved completely in simplifier and optimal solution retrieved
- bug in spxbasis.cpp: when basis is detected as singular during update step, now we refactorize to double check; if the
singularity is not resolved we exit with correct status
- bug in simplifier postsolving: empty columns may not only be fixed to bounds but also to zero for free variable
- bug in SPxSolver::setPricer() appearing when switching pricers in initialized state
- bug in simplifier SPxMainSM yielding incorrect dual values and reduced costs for maximization problems
- bug in simplifier SPxMainSM for postsolving of free columns with zero objective
- bug in factorization triggered by column singletons with nonzeros below numerical zero tolerance
11 Sep 2009 - Release 1.4.2
===========================
code quality & new features:
- in class SoPlex, enabled getBasis... methods also when simplifier is present and basis suboptimal, but regular
- added check for stalling in main solve loop SPxSolver::solve(): if objective value and shift haven't changed after a
large number of iterations, switch algorithm; this fixes a timeout of SCIP/SoPlex on instance glass4 (MIPLIB2003)
interface:
- small change in spxmpswrite.cpp: always write the default upper bound of 1.0 for integer variables, since CPLEX seems
to set infinity otherwise
- added method 'std::string statistics() const' to classes SoPlex, SPxSolver, SPxBasis, SLinSolver, SLUFactor, returning
statistical information like number of factorizations etc.
- added method 'Real cumulativeTime() const' to class SPxSolver returning time used in all calls to method solve() since
beginning or last call of 'resetCumulativeTime()'
- implemented copy constructors and assignment operators for almost all classes, especially SoPlex and SPxSolver
- the set methods (for pricer, ratiotester, ...) in classes SoPlex and SPxSolver now have an optional boolean parameter
(false by default) controlling whether the memory of the object handed over shall be managed internally; this does not
affect the default behaviour of the interface methods
- the constructor of class SPxSolver does not have parameters for pricer, ratiotester and starter, anymore; use the set
methods instead
fixed bugs:
- now correctly resetting basis status to REGULAR in method SPxSolver::loadBasis()
- memory bug in function memPack() in svset.cpp; this method is now called to clean up memory, in case too much memory
is wasted; this removes a large increase of the use of memory as tested in tests/testMemory.cpp
- bug in formulas for triggering refactorizations after increased relative/absolute fill
- memory leak in class SPxMainSM when calling simplify() multiple times
- bug in class DataKey leading to a segmentation fault when reading LPs with more than 8,388,607 variables or
constraints: now the maximum number allowed is 536,870,911 and asserted in debug mode
- numerical bug in SPxFastRT leading to failed assert 'rep() == ROW' (in leave.cpp): in column/leaving algorithm, now we
only consider nonbasic variables for entering the basis
20 Feb 2009 - Release 1.4.1
===========================
code quality & new features:
- speed up of the simplifier due to combined removal of rows/columns and combined updates of lhs and rhs in some cases
- overworked spxfastrt.h, spxfastrt.cpp
- added method isDescValid() in class SPxBasis to check if a descriptor is valid w.r.t. to the bounds of the loaded LP;
use this to check whether previous/given basis can be loaded in SoPlex::solve()
interface:
- small changes in spxlpfwrite.cpp: make bounds LP-format conform, output 'generals' only if integer variables present
- added new method writeState() to classes SoPlex and SPxSolver for dumping LP, basis, and parameter settings at once
fixed bugs:
- bug in simplifier spxmainsm.cpp: wrong init for postsolving free column singletons fixed
- wrong bugfix in spxsolver.cpp:maxInfeas(): the leave and entering case were switched back
- endless loop triggered by different values of feasibility/optimality tolerance for pricer in spxsolve.cpp
- memory handling in class UnitVector
- bug when resolving in SoPlex::solve() after change of LP
30 Sep 2008 - Release 1.4.0
===========================
code quality & new features:
- split the main() method in example.cpp into several smaller methods
- created directory 'tests' that should contain small programs that illustrate a (former) bug in SoPlex or something
like unit tests
- added member function precision() in spxout.h
- changed main target to soplexmain (from example)
- soplexmain.cpp now contains the main function, the output of this function should be more readable than the one of
example; original output can be turned on with the "-C" command line argument, the example is now available as target
'example' in the makefile
- name unification:
+ moved SPxLP::readFile() from spxfileio.cpp to spxio.cpp
+ renamed SPxLP::dumpFile() to SPxLP::writeFile(), included default params, moved it from spxlpfwrite.cpp to spxio.cpp
- simplified postsolving in spxmainsm.cpp (do not recompute reduced costs)
- overworked spxbasis.cpp:readBasis(): fixed bug and removed superflous code at end
- added fixing of the basis status if ZERO/P_FREE is not appropriate anymore, see varStatusToBasisStatusCol() and
varStatusToBasisStatusRow() in spxsolver.cpp
- removed superfluous variable 'count' in spxparmultpr.cpp/h
- modifications in the simplifier SPxMainSM to avoid numerical problems
- modified reading and writing of basis: the row and column names are now passed as pointers; default names are
entered if the pointers are 0; the changes appear in soplexmain.cpp, spxbasis.cpp, soplex.cpp, soplex.h
interface:
- introduced exceptions for all unrecoverable errors (lack of memory, no pricer, singular basis, etc.) to allow a
calling application to clean up
- command line options "-V" and "-h" of example solver now output info on compilation flags
- replaced SPxLP::operator<< by SPxLP::writeLPF()
- implemented basic handling of reading a basis file (not yet thoroughly tested); the basic changes are in example.cpp,
spxbasis.cpp, soplex.cpp, soplex.h
- added option to produce shared libraries with x86-Linux (OPT=opt-shared)
- constructor interface for the scalers SPxEquiliSC and SPxGeometSC changed: parameter colFirst is removed since the
corresponding member is not used; printed names of the scalars are slightly changed; there are only 5 command line
options for the scalar left since two of the previous 6 options coincided
fixed bugs:
- two bugs in spxlpfread.cpp: variables appearing in the bounds and general/integer/binary section
- bug in changeElement(): it now works also in the initial phase (NO_PROBLEM); aA corresponding test file is
'tests/testChangeElement.cpp'
- bug in forest.cpp: use proper array access instead of pointer iteration
- bug concerning the termination value (objective limit) in spxsolve.cpp
- memory leak in postsolving of spxmainsm.cpp
- bug in spxchangebasis.cpp, addedRows(), addedCols(): need to call loadMatrixVecs in all cases, since the references to
the matrix might have been invalided due to reallocation
- bug with detecting infeasible instances if scaling theCoTest does not help
- bug in spxlpfread.cpp with bounds on binary (old bounds were discarded)
- bug in spxsolver.cpp:maxInfeas(): the leave and entering case were switched
- bug in SLUFactor::solveRight4update(): now variable n is storing ssvec's size correctly
- bug in leave.cpp, left over from previous correction dealing with infeasible instances
27 Aug 2007 - Release 1.3.2
===========================
interface:
- keyword RANGES added in MPS format output (spxmpswrite.cpp)
- extended FAQ
fixed bugs:
- bug in spxmainsm.cpp concerning index mapping
- bug in unsimplify() if the problem has vanished in preprocessing
- missing check for singular basis at the end of factorize() in spxsolver.cpp
- assertion bug in FreeZeroObjVariablePS::execute() of spxmainsm.cpp
- bug in SVector::dim(): has to return the dimension and not the maximal index
21 Sep 2006 - Release 1.3.1
===========================
code quality & new features:
- made writeBasis() and writeBasisFile() in SoPlex and SPxSolver const-methods
- added SPxLP::writeLPF() with some parameters to dump an LP in .lp-format with variable and row names; this also fixes
some output bugs; the still existing SPxLP::operator<< uses self-created default names
- check for necessary rounding added to spxmainsm
interface:
- restructured example.cpp and added a small example on constructing an LP using the callable library methods
fixed bugs:
- bug in spxmainsm: IDs of rows and columns are not used anymore
- bug in spxmainsm in connection with free variables
- bug in soplex method unsimplify(): check whether problem vanished in preprocessing added
- bug in spxlpfreader in connection with strange input (unconstrained rows)
- bug in ssvector: use ssvec.size() instead of b.size() (the former may be smaller due to flattened near-zero values)
03 Feb 2006 - Release 1.3.0
===========================
code quality & new features:
- a lot of stuff we have already forgotten about
- new preprocessing
- the number of warnings issued for ignored entries in MPS files is now limited to 1000 per file
- improved stability of textbook ratio tester
- added reference solution files for further LP suites
- extensive testing: coverage analysis and valgrind; new binary "exercise_LP_changes" performs kind of unit tests for LP
change interface
- some assertions may be converted to generate warnings if NDEBUG is given; compile with "-DWITH_WARNINGS" to achieve this
- revised the whole documentation and made it doxygen-compatible
- made one-argument constructors explicit, added virtual destructors, blocked copy constructors and assignment operators
- rewrote major parts of ssvector.cpp to avoid pointer arithmetic
- added the spxout class to redirect output depending on the error level, provided all messages with a unique code for
finding it with grep
- improvement of the LPF and MPS readers
- dropped all unused functions from the SLinSolver interface
interface:
- status output is now human-readable
- methods obj(), rhs(), lhs(), upper(), lower(), maxObj() of SPxLP that returned a non const reference renamed to *_w and
made protected; use changeXXX() methods outside the class
- methods nofCols(), nofRows(), nofNZEs() deleted; use nCols(), nRows(), nNzos() instead
- methods obj(), rhs(), lhs(), upper(), lower(), maxObj() of LPColSet and LPRowSet that returned a non const reference
renamed to *_w
- LPColSet: renamed obj() methods to maxObj()/maxObj_w(), renamed writable lower(), upper() methods to
lower_w()/upper_w(), removed SPxLP::obj()
- LPRowSet: renamed writable lhs()/rhs() methods to lhs_w()/rhs_w()
- SPxSolver has got an additional NOT_INIT (not properly initialized) status
- renamed SPxDefaultPR to SPxDantzigPR
fixed bugs:
- bug in SPxSumST
- bug in memory allocation of SVSet: now uses doubling strategy correctly; the bug led to serious performance
degradation on large LPs under Windows, which is now fixed
- some bugs in LP change interface
- segmentation fault in Forrest-Tomlin update
- bug in SSVector where zero vectors were not handled properly
- bug in example.cpp which resulting in ignoring the update type switch
10 Apr 2002 - Release 1.2.1
===========================
- interface of SoPlex::terminate() changed
- SoPlex::setRep(int) changed to SoPlex::setRep(Representation)
16 Jan 2002 - Release 1.2.0
===========================
(empty)
|