Approximate Inference Using Sampling

class pgmpy.inference.ApproxInference.ApproxInference(model)[source]

Initializes the Approximate Inference class.

Parameters:

model (Instance of pgmpy.models.DiscreteBayesianNetwork or pgmpy.models.DynamicBayesianNetwork)

Examples

>>> from pgmpy.utils import get_example_model
>>> model = get_example_model('alarm')
>>> infer = ApproxInference(model)
get_distribution(samples, variables, state_names=None, joint=True)[source]

Computes distribution of variables from given data samples.

Parameters:

  • samples (pandas.DataFrame) – A dataframe of samples generated from the model.

  • variables (list (array-like)) – A list of variables whose distribution needs to be computed.

  • state_names (dict (default: None)) – A dict of state names for each variable in variables in the form {variable_name: list of states}. If None, inferred from the data but is possible that the final distribution misses some states.

  • joint (boolean) – If joint=True, computes the joint distribution over variables. Else, returns a dict with marginal distribution of each variable in variables.

map_query(variables, n_samples=10000, samples=None, evidence=None, virtual_evidence=None, state_names=None, show_progress=True, seed=None)[source]

Finds the most probable state in the joint distribution of variables. Calculates the result by generating samples and calculating most probable states based on the probabilities.

Parameters:

  • variables (list) – List of variables for which the probability distribution needs to be calculated.

  • n_samples (int) – The number of samples to generate for computing the distributions. Higher n_samples results in more accurate results at the cost of more computation time.

  • samples (pd.DataFrame (default: None)) – If provided, uses these samples to compute the distribution instead of generating samples. samples must conform with the provided evidence and virtual_evidence.

  • evidence (dict (default: None)) – The observed values. A dict key, value pair of the form {var: state_name}.

  • virtual_evidence (list (default: None)) – A list of pgmpy.factors.discrete.TabularCPD representing the virtual/soft evidence.

  • state_names (dict (default: None)) – A dict of state names for each variable in variables in the form {variable_name: list of states}. If None, inferred from the data but is possible that the final distribution misses some states.

  • show_progress (boolean (default: True)) – If True, shows a progress bar when generating samples.

  • seed (int (default: None)) – Sets the seed for the random generators.

Returns:

MAP values – The most probable state of provided variables given the evidence.

Return type:

dict

Examples

>>> from pgmpy.utils import get_example_model
>>> from pgmpy.inference import ApproxInference
>>> from pgmpy.factors.discrete import State,TabularCPD
>>> model = get_example_model("alarm")
>>> infer = ApproxInference(model)
>>> print(infer.map_query(variables=["HISTORY", "CVP"]))
{'HISTORY': 'FALSE', 'CVP': 'NORMAL'}
>>> virtual_evidence_history = TabularCPD(variable='HISTORY', variable_card=2,values=[[0.99],[0.01]],
...                                  state_names={"HISTORY": ["TRUE", "FALSE"]})
>>> evidence = {'CVP':'NORMAL'}
>>> print(infer.map_query(variables=["HISTORY"], evidence=evidence, virtual_evidence=[virtual_evidence_history]))
{'HISTORY': 'TRUE'}
query(variables, n_samples=10000, samples=None, evidence=None, virtual_evidence=None, joint=True, state_names=None, show_progress=True, seed=None)[source]

Method for doing approximate inference based on sampling in Bayesian Networks and Dynamic Bayesian Networks.

Parameters:

  • variables (list) – List of variables for which the probability distribution needs to be calculated.

  • n_samples (int) – The number of samples to generate for computing the distributions. Higher n_samples results in more accurate results at the cost of more computation time.

  • samples (pd.DataFrame (default: None)) – If provided, uses these samples to compute the distribution instead of generating samples. samples must conform with the provided evidence and virtual_evidence.

  • evidence (dict (default: None)) – The observed values. A dict key, value pair of the form {var: state_name}.

  • virtual_evidence (list (default: None)) – A list of pgmpy.factors.discrete.TabularCPD representing the virtual/soft evidence.

  • state_names (dict (default: None)) – A dict of state names for each variable in variables in the form {variable_name: list of states}. If None, inferred from the data but is possible that the final distribution misses some states.

  • show_progress (boolean (default: True)) – If True, shows a progress bar when generating samples.

  • seed (int (default: None)) – Sets the seed for the random generators.

Returns:

Probability distribution – The queried probability distribution.

Return type:

pgmpy.factors.discrete.TabularCPD

Examples

>>> from pgmpy.utils import get_example_model
>>> from pgmpy.inference import ApproxInference
>>> model = get_example_model("alarm")
>>> infer = ApproxInference(model)
>>> infer.query(variables=["HISTORY"])
<DiscreteFactor representing phi(HISTORY:2) at 0x7f92d9f5b910>
>>> infer.query(variables=["HISTORY", "CVP"], joint=True)
<DiscreteFactor representing phi(HISTORY:2, CVP:3) at 0x7f92d9f77610>
>>> infer.query(variables=["HISTORY", "CVP"], joint=False)
{'HISTORY': <DiscreteFactor representing phi(HISTORY:2) at 0x7f92dc61eb50>,
 'CVP': <DiscreteFactor representing phi(CVP:3) at 0x7f92d915ec40>}