Samplers

By default Everest generates perturbations to estimate gradients using sampling code based on SciPy. It provides sampling from common distributions such as Gaussian, Uniform and Bernoulli, and some additional methods such as Sobol and Latin hypercube sampling.

The sampling method and options are specified in the sampler subsection of the controls configuration settings. The sampling method is selected using the method keyword. If the method keyword is missing, the backend will select a default method, which samples from a normal distribution.

The sampling methods in the SciPy backend support several options that can be passed using the backend_options keyword. Please consult the online SciPy manual for details for the options to those sampler that can be passed.

The sampler methods use a random number generator to generate stochastic samples. The seed for that generator can be set using the seed keyword. If the seed keyword is not provided, the more general Everest seed, as configured by the random_seed keyword is used instead.

Everest offers the possibility to specify another backend for sampling with the optional backend keyword in the sampler` subsection of the ‘controls’ configuration settings. Everest will check if the requested backend is installed, and if the method specified by the method keyword is supported. By default, if the backend keyword is missing, Everest will select the scipy backend. Additional backends may be installed in your local installation, consult your systems manager for options.

Example

optimization:
    sampler:
        method: sobol
        seed: 123

Another example using the SciPy backend, with some options added, and the seed taken from the global Everest seed:

optimization:
    sampler:
        method: sobol
        backend_options:
            scramble: False