Distributions

A Distribution is a distribution over the parameter vector of a policy. It is the object the black-box optimization algorithms learn: at the beginning of an episode they sample a parameter vector from it, run the episode with the resulting policy, and update the distribution from the return that was obtained.

A distribution can be contextual, i.e. conditioned on a context vector built from the initial state and the episode info, in which case sampling and updating are performed per context.

Distribution	Interface for Distributions to represent a generic probability distribution.
GaussianDistribution	Gaussian distribution with fixed covariance matrix.
GaussianDiagonalDistribution	Gaussian distribution with diagonal covariance matrix.
GaussianCholeskyDistribution	Gaussian distribution with full covariance matrix.

• Distribution interface
• Gaussian distributions
• Torch distributions