CHANGELOG.md

Changelog

All notable changes to this project will be documented in this file. This project adheres to Semantic Versioning.

Unreleased

Added

• TuRBO is now implemented as a Procedure in the optimisation module. The code, and an example use, can be found in the examples folder as well.
• Additional tests on the shape of input data for test functions and on the shape of output data from procedures. This change was made to accomodate the usage of the TuRBO package.
• Argument verbose to Experiment initialisation method. This method controls the output of intermediate results (#samples for the #procedure_calls). See documentation of init of the Experiment class for more information.
• The dimensionality of hidden TestFunctions can now be changed with the dimensionality argument at initialisation of the class.
• The ParticleFilter now implements a callback method that allows for the inclusion of a callback function in each sampling iteration (except for the first one). See the documentation on the wiki and in the docstring for more information.
• A MSSM7 function has been added to the functions module. This function calls a trained neural network to evaluate the likelihood in a 12-dimensional space (which includes 5 nuisance parameters). The algorithm was trained on data from the Gambit collaboration.
• The PyScannerBit optimiser previously always inverted the test function before optimising it, as the optimiser all functions were to be minimised. With the introduction of the MSSM7 test function this assumption has become invalid, as MSSM7 needs to be maximised. The PyScannerBit optimiser therefore now implements the invert_function argument at construction, which indicates whether or not the test function should be inverted or not (default is True).
• Example code for pyscannerbit has been added to examples folder.

Changed

• As pygmo cannot be installed through the pip installer, it has been removed from installation requirements in the setup.py file. This will only yield error messages when the pygmo package is actually requested by the implemented pygmo Procedure.
• The weighing_deterministic_linear function in the particle filter's linear function was based on sample order, not on function value. This has been changed.
• To accomodate the implementation of callbacks in the particle filter, the selector methods don't select data directly, but instead output (indices, samples, values), where indices are the indices of the samples and values selected.
• Changed particle filter implementation to the one in https://github.com/bstienen/particlefilter.
• Plotting style now puts the grid at the lowest z-order, such that e.g. scatter markers are fully visible.

Fixed

• Some print commands that were left from a debugging era are now removed.
• When using the invert method on a wrapped function, the original function was also inverted. This was solved by having the get_simple_interface() and get_simple_interface_with_scan methods use copies of the original function.
• The weighing_stochastic_linear function in the particle filter optimisation method could occasionally raise errors related to nan probabilities. This is now fixed.
• The gaussian constructors needed for the particle filter could return negative results. As the output represents standard deviations, this is unexpected behaviour. The absolute value of the standard deviation is now returned instead.
• The particle filter did not implement a way to keep the best points of the previous iteration for the current one. This is now implemented through the survival_rate argument (default=0.2).
• The pygmo interface contained an error, which caused an incorrection function value to be reported in some instances. The coordinate of the best found minimum was correctly reported, but the function value is now fixed so that is actually correctsponds to the value at this coordinate.

Version 0.2.0 (Monday February 17th, 2020)

Added

• A submodule results that implements functionality to visualise or format the results of one or more experiments into neat little plots and tables. Accompanying this new submodule is a new example script.
• The Experiment class now is an abstract class with abstract methods that function as events. This allows users to more easily create Experiment classes that extend the functionality of the Experiment class.
• The experiment.yaml log now also contains entries on the experiment type (e.g. OptimisationExperiment) and the number of calls for the derivative of the TestFunction.
• The interface to TestFunctions was not working for all users. An interface class (SimpleFunctionWrapper) has been added that allows the user to provide the input parameters each as a separate argument. The interface can be activated by calling the get_simple_interface method of a TestFunction. An example using the different interfaces has been added (testfunction_interfaces.py).
• Add methods to the TestFunctions that check properties of the function. These include is_differentiable, is_bounded and get_dimensionality.
• Subpackage for algorithm implementations. RandomOptimisation and RejectionSampling (from the example scripts) have been added as examples of how to add implementations.
• Unit tests for the entire package are added.
• Testfunctions from precompiled binaries were added, so that the exact formula for them is unknown.
• Several optimisation procedures were added.

Changed

• The examples now use relative path indications for storage of their results, so that the examples can be used out of the box without errors.
• The package indicator in setup.py now is more general, allowing for easier inclusion of submodules.
• Procedures should now implement a method called check_testfunction that returns a boolean, indicating if the provided test function can be used in the procedure.
• The ranges of the Block testfunction now go all the way to -np.inf and np.inf, as is in accordance with the configuration of the function groups in the FunctionFeeder.

Fixed

• An error was raised when GaussianShells was logged, as this function stores its parameters internally as numpy arrays. These don't translate well to .yaml files. This has been solved by forcing the storage of function properties as lists if they are numpy arrays.
• Updated outdated docstrings at various locations in the package
• Fixed error causing testfunctions not to work with Pandas dataframes. They work with dataframes now, as was intended.

---

Version 0.1.1 (Tuesday August 20th, 2019)

Added

• A get_ranges method in the TestFunction class. It returns the ranges of the test function, taking a leeway parameter epsilon provided to the method into account: all minima will be raised by epsilon, all maxima will be reduced by it.

• The TestFunction __call__ and check_ranges methods now also implement the epsilon argument (see first addition above).

• Users requested a smaller range for the Easom test function. Although the range [-100, 100] is normal for this function, to accommodate this wish the boundaries of this application box are now configurable at initialisation of the test function.

• The plot_testfunction example script has been splitted into two separate scripts: one for 1d and another for 2d test functions.

• When an experiment ends, the results (i.e. the elapsed time and the number of total function calls) are stored in the experiment.yaml file. Note that the logged time includes also the overhead introduced by the HDS package itself.

• The following test functions are added:

  • BreitWigner
  • Reciprocal

Changed

• Increased the readability of the error message given when a test function is queried for its value outside of the box defined by its ranges parameter.
• The utils.get_time method now returns time in seconds

Fixed

• Ackley, Easom and Sphere test functions returned data in an incorrect shape. This has been corrected.
• The GaussianShells test function mapped multi-point input to a single output value. This has been fixed.