pyAgrum-nightly: Probabilistic Graphical Models

Common errors

• IOError: [pyAgrum] I/O Error: Stream states flags are not all unset

  cause This error often occurs when attempting to save a Bayesian network or other graphical model to a file, but the specified output directory does not exist.
  fix Ensure that the directory path provided to save functions (e.g., `gum.saveBN(bn, 'path/to/directory/filename.bif')`) exists before calling the function. Create the directory if it doesn't exist.

• DatabaseError: [pyAgrum] Database error: The conditioning set <COL=0, RA=4, INC=0> for target node PARTNER never appears in the database.

  cause This error typically arises during Bayesian network learning or inference when using data, especially with small datasets or specific data splits, where a particular combination of variable values (a 'conditioning set') exists in the query/test set but was never observed in the training data.
  fix Consider using 'smoothing' or adding 'priors' during the learning phase (e.g., using `BNLearner.useAprioriSmoothing()`) to handle unseen combinations. Also, evaluate the representativeness of your dataset and data splitting strategy.

Warnings

• breaking The main package import name changed from `pyAgrum` to `pyagrum` (all lowercase) starting from version 2.0.0. Older code using `import pyAgrum as gum` will fail.
  Fix: Update all `import pyAgrum` statements to `import pyagrum`.
• gotcha When defining Conditional Probability Tables (CPTs), especially for complex networks, directly assigning values can lead to indexing errors. pyAgrum recommends using dictionaries for clarity and to prevent common mistakes when introducing data.
  Fix: Always use dictionary-based assignments for CPTs (e.g., `bn.cpt("s")[{"c": 0}] = [0.5, 0.5]`) instead of direct list assignment, unless you are absolutely sure of the indexing order.

Install

• pip install pyAgrum-nightly Install latest nightly build

Imports

• pyagrum
  wrong:

  import pyAgrum as gum

  correct:

  import pyagrum as gum

  Since pyAgrum 2.0.0, the package name follows PEP8 rules and is lowercase. Using 'pyAgrum' will result in an ImportError.

Quickstart

This quickstart demonstrates how to create a simple Bayesian network (the 'Water Sprinkler' example), define its topology and conditional probability tables, and perform inference with evidence. It highlights the use of `pyagrum` (lowercase) for imports and dictionary-based CPT assignments.

import pyagrum as gum

# Create a Bayesian Network
bn = gum.BayesNet("WaterSprinkler")

# Add variables
id_c = bn.add(gum.LabelizedVariable("c", "cloudy ?", 2))
id_s = bn.add(gum.LabelizedVariable("s", "sprinkler ?", 2))
id_r = bn.add(gum.LabelizedVariable("r", "rain ?", 2))
id_w = bn.add(gum.LabelizedVariable("w", "wet grass ?", 2))

# Add arcs (dependencies)
bn.addArc(id_c, id_s)
bn.addArc(id_c, id_r)
bn.addArc(id_s, id_w)
bn.addArc(id_r, id_w)

# Define Conditional Probability Tables (CPTs) using dictionaries
bn.cpt("c").fillWith([0.5, 0.5])
bn.cpt("s")[{"c": 0}] = [0.5, 0.5]
bn.cpt("s")[{"c": 1}] = [0.9, 0.1]
bn.cpt("r")[{"c": 0}] = [0.8, 0.2]
bn.cpt("r")[{"c": 1}] = [0.2, 0.8]
bn.cpt("w")[{"s": 0, "r": 0}] = [1, 0]
bn.cpt("w")[{"s": 0, "r": 1}] = [0.1, 0.9]
bn.cpt("w")[{"s": 1, "r": 0}] = [0.1, 0.9]
bn.cpt("w")[{"s": 1, "r": 1}] = [0.01, 0.99]

# Perform inference
ie = gum.LazyPropagation(bn)
ie.setEvidence({"w": 1}) # Wet grass is true
ie.makeInference()

# Get posterior probabilities
posterior_c = ie.posterior("c")
print(f"P(c|w=1): {posterior_c}")