Distributions

Distributions for Stochastic Policy in RLzoo

Definition of parametrized distributions. Adapted from openai/baselines

class rlzoo.common.distributions.Categorical(ndim, logits=None)

Bases: rlzoo.common.distributions.Distribution

Creates a categorical distribution

entropy()

Calculate the entropy of distribution.

get_param()

greedy_sample()

Get actions greedily

kl(logits)

Args:

logits (tensor): logits variables of another distribution

logp(x)

Calculate log probability of a sample.

ndim

neglogp(x)

Calculate negative log probability of a sample.

sample()

Sample actions from distribution, using the Gumbel-Softmax trick

set_param(logits)

Args:

logits (tensor): logits variables to set

class rlzoo.common.distributions.DiagGaussian(ndim, mean_logstd=None)

Bases: rlzoo.common.distributions.Distribution

Creates a diagonal Gaussian distribution

entropy()

Calculate the entropy of distribution.

get_param()

Get parameters

greedy_sample()

Get actions greedily/deterministically

kl(mean_logstd)

Args:

mean_logstd (tensor): mean and logstd of another distribution

logp(x)

Calculate log probability of a sample.

ndim

neglogp(x)

Calculate negative log probability of a sample.

sample()

Get actions in deterministic or stochastic manner

set_param(mean_logstd)

Args:

mean_logstd (tensor): mean and log std

class rlzoo.common.distributions.Distribution

Bases: object

A particular probability distribution

entropy()

Calculate the entropy of distribution.

kl(*parameters)

Calculate Kullback–Leibler divergence

logp(x)

Calculate log probability of a sample.

neglogp(x)

Calculate negative log probability of a sample.

sample(*args, **kwargs)

Sampling from distribution. Allow explore parameters.

set_param(*args, **kwargs)

rlzoo.common.distributions.expand_dims(func)

rlzoo.common.distributions.make_dist(ac_space)

Get distribution based on action space

Parameters:ac_space – gym.spaces.Space