SobolSampler

class macrostat.sample.sobol.SobolSampler(model: Model, bounds: dict | None = None, sample_power: int = 10, sobol_seed: int = 0, logspace: bool = False, worker_function: callable = <function timeseries_worker>, simulation_args: tuple = (), output_folder: str = 'sobol_samples', cpu_count: int = 1, batchsize: int = None, output_filetype: str = 'csv', output_compression: str | None = None)

Bases: BaseSampler

A Sobol sequence sampler for efficient parameter space exploration.

The SobolSampler class implements quasi-random low-discrepancy sequence sampling to systematically explore a model’s parameter space. The benefit of Sobol sequences is that even the sub-samples are uniform distributed, which is not the case for random sampling from a uniform distribution.

The sampler works by generating a Sobol sequence in the unit hypercube [0,1]^d, transforming the sequence to the desired parameter ranges and creating model instances with the sampled parameters.

Example

>>> model = MyModel()
>>> bounds = {
...     'param1': (0.1, 1.0),
...     'param2': (1.0, 10.0)
... }
>>> sampler = SobolSampler(
...     model=model,
...     bounds=bounds,
...     sample_power=8,  # 2^8 = 256 samples
...     logspace=True,   # Sample in log space
...     cpu_count=4      # Use 4 CPUs
... )
>>> sampler.sample()

generate_parameters()

Generate points in the parameterspace for the parallel processor based on a Sobol sequence.

Here the scipy.stats.qmc.Sobol class is used to generate the Sobol sequence, specifically the random_base2 method is used to generate the samples, as it is has slightly better space filling properties than with a custom number of samples.

Returns:

DataFrame containing the Sobol points in the parameterspace. Rows are the samples, columns are the parameters

Return type:

pd.DataFrame