1 NLopt 2.3.1 (16 September 2012)
3 * Bug fix: COBLYA should return ROUNDOFF_LIMITED rather than XTOL_REACHED
4 if the trust-region radius underflows to zero; thanks to David Liu.
6 * Bug fix: incorrect return value from set min/max objective.
8 * Handle case of negative rescalings (from negative dx) in COBYLA and BOBYQA;
9 thanks to Alexander Law for the bug report
11 NLopt 2.3 (20 July 2012)
13 * In Matlab/Octave interface, make returning NaN from the objective
14 equivalent to nlopt_force_stop; thanks to Norman Violet for the suggestion.
16 * Added CCSA-quadratic (NLOPT_LD_CCSAQ), similar to MMA.
18 * Added interface for supplying a preconditioner (approximate Hessian);
19 currently only supported in CCSAQ.
21 * When adding mconstraints, allow tol==NULL as synonym for zero tolerances.
23 * Added missing NLOPT_LD_SLSQP constant in Matlab/Octave.
25 * Lower tolerance for dual optimization in MMA/CCSAQ; thanks to
26 Christophe Leruste for the problem report
28 * Fixed bug in timer, thanks to William Vaughn for the patch.
30 * Bug fix to convergence test in sbplx; thanks to Douglas Bates.
32 NLopt 2.2.4 (9 June 2011)
34 * Fixed linking problem for C++ and Python shared libraries on MacOS X;
35 thanks to Volker Lorrmann for the bug report.
37 NLopt 2.2.3 (8 June 2011)
39 * Fixed additional re-entrancy problem in Luksan algorithms missed
40 in the previous version; thanks to Gert Wollny for the bug report.
42 * Fixed set/get vector_storage in Fortran interface for Fortran
43 compilers with all-uppercase linkage.
45 NLopt 2.2.2 (26 May 2011)
47 * Added set_vector_storage API to modify the memory usage and the
48 subspace dimension for low-storage quasi-Newton methods.
50 * Fixed missing support for maxtime stopping criteria in Luksan and
51 ORIG_DIRECT algorithms; thanks to Jurgen Werner for the bug report.
53 * Fixed algorithms to support the case where the lower and upper bounds
54 are equal for some variables (which effectively eliminates those
55 variables from optimization).
57 * Added missing xtol check to SLSQP, which caused erroneous ROUNDOFF_LIMITED
58 error codes to be returned; thanks to Alexander Riess for the bug report.
60 * Fixed slight overcounting of function evaluations for maxeval check
63 * Fixed deprecated API to support xtol_abs == NULL for backward
64 compatibility (thanks to Francesco Biscani for the bug report).
66 * Made Luksan algorithms (e.g. LBFGS and other quasi-Newton methods)
67 re-entrant; thanks to Gert Wollny for the bug report. (Fixes
70 * Fixed bug in DIRECT (not ORIG_DIRECT), where a typo caused suboptimal
71 convergence in some cases; thanks to Sinisa Hristov for the bug report.
73 NLopt 2.2.1 (6 September 2010)
75 * If you compile nlopt.h with the NLOPT_DLL_EXPORT symbol #defined,
76 it now uses the dllexport directive (under Windows), useful for
77 compiling an NLopt DLL under Microsoft compilers; thanks to Benoit
78 Scherrer for the suggestion.
80 * Handle case where copysign function is missing, e.g. on Windows;
81 thanks to Benoit Scherrer for the bug report.
83 * Remove C99-style mixed declaration and code in a couple files, so
84 that code compiles in C89; thanks to Benoit Scherrer for the bug report.
86 * Removed a few compiler warnings under Microsoft compilers; thanks
87 to Benoit Scherrer for the bug report.
89 * Export nlopt_get_algorithm_name function on Windows; thanks to Ofek
90 Shilon for the bug report.
92 * Don't use dllimport directive with lcc on Windows (which doesn't
93 support it); thanks to Laurent Vanbeylen for the bug report.
95 * Update Nodedal README directory to indicate that Nocedal's LBFGS code
96 is now available under the GPL, and therefore may be distributed with
97 a future NLopt version (although Luksan's LBFGS code already works well).
99 * Bug fix in set/get_xtol_abs; thanks to David Rivest-Henault for the report.
101 NLopt 2.2 (15 July 2010)
103 * Added SLSQP algorithm for gradient-based local optimization with
104 nonlinear constraints via sequential quadratic programming, based
105 on the implementation by Dieter Kraft that was adapted for SciPy.
107 * Modified BOBYQA and COBYLA algorithms to support unequal initial
108 step sizes in different directions; thanks to Tom Fiddaman for pointing
109 out the need for this in the case where different directions have
110 very different scales.
112 * Added Python module docstring; thanks to Sebastian Walter for the suggestion.
114 * Added GUILE_INSTALL_DIR variable to allow the user to change the
115 Guile installation directory.
117 * Added Fortran interface for vector-valued constraints.
119 * Throw correct exceptions in Python for the add_*constraint functions;
120 thanks to Dmitrey Kroshko for the bug report.
122 * Support forced stop and exceptions in ORIG_DIRECT algorithm.
124 * Remove arbitrary 1e20 upper bound on function values from ORIG_DIRECT
127 * Bugfix in C++ interface (and some other language front-ends) when
128 deallocating the nlopt_opt object in cases like MLSL where
129 local_optimizer is used; thanks to Jurgen Werner for the bug report.
131 NLopt 2.1.2 (8 July 2010)
133 * The Python mconstraint (vector-valued constraint) functions
134 now pass a 2-dimensional array for the gradient argument, rather
135 than a flattened 1d array.
137 * Improved handling of exceptions and forced stops for constrained
138 optimization, making sure that no constraints are evaluated after
141 * Return an NLOPT_INVALID_ARGS error if more than n equality constraints
142 are added in an n-dimensional problem.
144 * Fix bug that could cause spurious NLOPT_INVALID_ARGS errors when
145 adding constraints under rare circumstances.
147 * Eliminate a few small memory leaks that could occur under error conditions.
149 NLopt 2.1.1 (7 July 2010)
151 * More robust configure check for Python include directories, via
152 distutils.sysconfig.get_python_inc() and numpy.get_include().
153 Thanks to Nathaniel Smith for the tip.
155 * Bug fix in Guile interface: added missing prefix to nlopt-version-major
158 NLopt 2.1 (6 July 2010)
160 * New vector-valued constraint feature; thanks to Dmitrey Kroshko of OpenOpt
163 * COBYLA now accepts equality constraints.
165 * Guard against multiple inclusion in nlopt.hpp; thanks to Saul
166 Thurrowgood for the suggestion.
168 NLopt 2.0.2 (17 June 2010)
170 * Fixed compilation failure in Microsoft Visual Studio, due to
171 incorrect usage of __stdcall keyword; thanks to Dave Katz for the
174 NLopt 2.0.1 (16 June 2010)
176 * Bug fix in Fortran API (for nlo_get_ functions returning arrays).
178 * Fixed buggy compilation with MinGW.
180 NLopt 2.0 (15 June 2010)
182 * New C API, that works by creating an nlopt_opt "object" and then calling
183 functions to set the optimization parameters -- much more extensible
184 than the old API (which is preserved for backwards compatibility).
185 (Updated Fortran, Matlab, and GNU Octave wrappers as well.)
187 * C++ nlopt.hpp wrapper around C API, allowing namespaces, object
188 constructors/destructors, std::vector<double>, and exceptions
191 * New nlopt wrappers callable from Python and GNU Guile, generated
192 with the help of SWIG.
194 * New 'man nlopt' manual page documenting new API.
196 * New AUGLAG algorithm(s) implementing an augmented-Lagrangian method
197 proposed by Birgin and Martinez (2008), which supports nonlinear
198 equality and inequality constraints "wrapped around" other
199 local/global optimization methods.
201 * Added API for nonlinear equality constraints (currently only
202 supported by AUGLAG and ISRES algorithms).
204 * Support inequality constraints directly in ORIG_DIRECT algorithms
205 (no need to return NaN when constraint is violated).
207 * Inequality/equality constraints now have optional tolerances that
208 are used as conditions in stopping criteria.
210 * Pseudo-randomize simplex steps in COBYLA algorithm, improving robustness
211 by avoiding accidentally taking steps that don't improve conditioning
212 (which seems to happen sometimes with active bound constraints). The
213 algorithm remains deterministic (a deterministic seed is used), however.
215 * Allow COBYLA to increase the trust-region radius if the predicted improvement
216 was approximately right and the simplex is OK, following a suggestion
217 in the SAS manual for PROC NLP that seems to improve convergence speed.
219 * Added nlopt_force_stop function to force a (graceful) halt to
220 the optimization, and corresponding NLOPT_FORCED_STOP return code.
222 * Improved thread-safety in random-number generation: thread-local
223 storage is used for random-number state, on compilers that support
224 it (e.g. gcc, Intel, Microsoft), to make the generation thread-safe.
225 In this case, the random-number seed must be set per-thread.
227 * Return an error in global-search algorithms if the domain is not finite.
229 * Use stdcall convention on Windows; thanks to Alan Young for the suggestion.
231 * Added missing absolute-tolerance criteria in Luksan algorithms; thanks
232 to Greg Nicholas for the bug report.
234 * Fixed compilation under C++, and use C++ compiler for everything in
235 --with-cxx mode; thanks to Greg Nicholas for the bug report.
237 * In MMA, only stop at minf_max/stopval if the point is feasible.
239 * Fix Matlab mex file to not include unnecessary nlopt-util.h file,
240 simplifying Windows compilation.
242 NLopt 1.2 (18 November 2009)
244 * Added Powell's BOBYQA algorithm for box-constrained optimization
245 without derivatives, an improvement on NEWUOA.
247 * Added ISRES genetic algorithm, supporting nonlinearly constrained
250 * New functions nlopt_{set/get}_stochastic_population to provide
251 optional greater control over the random "population" sizes in
252 stochastic algorithms (although it still has a sensible default).
254 * Bug fix: remove extraneous text accidentally included in nlopt.f Fortran
257 * Bug fix: configure script now correctly handles Matlab installation
258 when MEX_INSTALL_DIR is specified manually by the user.
260 NLopt 1.1 (12 November 2009)
262 * configure script detects whether --enable-shared is required
263 in order to compile Matlab and Octave plugins (as is the case
264 on x86_64), and disables compilation of those plugins if
265 --enable-shared is not used.
267 * Added --without-octave and --without-matlab configure options to
268 disable Octave and Matlab plugins, respectively.
270 * Modified COBYLA algorithm to have better support for bound
273 * Added new NLOPT_ROUNDOFF_LIMITED failure code to indicate
274 cases in which optimization breaks down due to roundoff errors,
275 in which case it is possible that useful results were obtained.
277 * Experimental support for nonlinear equality constraints via
278 augmented-Lagrangian method.
280 * Support for compiling under Windows (ideally with MinGW) as a
281 DLL, although you have to manually add #define NLOPT_DLL
282 to nlopt.h *after* installing (after compiling NLopt).
284 * Added several checks for roundoff-related breakdown to NEWUOA code.
286 * When only a relative error tolerance is specified, no longer
287 fails to halt when exact convergence to zero is obtained.
289 * Workaround for incompatible qsort_r functions in BSD and GNU libc
290 by always using my own version; thanks to Wendy Vandoolaeghe
291 and Philippe Preux for the bug report and explanation.
293 * Workaround for gcc 3.4.x conflict with HUGE_VAL definition in Solaris
296 * Better identification of Matlab-plugin installation directory.
298 * Fixed identification of Octave-plugin installation directory for
299 recent Octave versions.
301 NLopt 1.0.1 (13 Nov. 2008)
303 * Allow user to override Matlab-plugin installation directory with
306 * Bug fix in my DIRECT code that prevented convergence (DIRECT-L unaffected).
308 * MLSL needs a nonzero default ftol_rel and/or xtol_rel to ensure that
309 its local searches terminate; use roughly machine precision as defaults.
311 NLopt 1.0 (11 Nov. 2008)
313 * Initial public release.