stable_learning_control.algos.pytorch.policies.lyapunov_actor_twin_critic

Lyapunov (soft) actor twin critic policy.

This module contains a modified Pytorch implementation of the Lyapunov Actor-Critic policy of Han et al. 2020. Like the original SAC algorithm, this LAC variant uses two critics instead of one to mitigate a possible underestimation bias, while the original LAC only uses one critic.

Module Contents

Classes

LyapunovActorTwinCritic

Lyapunov (soft) Actor-(twin Critic) network.

Attributes

HIDDEN_SIZES_DEFAULT
ACTIVATION_DEFAULT
OUTPUT_ACTIVATION_DEFAULT

stable_learning_control.algos.pytorch.policies.lyapunov_actor_twin_critic.HIDDEN_SIZES_DEFAULT

stable_learning_control.algos.pytorch.policies.lyapunov_actor_twin_critic.ACTIVATION_DEFAULT

stable_learning_control.algos.pytorch.policies.lyapunov_actor_twin_critic.OUTPUT_ACTIVATION_DEFAULT

class stable_learning_control.algos.pytorch.policies.lyapunov_actor_twin_critic.LyapunovActorTwinCritic(observation_space, action_space, hidden_sizes=HIDDEN_SIZES_DEFAULT, activation=ACTIVATION_DEFAULT, output_activation=OUTPUT_ACTIVATION_DEFAULT)

Bases: torch.nn.Module

Lyapunov (soft) Actor-(twin Critic) network.

self.pi

The squashed gaussian policy network (actor).

Type:

SquashedGaussianActor

self.L

The first soft L-network (critic).

Type:

LCritic

self.L2

The second soft L-network (critic).

Type:

LCritic

Initialise the LyapunovActorTwinCritic object.

Parameters:

• observation_space (gym.space.box.Box) – A gymnasium observation space.

• action_space (gym.space.box.Box) – A gymnasium action space.

• hidden_sizes (Union[dict, tuple, list], optional) – Sizes of the hidden layers for the actor. Defaults to (256, 256).

• activation (Union[dict, torch.nn.modules.activation], optional) – The (actor and critic) hidden layers activation function. Defaults to torch.nn.ReLU.

• output_activation (Union[dict, torch.nn.modules.activation], optional) – The (actor and critic) output activation function. Defaults to torch.nn.ReLU for the actor and nn.Identity for the critic.

Note

It is currently not possible to set the critic output activation function when using the LyapunovActorTwinCritic. This is since it by design requires the critic output activation to by of type torch.square().

forward(obs, act, deterministic=False, with_logprob=True)

Performs a forward pass through all the networks (Actor and both L critics).

Parameters:

• obs (torch.Tensor) – The tensor of observations.

• act (torch.Tensor) – The tensor of actions.

• deterministic (bool, optional) – Whether we want to use a deterministic policy (used at test time). When true the mean action of the stochastic policy is returned. If false the action is sampled from the stochastic policy. Defaults to False.

• with_logprob (bool, optional) – Whether we want to return the log probability of an action. Defaults to True.

Returns:

tuple containing:

• pi_action (torch.Tensor): The actions given by the policy.

• logp_pi (torch.Tensor): The log probabilities of each of these actions.

• L (torch.Tensor): First critic L values.

• L2 (torch.Tensor): Second critic L values.

Return type:

(tuple)

Note

Useful for when you want to print out the full network graph using TensorBoard.

act(obs, deterministic=False)

Returns the action from the current state given the current policy.

Parameters:

• obs (torch.Tensor) – The current observation (state).

• deterministic (bool, optional) – Whether we want to use a deterministic policy (used at test time). When true the mean action of the stochastic policy is returned. If False the action is sampled from the stochastic policy. Defaults to False.

Returns:

The action from the current state given the current policy.

Return type:

numpy.ndarray