rlgt.environments.local_environments.LocalFlipEnvironment

class documentation

class LocalFlipEnvironment(GraphEnvironment):

Constructor: LocalFlipEnvironment(graph_invariant, graph_order, episode_length, flip_only, ...)

This class inherits from the GraphEnvironment class and models a graph building game used for constructing 2-edge-colored looped complete graphs in which the edges (resp. arcs) are initially fully colored in some manner, and the agent moves from one vertex to another, thereby traversing existing edges (resp. arcs) and potentially flipping them. More precisely, in each step, the agent is located at a vertex and selects an incident edge (resp. an outgoing arc), traverses it, and moves to the opposite endpoint. While traversing an edge (resp. arc), the agent may flip its proper edge color. The user can select the graph order, choose whether the graphs are directed or undirected, and specify whether loops are allowed. Additionally, the user can configure the mechanism that controls how the initial fully colored graphs are generated, which may be either deterministic or nondeterministic. The user can also select the vertex at which the agent starts the potential flipping procedure.

The RL tasks in this environment are continuing, and the total number of actions to be performed, i.e., the episode length, is configurable.

Each state is represented by a binary numpy.ndarray of type numpy.uint8 and length flattened_length + graph_order, where graph_order is the configured graph order and flattened_length is the flattened length of the graphs to be constructed. In the state vectors, the first flattened_length bits indicate which of the flattened_length edges (resp. arcs) are currently of color 1. The edges (resp. arcs) are ordered according to the selected flattened ordering, which may be either row-major or clockwise. The final graph_order bits form a one-hot encoding of the vertex at which the agent is currently located.

The environment supports two action-handling modes. If flip_only is set to False (the default), each action is represented by a numpy.int32 integer between 0 and 2 * graph_order - 1. For an action value a, the quantity a % graph_order specifies the vertex to which the agent moves, while a // graph_order determines whether the traversed edge (resp. arc) is flipped. A value of 1 indicates that the proper edge color is changed, whereas 0 indicates that it remains unchanged. If flip_only is set to True, then every traversed edge (resp. arc) must be flipped. In this case, each action is a numpy.int32 integer between 0 and graph_order - 1 specifying the destination vertex, with the traversed edge (resp. arc) necessarily flipped. If loops are not allowed, then the agent cannot move to its current vertex.

Method	`__init__`	This constructor initializes an instance of the `LocalFlipEnvironment` class.
Method	`episode_length.setter`	This setter allows the user to potentially reconfigure the episode length between two independent batches of episodes. It should not be used while a batch of episodes is currently in progress.
Method	`state_batch_to_graph_batch`	This abstract method must be implemented by any concrete subclass. It extracts the batch of underlying graphs corresponding to a provided batch of states. Implementations must return a `Graph` object containing the graphs corresponding to each row in ...
Instance Variable	`initial_graph_generator`	A `GraphGenerator` function that defines how the underlying fully colored graphs are generated for the initial states. This attribute may be reconfigured between independent batches of episodes.
Instance Variable	`starting_vertex`	A nonnegative `int` below the configured graph order that determines the vertex at which the agent should start the potential flipping procedure. This attribute may be reconfigured between independent batches of episodes.
Property	`action_mask`	This abstract property must be implemented by any concrete subclass. It must return `None` if no episodes are currently being run in parallel, or if every action is available in every current state. Otherwise, it must return a two-dimensional ...
Property	`action_number`	This abstract property must be implemented by any concrete subclass. It must return the total number of distinct actions that can be executed in the environment, as a positive `int`.
Property	`episode_length`	This abstract property must be implemented by any concrete subclass. It must return the predetermined common length of all episodes run in parallel, i.e., the total number of actions executed in each episode, as a positive ...
Property	`is_continuing`	This abstract property must be implemented by any concrete subclass. It must return a `bool` indicating whether the environment is continuing (`True`) or episodic (`False`).
Property	`state_dtype`	This abstract property must be implemented by any concrete subclass. It must return the data type of the one-dimensional `numpy.ndarray` vectors that represent states, as a `numpy.dtype`.
Property	`state_length`	This abstract property must be implemented by any concrete subclass. It must return the number of entries in each state vector, i.e., the length of the one-dimensional `numpy.ndarray` vectors that represent states, as a positive ...
Method	`_initialize_batch`	This abstract method must be implemented by any concrete subclass. It must initialize a batch of episodes of the specified size and update the `_state_batch` and `_status` attributes so that they represent the newly initialized batch.
Method	`_transition_batch`	This method applies a batch of actions to the current batch of states and updates the `_state_batch` and `_status` attributes to reflect the resulting states and the updated batch status.
Instance Variable	`_allow_loops`	A `bool` indicating whether loops are allowed in the graphs to be constructed.
Instance Variable	`_current_vertices`	Either `None` or a `numpy.ndarray` of type `numpy.int32` specifying the vertex where the agent is currently located in each episode run in parallel.
Instance Variable	`_episode_length`	A positive `int` specifying the episode length, i.e., the total number of actions in each episode.
Instance Variable	`_flattened_length`	A positive `int` equal to the flattened length of the graphs to be constructed.
Instance Variable	`_flattened_ordering`	An item of the `FlattenedOrdering` enumeration specifying the edge (resp. arc) ordering (row-major or clockwise).
Instance Variable	`_flip_only`	A `bool` indicating whether all traversed edges (resp. arcs) must be flipped.
Instance Variable	`_graph_order`	A positive `int` that describes the order of the graphs to be constructed.
Instance Variable	`_is_directed`	A `bool` indicating whether the graphs to be constructed are directed or undirected.
Instance Variable	`_state_batch`	See the description of the `GraphEnvironment._state_batch` attribute.
Instance Variable	`_status`	See the description of the `GraphEnvironment._status` attribute.
Instance Variable	`_step_count`	Either `None` or a nonnegative `int` counting how many actions have been executed in the current batch of episodes. When `_step_count` equals `_episode_length`, the episode has reached a final state. This attribute is updated after each call to ...

Inherited from GraphEnvironment:

Method	`reset_batch`	This method initializes a batch of episodes of a specified size and returns the resulting batch of states, the corresponding values of the selected graph invariant (if computed), and the status of the batch of episodes...
Method	`state_to_graph`	This method extracts the underlying graph corresponding to a single state.
Method	`step_batch`	This method applies a batch of actions to the current batch of episodes and returns the resulting batch of states, the corresponding values of the selected graph invariant (if computed), and the updated status of the batch...
Instance Variable	`sparse_setting`	A `bool` indicating whether the graph invariant values should be computed only for the final batch of actions.
Instance Variable	`__graph_batch`	Either `None` or a `Graph` object representing the current batch of underlying graphs. This attribute is updated only when required by the sparse setting.
Instance Variable	`__graph_invariant`	A `GraphInvariant` function specifying the graph invariant to be maximized.
Instance Variable	`__graph_invariant_batch`	Either `None` or a one-dimensional `numpy.ndarray` of type `numpy.float32` containing the current batch of graph invariant values. As with `__graph_batch`, this attribute is updated only when required by the sparse setting.
Instance Variable	`__graph_invariant_diff`	Either `None`, indicating that graph invariant values are always computed directly using `__graph_invariant`, or a `GraphInvariantDiff` function used to incrementally update invariant values after state transitions.

def __init__(self, graph_invariant: GraphInvariant, graph_order: int, episode_length: int | None = None, flip_only: bool = False, flattened_ordering: FlattenedOrdering = FlattenedOrdering.ROW_MAJOR, is_directed: bool = False, allow_loops: bool = False, initial_graph_generator: GraphGenerator | None = None, starting_vertex: int = 0, graph_invariant_diff: GraphInvariantDiff | None = None, sparse_setting: bool = False): ¶

overrides rlgt.environments.graph_environment.GraphEnvironment.__init__

This constructor initializes an instance of the LocalFlipEnvironment class.

Parameters
graph_invariant:`GraphInvariant`	A `GraphInvariant` function that computes the graph invariant values associated with a batch of underlying graphs. These values are the quantities to be maximized by the environment.
graph_order:`int`	A positive `int` (not below 2) that represents the graph order of the graphs to be constructed.
episode_length:`int \| None`	Either `None`, or a positive `int` specifying the number of actions in each episode. If `None`, the episode length defaults to the flattened length of the graphs to be constructed. The default value is `None`.
flip_only:`bool`	A `bool` specifying whether the traversed edges (resp. arcs) must be flipped. The default value is `False`.
flattened_ordering:`FlattenedOrdering`	An item of the `FlattenedOrdering` enumeration specifying whether the edges (resp. arcs) are ordered row-major or clockwise. The default value is `FlattenedOrdering.ROW_MAJOR`.
is_directed:`bool`	A `bool` indicating whether the graphs to be constructed are directed. The default value is `False`.
allow_loops:`bool`	A `bool` indicating whether loops are allowed in the graphs to be constructed. The default value is `False`.
initial_graph_generator:`GraphGenerator \| None`	Either `None` or a `GraphGenerator` function that determines how the initial fully colored graphs are generated for the batch of initial states. If `None`, all edges (resp. arcs) in all graphs are initially colored with color 0. The default value is `None`.
starting_vertex:`int`	A nonnegative `int` strictly less than `graph_order` specifying the vertex at which the agent starts the potential flipping procedure. The default value is 0.
graph_invariant_diff:`GraphInvariantDiff \| None`	Either `None`, indicating that graph invariant values are always computed directly using `graph_invariant`, or a `GraphInvariantDiff` function that computes element-wise differences of the graph invariant values when the environment transitions from one batch of underlying graphs to another. The default value is `None`.
sparse_setting:`bool`	A `bool` indicating whether the sparse setting is enabled. If set to `True`, the graph invariant values are computed only for the final batch of actions. Otherwise, the graph invariant values are computed after every batch of actions. The default value is `False`.

@episode_length.setter

def episode_length(self, episode_length: int): ¶

This setter allows the user to potentially reconfigure the episode length between two independent batches of episodes. It should not be used while a batch of episodes is currently in progress.

Parameters
episode_length:`int`	A positive `int` specifying the new episode length.

def state_batch_to_graph_batch(self, state_batch: np.ndarray) -> Graph: ¶

overrides rlgt.environments.graph_environment.GraphEnvironment.state_batch_to_graph_batch

This abstract method must be implemented by any concrete subclass. It extracts the batch of underlying graphs corresponding to a provided batch of states. Implementations must return a Graph object containing the graphs corresponding to each row in state_batch, preserving the row order. This method must be pure and must not modify any attributes of the class instance.

Parameters
state_batch:`np.ndarray`	A two-dimensional `numpy.ndarray` whose rows represent individual states from which the underlying graphs are to be extracted.

Returns
`Graph`	A `Graph` object representing the extracted batch of graphs.

initial_graph_generator: GraphGenerator = ¶

A GraphGenerator function that defines how the underlying fully colored graphs are generated for the initial states. This attribute may be reconfigured between independent batches of episodes.

starting_vertex: int = ¶

A nonnegative int below the configured graph order that determines the vertex at which the agent should start the potential flipping procedure. This attribute may be reconfigured between independent batches of episodes.

@property

action_mask: np.ndarray | None = ¶

overrides rlgt.environments.graph_environment.GraphEnvironment.action_mask

This abstract property must be implemented by any concrete subclass. It must return None if no episodes are currently being run in parallel, or if every action is available in every current state. Otherwise, it must return a two-dimensional numpy.ndarray matrix a of type bool whose entry a[i, j] is True if and only if action j is available in the current state of the i-th episode.

@property

action_number: int = ¶

overrides rlgt.environments.graph_environment.GraphEnvironment.action_number

This abstract property must be implemented by any concrete subclass. It must return the total number of distinct actions that can be executed in the environment, as a positive int.

@property

episode_length: int = ¶

overrides rlgt.environments.graph_environment.GraphEnvironment.episode_length

This abstract property must be implemented by any concrete subclass. It must return the predetermined common length of all episodes run in parallel, i.e., the total number of actions executed in each episode, as a positive int.

@property

is_continuing: bool = ¶

overrides rlgt.environments.graph_environment.GraphEnvironment.is_continuing

This abstract property must be implemented by any concrete subclass. It must return a bool indicating whether the environment is continuing (True) or episodic (False).

@property

state_dtype: np.dtype = ¶

overrides rlgt.environments.graph_environment.GraphEnvironment.state_dtype

This abstract property must be implemented by any concrete subclass. It must return the data type of the one-dimensional numpy.ndarray vectors that represent states, as a numpy.dtype.

@property

state_length: int = ¶

overrides rlgt.environments.graph_environment.GraphEnvironment.state_length

This abstract property must be implemented by any concrete subclass. It must return the number of entries in each state vector, i.e., the length of the one-dimensional numpy.ndarray vectors that represent states, as a positive int.

def _initialize_batch(self, batch_size: int): ¶

overrides rlgt.environments.graph_environment.GraphEnvironment._initialize_batch

This abstract method must be implemented by any concrete subclass. It must initialize a batch of episodes of the specified size and update the _state_batch and _status attributes so that they represent the newly initialized batch.

Parameters
batch_size:`int`	The number of episodes to initialize in the batch, given as a positive `int`.

def _transition_batch(self, action_batch: np.ndarray): ¶

overrides rlgt.environments.graph_environment.GraphEnvironment._transition_batch

This method applies a batch of actions to the current batch of states and updates the _state_batch and _status attributes to reflect the resulting states and the updated batch status.

Parameters
action_batch:`np.ndarray`	A one-dimensional `numpy.ndarray` of type `numpy.int32` containing the actions to be applied. The length of `action_batch` must match the number of states in `_state_batch`.

Note
If loops are not allowed and an action attempts to traverse a loop, a `RuntimeError` is raised.

_allow_loops: bool = ¶

A bool indicating whether loops are allowed in the graphs to be constructed.

_current_vertices: np.ndarray | None = ¶

Either None or a numpy.ndarray of type numpy.int32 specifying the vertex where the agent is currently located in each episode run in parallel.

_episode_length = ¶

A positive int specifying the episode length, i.e., the total number of actions in each episode.

_flattened_length: int = ¶

A positive int equal to the flattened length of the graphs to be constructed.

_flattened_ordering: FlattenedOrdering = ¶

An item of the FlattenedOrdering enumeration specifying the edge (resp. arc) ordering (row-major or clockwise).

_flip_only: bool = ¶

A bool indicating whether all traversed edges (resp. arcs) must be flipped.

_graph_order: int = ¶

A positive int that describes the order of the graphs to be constructed.

_is_directed: bool = ¶

A bool indicating whether the graphs to be constructed are directed or undirected.

_state_batch = ¶

overrides rlgt.environments.graph_environment.GraphEnvironment._state_batch

See the description of the GraphEnvironment._state_batch attribute.

_status = ¶

overrides rlgt.environments.graph_environment.GraphEnvironment._status

See the description of the GraphEnvironment._status attribute.

_step_count: int | None = ¶

Either None or a nonnegative int counting how many actions have been executed in the current batch of episodes. When _step_count equals _episode_length, the episode has reached a final state. This attribute is updated after each call to GraphEnvironment.reset_batch or GraphEnvironment.step_batch.