ontolearn.learners.drill

Classes

Drill	Neuro-Symbolic Class Expression Learning (https://www.ijcai.org/proceedings/2023/0403.pdf)
DrillHeuristic	Heuristic in Convolutional DQL concept learning.
DrillNet	A neural model for Deep Q-Learning.

Module Contents

class ontolearn.learners.drill.Drill(knowledge_base, path_embeddings: str = None, refinement_operator: LengthBasedRefinement = None, use_inverse: bool = True, use_data_properties: bool = True, use_card_restrictions: bool = True, use_nominals: bool = True, min_cardinality_restriction: int = 2, max_cardinality_restriction: int = 5, positive_type_bias: int = 1, quality_func: Callable = None, reward_func: object = None, batch_size=None, num_workers: int = 1, iter_bound=None, max_num_of_concepts_tested=None, verbose: int = 0, terminate_on_goal=None, max_len_replay_memory=256, epsilon_decay: float = 0.01, epsilon_min: float = 0.0, num_epochs_per_replay: int = 2, num_episodes_per_replay: int = 2, learning_rate: float = 0.001, max_runtime=None, num_of_sequential_actions=3, stop_at_goal=True, num_episode: int = 10)

Bases: ontolearn.base_concept_learner.RefinementBasedConceptLearner

Neuro-Symbolic Class Expression Learning (https://www.ijcai.org/proceedings/2023/0403.pdf)

name = 'DRILL'

verbose

learning_problem = None

num_workers

learning_rate

num_episode

num_of_sequential_actions

num_epochs_per_replay

max_len_replay_memory

epsilon_decay

epsilon_min

batch_size

num_episodes_per_replay

seen_examples

pos: FrozenSet[owlapy.owl_individual.OWLNamedIndividual] = None

neg: FrozenSet[owlapy.owl_individual.OWLNamedIndividual] = None

positive_type_bias

start_time = None

goal_found = False

search_tree

stop_at_goal

epsilon = 1

quality_func

initialize_training_class_expression_learning_problem(pos: FrozenSet[owlapy.owl_individual.OWLNamedIndividual], neg: FrozenSet[owlapy.owl_individual.OWLNamedIndividual]) → RL_State

Initialize

rl_learning_loop(pos_uri: FrozenSet[owlapy.owl_individual.OWLNamedIndividual], neg_uri: FrozenSet[owlapy.owl_individual.OWLNamedIndividual]) → List[float]

Reinforcement Learning Training Loop

Initialize RL environment for a given learning problem (E^+ pos_iri and E^- neg_iri )

Training:

2.1 Obtain a trajectory: A sequence of RL states/DL concepts T, Person, (Female and

orall hasSibling Female).

Rewards at each transition are also computed

train(dataset: Iterable[Tuple[str, Set, Set]] | None = None, num_of_target_concepts: int = 1, num_learning_problems: int = 1)

Training RL agent (1) Generate Learning Problems (2) For each learning problem, perform the RL loop

save(directory: str = None) → None

save weights of the deep Q-network

load(directory: str = None) → None

load weights of the deep Q-network

fit(learning_problem: PosNegLPStandard, max_runtime=None)

Run the concept learning algorithm according to its configuration.

Once finished, the results can be queried with the best_hypotheses function.

init_embeddings_of_examples(pos_uri: FrozenSet[owlapy.owl_individual.OWLNamedIndividual], neg_uri: FrozenSet[owlapy.owl_individual.OWLNamedIndividual])

create_rl_state(c: owlapy.class_expression.OWLClassExpression, parent_node: RL_State | None = None, is_root: bool = False) → RL_State

Create an RL_State instance.

compute_quality_of_class_expression(state: RL_State) → None

Compute Quality of owl class expression. # (1) Perform concept retrieval # (2) Compute the quality w.r.t. (1), positive and negative examples # (3) Increment the number of tested concepts attribute.

apply_refinement(rl_state: RL_State) → Generator

Downward refinements

select_next_state(current_state, next_rl_states) → Tuple[RL_State, float]

sequence_of_actions(root_rl_state: RL_State) → Tuple[List[Tuple[RL_State, RL_State]], List[SupportsFloat]]

Performing sequence of actions in an RL env whose root state is ⊤

form_experiences(state_pairs: List, rewards: List) → None

Form experiences from a sequence of concepts and corresponding rewards.

state_pairs - A list of tuples containing two consecutive states. reward - A list of reward.

Gamma is 1.

Return X - A list of embeddings of current concept, next concept, positive examples, negative examples. y - Argmax Q value.

learn_from_replay_memory() → None

Learning by replaying memory.

update_search(concepts, predicted_Q_values=None)

@param concepts: @param predicted_Q_values: @return:

get_embeddings_individuals(individuals: List[str]) → torch.FloatTensor

get_individuals(rl_state: RL_State) → List[str]

assign_embeddings(rl_state: RL_State) → None

Assign embeddings to a rl state. A rl state is represented with vector representation of all individuals belonging to a respective OWLClassExpression.

save_weights(path: str = None) → None

Save weights DQL

exploration_exploitation_tradeoff(current_state: AbstractNode, next_states: List[AbstractNode]) → AbstractNode

Exploration vs Exploitation tradeoff at finding next state. (1) Exploration. (2) Exploitation.

exploitation(current_state: AbstractNode, next_states: List[AbstractNode]) → RL_State

Find next node that is assigned with highest predicted Q value.

1. Predict Q values : predictions.shape => torch.Size([n, 1]) where n = len(next_states).

2. Find the index of max value in predictions.

3. Use the index to obtain next state.

4. Return next state.

predict_values(current_state: RL_State, next_states: List[RL_State]) → torch.Tensor

Predict promise of next states given current state.

Returns:

Predicted Q values.

static retrieve_concept_chain(rl_state: RL_State) → List[RL_State]

generate_learning_problems(num_of_target_concepts, num_learning_problems) → List[Tuple[str, Set, Set]]

Generate learning problems if none is provided.

Time complexity: O(n^2) n = named concepts

learn_from_illustration(sequence_of_goal_path: List[RL_State])

Parameters:

sequence_of_goal_path – ⊤,Parent,Parent ⊓ Daughter.

best_hypotheses(n=1, return_node: bool = False) → owlapy.class_expression.OWLClassExpression | List[owlapy.class_expression.OWLClassExpression]

Get the current best found hypotheses according to the quality.

Parameters:

n – Maximum number of results.

Returns:

Iterable with hypotheses in form of search tree nodes.

clean()

Clear all states of the concept learner.

next_node_to_expand() → RL_State

Return a node that maximizes the heuristic function at time t.

downward_refinement(*args, **kwargs)

Execute one refinement step of a refinement based learning algorithm.

Parameters:

node (_N) – the search tree node on which to refine.

Returns:

Refinement results as new search tree nodes (they still need to be added to the tree).

Return type:

Iterable[_N]

show_search_tree(heading_step: str, top_n: int = 10) → None

A debugging function to print out the current search tree and the current n best found hypotheses to standard output.

Parameters:

• heading_step – A message to display at the beginning of the output.

• top_n – The number of current best hypotheses to print out.

terminate_training()

class ontolearn.learners.drill.DrillHeuristic(pos=None, neg=None, model=None, mode=None, model_args=None)

Heuristic in Convolutional DQL concept learning. Heuristic implements a convolutional neural network.

score(node, parent_node=None)

Compute heuristic value of root node only

apply(node, parent_node=None)

Assign predicted Q-value to node object.

class ontolearn.learners.drill.DrillNet(args)

Bases: torch.nn.Module

A neural model for Deep Q-Learning.

An input Drill has the following form:

1. Indexes of individuals belonging to current state (s).

2. Indexes of individuals belonging to next state (s_prime).

3. Indexes of individuals provided as positive examples.

4. Indexes of individuals provided as negative examples.

Given such input, we from a sparse 3D Tensor where each slice is a **** N * by ***D* where N is the number of individuals and D is the number of dimension of embeddings. Given that N on the current benchmark datasets < 10^3, we can get away with this computation. By doing so we do not need to subsample from given inputs.

loss

conv1

size_of_fc1

fc1

fc2

init()

forward(X: torch.FloatTensor)

X n by 4 by d float tensor