garage.tf.policies.categorical_lstm_policy module

Categorical LSTM Policy.

A policy represented by a Categorical distribution which is parameterized by a Long short-term memory (LSTM).

class CategoricalLSTMPolicy(env_spec, name='CategoricalLSTMPolicy', hidden_dim=32, hidden_nonlinearity=<function tanh>, hidden_w_init=<tensorflow.python.ops.init_ops_v2.GlorotUniform object>, hidden_b_init=<tensorflow.python.ops.init_ops_v2.Zeros object>, recurrent_nonlinearity=<function sigmoid>, recurrent_w_init=<tensorflow.python.ops.init_ops_v2.GlorotUniform object>, output_nonlinearity=None, output_w_init=<tensorflow.python.ops.init_ops_v2.GlorotUniform object>, output_b_init=<tensorflow.python.ops.init_ops_v2.Zeros object>, hidden_state_init=<tensorflow.python.ops.init_ops_v2.Zeros object>, hidden_state_init_trainable=False, cell_state_init=<tensorflow.python.ops.init_ops_v2.Zeros object>, cell_state_init_trainable=False, state_include_action=True, forget_bias=True, layer_normalization=False)

Bases: garage.tf.policies.policy.StochasticPolicy

Categorical LSTM Policy.

A policy represented by a Categorical distribution which is parameterized by a Long short-term memory (LSTM).

It only works with akro.Discrete action space.

Parameters:

• env_spec (garage.envs.env_spec.EnvSpec) – Environment specification.
• name (str) – Policy name, also the variable scope.
• hidden_dim (int) – Hidden dimension for LSTM cell.
• hidden_nonlinearity (callable) – Activation function for intermediate dense layer(s). It should return a tf.Tensor. Set it to None to maintain a linear activation.
• hidden_w_init (callable) – Initializer function for the weight of intermediate dense layer(s). The function should return a tf.Tensor.
• hidden_b_init (callable) – Initializer function for the bias of intermediate dense layer(s). The function should return a tf.Tensor.
• recurrent_nonlinearity (callable) – Activation function for recurrent layers. It should return a tf.Tensor. Set it to None to maintain a linear activation.
• recurrent_w_init (callable) – Initializer function for the weight of recurrent layer(s). The function should return a tf.Tensor.
• output_nonlinearity (callable) – Activation function for output dense layer. It should return a tf.Tensor. Set it to None to maintain a linear activation.
• output_w_init (callable) – Initializer function for the weight of output dense layer(s). The function should return a tf.Tensor.
• output_b_init (callable) – Initializer function for the bias of output dense layer(s). The function should return a tf.Tensor.
• hidden_state_init (callable) – Initializer function for the initial hidden state. The functino should return a tf.Tensor.
• hidden_state_init_trainable (bool) – Bool for whether the initial hidden state is trainable.
• cell_state_init (callable) – Initializer function for the initial cell state. The functino should return a tf.Tensor.
• cell_state_init_trainable (bool) – Bool for whether the initial cell state is trainable.
• state_include_action (bool) – Whether the state includes action. If True, input dimension will be (observation dimension + action dimension).
• forget_bias (bool) – If True, add 1 to the bias of the forget gate at initialization. It’s used to reduce the scale of forgetting at the beginning of the training.
• layer_normalization (bool) – Bool for using layer normalization or not.

build(state_input, name=None)

Build policy.

Parameters:

• state_input (tf.Tensor) – State input.
• name (str) – Name of the policy, which is also the name scope.

Returns:

Policy distribution. tf.Tensor: Step output, with shape \((N, S^*)\) tf.Tensor: Step hidden state, with shape \((N, S^*)\) tf.Tensor: Step cell state, with shape \((N, S^*)\) tf.Tensor: Initial hidden state, used to reset the hidden state

when policy resets. Shape: \((S^*)\)

tf.Tensor: Initial cell state, used to reset the cell state

when policy resets. Shape: \((S^*)\)

Return type:

tfp.distributions.OneHotCategorical

clone(name)

Return a clone of the policy.

It only copies the configuration of the primitive, not the parameters.

Parameters:name (str) – Name of the newly created policy. It has to be different from source policy if cloned under the same computational graph. Returns:Newly cloned policy. Return type:garage.tf.policies.CategoricalLSTMPolicy

distribution

Policy distribution.

Returns:Policy distribution. Return type:tfp.Distribution.OneHotCategorical

get_action(observation)

Return a single action.

Parameters:observation (numpy.ndarray) – Observations. Returns:Action given input observation. dict(numpy.ndarray): Distribution parameters. Return type:int

get_actions(observations)

Return multiple actions.

Parameters:observations (numpy.ndarray) – Observations. Returns:Actions given input observations. dict(numpy.ndarray): Distribution parameters. Return type:list[int]

input_dim

Dimension of the policy input.

Type:int

reset(do_resets=None)

Reset the policy.

Note

If do_resets is None, it will be by default np.array([True]), which implies the policy will not be “vectorized”, i.e. number of paralle environments for training data sampling = 1.

Parameters:do_resets (numpy.ndarray) – Bool that indicates terminal state(s).

state_info_specs

State info specifcation.

Returns:

keys and shapes for the information related to the

policy’s state when taking an action.

Return type:List[str]

vectorized

Vectorized or not.

Returns:True if primitive supports vectorized operations. Return type:Bool