sampling#

Sampling methods.

This module contains a collection of sampling methods. The core principle is to define a function that returns a single observation (either via return or yield) from a probability measure. Then, using the decorator sample_generator, a function that returns a single observation can be converted into a generator, that can then be sampled using islice.

Example

Sample the stochastic kernel of a dynamical system (i.e. the state transition probability kernel).

>>> from gym_socks.envs.integrator import NDIntegratorEnv
>>> from gym_socks.policies import RandomizedPolicy
>>> env = NDIntegratorEnv(2)
>>> sample_space = gym.spaces.Box(low=-1, high=1, shape=(2,), dtype=np.float32)
>>> sampler = default_sampler(
...     system=env, policy=RandomizedPolicy(env), sample_space=sample_space
... )
>>> S = sample(sampler=sampler, sample_size=100)

The main reason for this setup is to allow for “observation functions” which have different structures, e.g.

  • functions that return an observation,

  • infinite generators that yield observations, and

  • finite generators that yield observations.

Submodules#