CHANGELOG

2023-06-14  Giacomo Nannicini <g.nannicini@usc.edu>
	Documentation update.
	Version 4.2.6 released.
	
2022-11-20  Giacomo Nannicini <g.nannicini@usc.edu>
	Increased parallelization efficiency during initialization phase.
	Version 4.2.5 released.
	
2022-03-16  Giacomo Nannicini <nannicini@us.ibm.com>
	Changed order of imports to try to enforce OMP_NUM_THREADS=1 when
	testing.
	Version 4.2.4 released.

2022-02-25  Giacomo Nannicini <nannicini@us.ibm.com>
	Version 4.2.3 released.
	
2022-02-24  Giacomo Nannicini <nannicini@us.ibm.com>
	Added tox to the test workflow.
	Moved tests to nose2.
	* rbfopt_algorithm.py (optimize_parallel): fixed bug that could
	occur when using many threads and the noisy function evaluatio.

2021-02-16  Giacomo Nannicini <nannicini@us.ibm.com>
	Version 4.2.2 released.
	* rbfopt_aux_problems.py (ga_optimize): fixed bug that ignored
	integrality restrictions in some cases where categorical variables
	were at the end of array of variables.

2020-08-13  Giacomo Nannicini <nannicini@us.ibm.com>
	Version 4.2.1 released.
	* rbfopt_settings.py (set_auto_parameters): fixed bug that turned
	on affine scaling on certain conditions.
	* rbfopt_utils.py (get_lhd_maximin_points, get_lhd_corr_points):
	fixed bug that modified bounds on integer variables.
	* Changed:
	rbfopt_utils.transform_domain() functions now require list of
	integer variables as one of the arguments.
	
2020-06-26  Giacomo Nannicini <nannicini@us.ibm.com>
	Version 4.2.0 released.
	* Added:
	Support for categorical variables (variable type 'C', in addition
	to 'R' for real and 'I' for integer). These variables are treated
	differently from integer variables, but they still need to have
	lower and upper bounds; e.g., a categorical variable with bounds
	[0, 2] can take on 3 possible values: 0, 1, 2.
	* Changed:
	Renamed parameter names starting with 'tr' to 'ref', for better
	consistency with the literature.

2019-04-08  Giacomo Nannicini  <nannicini@us.ibm.com>
	Version 4.1.1 released.
	* rbfopt_utils.py (initialize_nodes): fixed bug with sample_size
	not cast to integer.
	
2019-03-14  Giacomo Nannicini  <nannicini@us.ibm.com>
	Version 4.1.0 released.
	* Added:
	Finer control of the optimization process, via several options:
	init_include_midpoint, init_sample_fraction, do_local_search.
	* Changed:
	More efficient initialization phase, requiring fewer points.

2019-03-05  Giacomo Nannicini  <nannicini@us.ibm.com>
	* rbfopt_utils.py (initialize_nodes): more flexibility in choosing
	the number of sample points, and adding the midpoint.
	* rbfopt_utils.py (init_points_cleanup): added function to
	eliminate duplicate points in initialization.
	* rbfopt_algorithm.py (RbfoptAlgorithm): added several ways to
	customize the search: early start, skipping local search.
	* rbfopt_settings.py: added options init_include_midpoint,
	init_sample_fraction, do_local_search.

2018-10-06  Giacomo Nannicini  <nannicini@us.ibm.com>
	Version 4.0.3 released.
	* Added:
	More ways for the user to control the initial points and
	optimization process.
	Ability to control debug messages from solvers.
	* Changed:
	Format of the output log is more compact.
	Improved management of failed/discarded iterations.

2018-09-23  Giacomo Nannicini  <nannicini@us.ibm.com>
	* rbfopt_algorithm.py (update_log): changed log format.
	* rbfopt_algorithm.py (RbfoptAlgorithm): constructor now accepts
	do_init_strategy option.
	* rbfopt_algorithm.py (restart): added distance check for
	user-provided points.
	* rbfopt_settings.py : added options max_cycles and
	thresh_unlimited_refinement_stalled.

2018-08-29  Giacomo Nannicini  <nannicini@us.ibm.com>
	* rbfopt_algorithm.py (restoration_search_parallel): deleted.
	* rbfopt_algorithm.py (optimize_parallel): solved bug in parallel
	optimizer when evaluating RBF model (for temp_node_pos) after
	restarts.
	* rbfopt_algorithm.py (discarded_list): discarded points are now
	evaluated on a rolling basis.

2018-05-25  Giacomo Nannicini  <nannicini@us.ibm.com>
	Version 4.0.2 released.
	* Added:
	Detailed user manual.
	Improved testing tools.
	* Changed:
	Smaller formulation for the noisy interpolation models.
	Improved stability of the restoration step.

2018-05-24  Giacomo Nannicini  <nannicini@us.ibm.com>
	* rbfopt_algorithm.py (refinement_step): fixed bug in node
	deletion.
	* rbfopt_degreeX_models.py: reduced size of noisy interpolation
	models.
	* tests/test_rbfopt_algorithm.py: now only contains relatively
	fast tests; the slow tests have been isolated and are not run by
	default.
	* setup.cfg: added for fine-grained test definition.
	* setup.py: fixed possible infinite loop on Windows when using
	multiprocessing.
	
2018-03-09  Giacomo Nannicini  <nannicini@us.ibm.com>
	Version 4.0.1 released.
	* rbfopt_user_black_box.py: fixed naming of abstract methods.

2017-09-05  Giacomo Nannicini  <nannicini@us.ibm.com>
	Version 4.0.0 released. List of major changes:
	* Added:
	Option max_cross_validations to limit the total number of cross
	validations performed.
	* Changed:
	Improved numerical stability for nonlinear models in Gutmann's
	algorithm.
	Cross validation now uses Gaussian RBFs and different values of
	the shape parameter.

2017-08-31  Giacomo Nannicini  <nannicini@us.ibm.com>
	* rbfopt_cl_interface.py, rbfopt_test_interface.py, __init__.py:
	Set environment variable OMP_NUM_THREADS=1 to avoid problems with
	OpenBLAS parallel optimizer.
	* README: Updated documentation to reflect above change.

2017-08-24  Giacomo Nannicini  <nannicini@us.ibm.com>
	Version 4.0.0 alpha. List of major changes:
	* Added:
	Support for Gaussian radial basis function
	(rbfopt_degreem1_models).
	Automatic selection of the shape parameter
	(rbfopt_utils.get_best_rbf_model)
	* Changed:
	(RbfoptBlackBox) interface now requires returning var_type instead
	of a list of integer vars.
	(RbfoptBlackBox.evaluate_fast) is now called evaluate_noisy, and
	requires returning an estimate of the error.

2017-08-22  Giacomo Nannicini  <nannicini@us.ibm.com>
	List of major changes:
	* Added:
	Installation script setup.py.
	Class to construct RbfoptBlackBox from user data
	(rbfopt_user_black_box).
	The shape of the multiquadric RBF can now be set at runtime using
	the rbf_shape_parameter option in RbfoptSettings.
	The path to the solvers bonmin and ipopt can now be set at runtime
	using minlp_solver_path and nlp_solver_path in RbfoptSettings.
	* Changed:
	Directory structure reorganized to allow installation as a module
	in Python's site-packages.
	(rbfopt_black_box) BlackBox renamed to RbfoptBlackBox.
	(rbfopt_cl_interface.py) is now a script and accepts reading
	black-box definitions from a path.
	(rbfopt_test_interface.py) is now a script.
	Documentation and examples are now distributed within the rbfopt
	module.
	README updated with new installation instructions and minimal
	examples.
	* Removed:
	Module rbfopt_config is no longer necessary.

2017-08-18  Giacomo Nannicini  <nannicini@us.ibm.com>
	Version 3.0.2 released. List of major changes:
	* Added:
	Model improvement search for refinement phase.
	* Changed:
	Refinement phase now happens regardless of finding an improved
	solution.
	Pyomo models changed for numerical stability with default Ipopt
	and Bonmin as downloaded from AMPL's website.
	
2017-06-26  Giacomo Nannicini  <nannicini@us.ibm.com>	
	Version 3.0.1 released. List of major changes:
	* Changed:
	Faster and improved refinement phase.
	Several default settings have been changed to exploit the
	refinement phase further.
	Fixed bug in parallel optimizer.
	
2017-06-14  Giacomo Nannicini  <nannicini@us.ibm.com>	
	* rbfopt_refinement (init_trust_region): Now uses only n+1 points
	rather 2n+1 for the initial trust region model.
	* rbfopt_settings.py (RbfoptSettings): Changed default settings
	for eps_impr, refinement_frequency, max_stalled_cycles to favor
	local search.
	* rbfopt_algorithm.py (optimize_parallel): Fixed bug that could
	occur in parallel optimizer when many consecutive refinement
	steps failed on integer problems.
	
2017-06-09  Giacomo Nannicini  <nannicini@us.ibm.com>
	Version 3.0.0 released. List of major changes:
	* Added: 
	Refinement phase for improved local search: see option
	refinement_frequency and new options that start with tr_.
	Support for Python 3.
	* Changed:
	OptAlgorithm is now called RbfoptAlgoritm.
	RbfSettings is now called RbfoptSettings.
	Automatic model selection is now active by default, using fast
	linear algebra routines.
	Rbf functions in module rbfopt_utils now use Numpy, triggering
	several code optimization changes and major improvement in
	iteration speed.
	Several default settings have been changed.
	* Removed:
	Module rbfopt_model_selection is no longer necessary.
	Option model_selection_solver is no longer necessary.
	Option max_stalled_obj_fun_impr is subsumed by eps_impr.
	Cython modules and setup.py have been removed.

2017-06-08  Giacomo Nannicini  <nannicini@us.ibm.com>
	* rbfopt_utils.py (get_model_quality_estimate)
	(get_best_rbf_model): Functions moved to this module from the
	former model_selection module.	
	* rbfopt_model_selection.py: Module removed.
	* rbfopt_settings.py (RbfoptSettings): Many default settings have
	been updated, most notably rbf is now set to auto, and some
	settings have been deleted.
	
2017-06-03  Giacomo Nannicini  <nannicini@us.ibm.com>
	* rbfopt_utils.py: Major optimizations in the computation of
	radial basis functions.
	* setup.py: File removed, since there is no longer a speed
	advantage using Cython and compilation is therefore not necessary.
	
2017-05-31  Giacomo Nannicini  <nannicini@us.ibm.com>
	* rbfopt_refinement.py: The module provides functions to
	periodically run a trust-region based refinement phase. This is
	only supported for the serial search engine, at the moment.
	* rbfopt_settings.py (max_consecutive_local_searches): This
	parameter has been removed.
	* rbfopt_settings.py (RbfoptSettings): The class RbfSettings was
	renamed RbfoptSettings.
	* rbfopt_algorithm.py (RbfoptAlgorithm): The class OptAlgorithm
	was renamed RbfoptAlgorithm.

2017-05-22  Giacomo Nannicini  <nannicini@us.ibm.com>
	Upgraded code to support Python 3.

2017-03-15  Giacomo Nannicini  <nannicini@us.ibm.com>
	Version 2.1 released.
	* requirements.txt: Updated requirements. 
	* setup.py: Improved build system on Windows.

2017-02-16  Giacomo Nannicini  <nannicini@us.ibm.com>
	* rbfopt_algorithm.py (optimize_serial, optimize_parallel): The
	algorithm now checks if points generated after restart have been
	explored before restart. If so, it uses the previously evaluated
	value.
	* rbfopt_settings.py (max_consecutive_discarded): This setting is
	now intended to be per cpu, i.e. the actual value is obtained
	multiplying by the number of CPUs.

2017-02-09  Giacomo Nannicini  <nannicini@us.ibm.com>
	Major changes with the introduction of Numpy and Cython.
	* rbfopt_aux_problems.py (get_min_bump_node, get_bump_new_node):
	These functions are moved to aux_problems from utils, to allow
	faster performance using Cython.
	* rbfopt_utils.pyx, rbfopt_aux_problems.pyx: Hybrid Python/C
	version of these modules. At the moment only the RBF functions are
	written in real C.
	* setup.py: Created, for automatic compilation in Cython.
	
2016-10-02  Giacomo Nannicini  <nannicini@us.ibm.com>
	* rbfopt_cl_interface.py (register_options): Options are now
	grouped into Algorithmic settings and Execution settings.
	* rbfopt_cl_interface.py (rbfopt_cl_interface): Added option to
	read initialization points from file.
	* rbfopt_algorithm.py (restart): Now adds user-provided points to
	the initial sample. It is no longer possible to provide a full
	initial sample to the algorithm, and employ user-provided points
	instead of the generated sample.
	
2016-07-25  Giacomo Nannicini  <nannicini@us.ibm.com>
	Version 2.0 released.
	* rbfopt_aux_problems.py (global_search, pure_global_search):
	Restructured to allow for 'genetic', 'sampling' and 'solver'
	solution methods..
	* rbfopt_aux_problem.py (MetricSRSMObj): The option
	modified_msrsm_score now determines the weight for the objective
	function component of the score: the classical MSRSM is obtained
	setting it to False, the default is True.

2016-07-20  Giacomo Nannicini  <nannicini@us.ibm.com>
	Removed the possibility of setting integer_vars to None. It now
	has to be empty if there are no integer variables.
	The argument integer_vars now always appears immediately after
	var_upper.
	* rbfopt_algorithm.py (__init__): Resolved a bug in handling fixed
	variables with integer variables.
	* rbfopt_aux_algorithm.py (GutmannHkObj): Improved numerical
	stability of formulas.

2016-07-19  Giacomo Nannicini  <nannicini@us.ibm.com>
	* rbfopt_settings.py: Added global_search_method option to allow
	for solution of auxiliary subproblems not based on solvers.
	* rbfopt_aux_problems.py (ga_optimize, ga_mate, ga_mutate): New
	functions for genetic algorithm, which can be used instead of the
	solvers.
	* rbfopt_aux_problems.py (MetricSRSMObj): Reworked for faster
	computation.
	* rbfopt_aux_problem.py (MaximinDistanceObj, GutmannHkObj)
	(GutmannMukObj): New classes to allow solution of auxiliary
	problems via genetic algorithms.

2016-07-14  Giacomo Nannicini  <nannicini@us.ibm.com>
	Updated dependency to Pyomo 4.3.11388.
	* rbfopt_config.py: Added MINLP_SOLVER_PATH and NLP_SOLVER_PATH
	options.
	* rbfopt_aux_problems.py: Now use PATH option.
	* rbfopt_algorithm.py: Using binary mode to load/save state files
	to fix issues on Windows.

2016-05-05  Giacomo Nannicini  <nannicini@us.ibm.com>
	* rbfopt_settings.py: Added save_state_interval and
	save_state_file settings.
	* rbfopt_algorithm.py (optimize_serial, optimize_parallel):
	Supports periodic state save according to the two options above.

2016-04-20  Giacomo Nannicini  <nannicini@us.ibm.com>
	* rbfopt_algorithm.py (__init__, optimize): Fixed variables no
	longer cause numerical issues and are correctly handled.

2016-03-17  Giacomo Nannicini  <nannicini@us.ibm.com>
	* rbfopt_algorithm.py (update_log): Log file format change to fit
	within 80 characters, and print elapsed time.
	
2016-03-16  Giacomo Nannicini  <nannicini@us.ibm.com>
	* rbfopt_algorithm.py (OptAlgorithm.optimize_parallel): Added
	parallel version of the optimization engine.
	* rbfopt_algorithm.py (OptAlgorithm.optimize_serial): Code for the
	serial (standard) optimizer was moved here.
	* rbfopt_algorithm.py (optimize): Now simply redirects to either
	the serial or parallel version.
	* rbfopt_black_box.py (BlackBox): Added abstract class to define a
	black box optimization problem.
	* rbfopt_black_box_example.py (BlackBox): Added an example of an
	implementation of the BlackBox abstract class.
	* rbfopt_utils.py: Added several utility functions for parallel
	optimization.
	* README: Documentation updated to discuss the parallel optimizer.
	
2016-03-07  Giacomo Nannicini  <nannicini@us.ibm.com>
	* rbfopt_algorithm.py (OptAlgorithm): Added class for optimization
	algorithm. Allows saving/loading algorithm state.
	* rbfopt.py (rbf_optimize): Now simply redirects to OptAlgorithm.
	
2016-02-28  Giacomo Nannicini  <nannicini@us.ibm.com>
	* test_rbfopt.py: Moved to tests/rbfopt_test_interface.py.
	* tests/: Added unit tests for all modules to documentation.
	* rbfopt.py (rbf_optimize): Several buxfixes.
	* rbfopt_utils.py (several functions): Several bugfixes.
	Version 1.2.0 released.
	
2015-11-21  Giacomo Nannicini  <nannicini@sutd.edu.sg>
	* rbfopt.py (rbf_optimize): Implementation of Metric SRSM method.
	* rbfopt_utils.py: Performance improvements. Bulk version of some
	utility functions (bulk_evaluate_rbf, bulk_get_min_distance).
	* rbfopt_aux_problems.py (maximize_one_over_mu): Renamed to
	pure_global_search.
	* rbfopt_aux_problems.py (maximize_h_k): Renamed to global_search.
	* rbfopt_aux_problems.py (MetricSRSMObj): New class to support the
	Metric SRSM method.

2015-10-26  Giacomo Nannicini  <nannicini@sutd.edu.sg>
	* README: Updated installation instructions for Windows systems.

2015-09-23  Giacomo Nannicini  <nannicini@sutd.edu.sg>
	* rbfopt.py (rbf_optimize): Minor bug fixing and performance
	improvements in handling the noisy oracle.

2015-07-29  Giacomo Nannicini  <nannicini@sutd.edu.sg>
	* rbfopt_model_selection.py: Introduced stand-alone module for
	model selection. Increased efficiency of cross-validation via LPs
	solved with Cplex or Clp.
	* rbfopt_settings.py (RbfSettings): Introduced unpacking of
	parameters from dictionary.
	* rbfopt_settings.py (RbfSettings): skip_targetval_clipping
	flipped and renamed to targetval_clipping.
	* rbfopt_cl_interface.py (rbfopt_cl_interface): Switched to
	dictionary to pass algorithmic parameters.
	* test_rbfopt: Switched to dictionary to pass algorithmic
	parameters.
	Version 1.1.0 released.

2015-07-17  Giacomo Nannicini  <nannicini@sutd.edu.sg>
	Switch to Pyomo 4. The relevant files have been updated to use the
	pyomo module rather than coopr.

2015-06-23  Giacomo Nannicini  <nannicini@sutd.edu.sg>
	* rbfopt_settings.py (RbfSettings): Added log_scaling_threshold
	option. Moved RbfSettings here from the rbfopt module.
	* rbfopt.py (rbf_optimize): Improved automatic function
	scaling. Improved efficiencyhen skipping infstep.
	* rbfopt_cl_interface.py (register_options): Parameters are now
	read automatically from the rbfopt_settings module.

2015-06-11  Giacomo Nannicini  <nannicini@sutd.edu.sg>
	* black_box.py: Introduced class BlackBox as a wrapper around the
	black-box methods.
	Improved docstrings following NumPy's format.
	Added documentation in the doc/ directory.
	Version 1.0.2 released.

2015-06-03  Giacomo Nannicini  <nannicini@sutd.edu.sg>
	* rbfopt.py (rbf_optimize): Better printout of initialization
	points.

2015-05-17  Giacomo Nannicini  <nannicini@sutd.edu.sg>
	* rbfopt_cl_interface.py: Added module for command line interface.
	* rbfopt_utils.py (initialize_nodes): Added test for linear
	independence of the generated points.
	* test_rbfopt.py (test_rbfopt): Function removed.
	* test_rbfopt.py: Module restructured to integrate with
	rbfopt_cl_interface.py.
	Version 1.0.1 released.

2015-05-03  Giacomo Nannicini  <nannicini@sutd.edu.sg>
	* rbfopt.py (RbfSettings): Added print_solver_output option.
	* rbfopt_aux_problems: Added support for print_solver_output.
	* rbfopt_aux_problems (maximize_h_k): Added initialization of
	auxiliary variables for the problem.
	* test_functions: Added several MINLP functions.

2014-12-05  Giacomo Nannicini  <nannicini@sutd.edu.sg>
	Version 1.0 released.

2014-11-19  Giacomo Nannicini  <nannicini@sutd.edu.sg>
	* rbfopt.py (rbf_optimize): Added support for redirection of
	output stream in rbfopt.py.
	* rbfopt_utils.py (transform_function_values): Fixed division by
	zero bug when fmin is zero.

2014-11-05  Giacomo Nannicini  <nannicini@sutd.edu.sg>
	* rbfopt_utils.py (transform_function_values): Fixed misspelled
	keyword 'a ffine'.

2014-09-24  Giacomo Nannicini  <nannicini@sutd.edu.sg>
	Version 0.9 released.