algorithms.util.action_perturbator

This module defines several classes to perturb (explore) actions.

An ActionPerturbator takes a continuous action, i.e., a vector of parameters, and returns a vector with same shape but different values. The difference between the original vector and the resulting one is called the perturbation, is often implemented as a random noise, and allows exploring the action space.

We have implemented 4 perturbators:

• Epsilon applies a uniform random noise to all dimensions.

• Gaussian applies a gaussian random noise to all dimensions.

• MultiDim has a probability to apply a uniform noise to each dimension.

• Identity performs no perturbation.

members:

Classes

ActionPerturbator()
EpsilonActionPerturbator(noise)	Implements a simple ε noise.
GaussianActionPerturbator(noise)	Implements a Gaussian (σ) noise.
IdentityAction()	Returns the same action (no perturbation).
MultiDimActionPerturbator(noise, probability)	Custom algorithm, with a probability to noise each dimension.