Creating a sampling procedure

The high-dimensional-sampling (hds) package allows users to define sampling procedures and test them on a wide variety of test functions in a neatly automated an logged manner. Both the experimental framework and the test functions are defined within the package, the only thing the user has to do is define the procedure itself.

Choosing the type of sampling

Within the hds project two different sampling tasks have been defined:

• Posterior sampling: taking samples from a distribution of which the probability density (i.e. the function value) is only accessible through these samples.
• Optimisation: Finding the optimum of an unknown function.

Implementing a procedure for either of these two can be done by creating a class that inherits from the procedures.Procedure class:

import high_dimensional_sampling as hds

class MyProcedure(hds.Procedure):
    ...

This class should implement at least the following four methods:

__init__(self)

This is Initialisation method in which anything can be done. At the very least it should define the self.store_parameters property as a list, containing the names of the parameters of the Procedure. These parameters will then autometically be stored when an experiment is run. If no procedure parameters exist, self.store_parameters should be an empty list.

Your implementation of this initialisation method is allowed to include extra input arguments, but all these input arguments should have default values implemented, i.e.

__init__(self, a=1, b=3.14, c=4)

As this method is an intialisation of the object, no parameters should be returned from it.

__call__(self, function)

The call method queries the class for the sampling of (a) new data point(s). It takes a TestFunction instance as input and should return the sampled data point(s) x and the labels for this(these) data point(s) y. The labels can be obtained by calling the test function. x should be a numpy array of shape (nDatapoints, nVariables), whereas y should be a numpy array of shape (nDatapoints, nOutputVariables). Note that nOutputVariables will often be 1.

Although it is not required, it is advisable to implement the __call__ method in such a way that it returns a single data point at a time. Whether of not this is possible depends of course on the implemented procedure, but following this guideline makes sure that the experiment is stopped as soon as possible and no further (possibly expensive) iterations are run. This makes comparison of different Procedures easier.

is_finished(self)

This method checks if the Procedure is finished and the experiment can be stopped. It should return a boolean, which is True is the experiment can be stopped, or False if it should continue. If True, this overrides the finish_line condition set in the Experiment.

reset(self)

This method should reset all internal variables conserning previous runs. For instance: if your implementation keeps track of the number of points sampled in a property called n_sampled, the reset() method should set n_sampled to 0. It is called every time a new testfunction is provided to the Procedure.

Please do note that it should not reset possible configuration parameters of the Procedure.

Example scripts

In the examples folder two example procedures are provided: one for rejection sampling, a procedure for posterior sampling, the other for random optimisation, a procedure for optimum finding.