Flatten inputs along dimension 1 and then pass through MLP.

Normalize action to in [-1, 1]. Optionally normalize observations and scale reward.

Usage: ``` policy = TanhGaussianPolicy(...) action, mean, log_std, _ = policy(obs) action, mean, log_std, _ = policy(obs, determ

A helper class since most policies have a tanh output activation.

Take a random discrete action with some probability.

This strategy implements the Ornstein-Uhlenbeck process, which adds time-correlated noise to the actions taken by the deterministic policy.

Replay buffer for environments whose observations are dictionaries, such as - OpenAI Gym GoalEnv environments. https://blog.openai.com/in

This class wraps an image-based environment with a VAE. Assumes you get flattened (channels,84,84) observations from wrapped_env. This class a

This strategy adds Gaussian noise to the action taken by the deterministic policy. Based on the rllab implementation.

Usage: ``` path_builder = PathBuilder() path.add_sample( observations=1, actions=2, next_observations=3,

Twin Delayed Deep Deterministic policy gradients

Represent distribution of X where X ~ tanh(Z) Z ~ N(mean, std) Note: this is not very numerically stable.

Deep Deterministic Policy Gradient

Object composed of a list of OutofGraphReplayBuffers.

With probability epsilon, take a completely random action. with probability 1-epsilon, add Gaussian noise to the action taken by a determ

Simple 1D LayerNorm.

The convolutional network used to compute the agent's Q-values. Attributes: num_actions: An integer representing the number of actions. con

A logged out-of-graph Replay Buffer for Prioritized Experience Replay.

Logs the replay buffer to disk everytime it's full.

Given a list of (x, t) value-time pairs, return value x at time t, and linearly interpolate between the two

A timer that goes off after the a fixed time period. The catch: you need to poll it and provide it the time! Usage: ``` timer =

Do a grid search over hyperparameters based on a predefined set of hyperparameters.

A simple wrapper to combiner multiple DeterministicHyperParameterSweeper's

An implementation of the DQN agent with fixed replay buffer(s).

MultiHeadDQNAgent with fixed replay buffer(s).

MultiNetworkDQNAgent with fixed replay buffer(s).

An implementation of the DQN agent with fixed replay buffer(s).

An implementation of the DQN agent with fixed replay buffer(s).

Object that handles running Dopamine experiments with fixed replay buffer.

Tests for Atari environment with various agents.

Automatically assign int IDs to hashable objects. Usage: ``` id_map = IntIdDict() print(id_map['a']) print(id_map['b'])

Same as RampUpSchedule but round output to an int

Linearly interpolate and then stop at a final value.

Return something ranging from [min_value + offset, max_value + offset], distributed with a log.

An implementation of the DQN agent with replay buffer logging to disk.

An implementation of the Quantile agent with replay buffer logging to disk.

A simpler interface for creating policies.

My own wrapper around MujocoEnv. The caller needs to declare

Multiple convolutional networks to compute Q-value estimates. Attributes: num_actions: An inteer representing the number of actions. num_ne

DQN agent with multiple heads.

Multi-head convolutional network to compute multiple Q-value estimates. Attributes: num_actions: An integer representing the number of actions.

DQN agent with multiple heads.

General policy interface.