Testing Uniform Random Samplers: Methods, Datasets and Protocols

The chi2 folder contains the statistical tests as well as the scripts necessary to build the singularity container that will contain the tests to run as well as the samplers under test.

The datasets folder contains various DIMACS files.

The utils folder contains the python d-DNNF parser that is used by the statistical tests as well as the code necessary to generate our synthetic datasets. rsat can be used to generate random k-CNF formulae and blast can be used to blast random variable assignments into an existing formula.

The results contains various python scripts used to generate the results in the paper as well as the csv files obtained during our experiments.

Testing a Sampler

cd chi2
make
./chi2.sif --test <statistical test> \
    --min_elem_per_cell <integer (default 10)> \
    -b <batch size (default 1000)> \
    -c <path to cnf file> \
    -s <sampler to test> \
    -a <significance level (usually 0.01)>

The available values for --test are

monobit
modbit
freq_var
sfpc
chisquared
birthday

The modbit test has an aditional parameter --modbit_q <integer>.

The birthday test has an additional parameter --bday_prob <float>. If this value is smaller than 1.0 then it is interpreted as the probability to observe a repetition. If it is greater than 1.0 then it is interpreted as the desired expected number of repetitions under the null hypothesis. This parameter influences the required sample size.

The available values for -s are

unigen3
spur
sts
smarch
quicksampler
cmsgen
kus
distaware
walksat
bddsampler

The current integration of BDDSampler expects that there exist a file <cnf file>.bdd which has been computed before running the test.

Adding a Sampler

To add a sampler, please modify the src/samplers.py file. You will need to write a function getSolutionsFromSampler(inputFile : str, numSolutions : int, newSeed : int) : [frozenset(int)] with inputFile the path to the DIMACS file, numSolutions the requested sample size and newSeed the random seed given to the function (this can increase reproducibility as well as reduce isses especially if the random number generator of the sampler is initialized with the current time). The function is expected to return a list of solutions to the formula. Each solution is expected to be a frozenset of integers with each integer being a literal in the usual DIMACS format.

Finally, please modify the function getSamplerFunction which should ideally remain at the end of the file. This function takes as input the string given to the program via the option -s and returns the appropriate getSolutionsFromSampler function depending on the string.

If your sampler has been installed within the container (preferrably base.def once the installation is stable enough), you should be able to test your sampler.

Adding a Test

To add a statistical test, please modify the src/tests.py file. You will need to write a function that performs the statistical test and takes as input an object of type Settings as defined in the src/util.py file. The Settings class contains various parameters given to the program such as batch size, the sampler function to be used (Settings.sampler_fn) as well as the d-DNNF representing the formula under test.

Once you have added the necessary function, you should add an entry to the getTestFunction(str) which should preferable remain at the end of the file. This function takes as input a string given to the program via the option --test and returns the appropriate perform_statistical_test(Settings) function.