Does inference follow arrows only?

No. Observed effects can update beliefs about upstream variables through the joint distribution.

Can it represent continuous variables?

Yes, using suitable local conditional distributions rather than only discrete tables.

Must its parameters have Bayesian priors?

No. The network representation does not mandate one particular parameter-estimation method.

AI Glossary / Statistics and probability / Intermediate

Bayesian network

Definition

A Bayesian network is a probabilistic model represented by a directed acyclic graph and a conditional probability distribution for each variable given its parents. Together, these specify a joint distribution through a product of local conditional distributions. The graph encodes conditional-independence assumptions that can make complex uncertainty easier to represent and reason about.

What it means

The nodes represent random variables, such as equipment state, sensor warning, and service interruption. Directed edges identify the parent variables used in each node's local probability model.

The graph alone is incomplete: numerical distributions are also required. A connection does not merely attach a probability to an arrow, and an absent direct edge does not automatically imply unconditional independence between its endpoints.

Bayesian networks support reasoning from observations to unobserved variables in either direction through the represented distribution. Despite the name, their parameters can be estimated using Bayesian or other statistical methods. They are not neural networks, and their arrows do not automatically establish causal effects without additional causal assumptions.

How it works

Choose clearly defined variables and possible values, then specify an acyclic parent structure. For each node X_i, provide P(X_i given its parents); a node with no parents has an unconditional distribution.

The joint probability is the product of these local distributions. For discrete variables, a conditional probability table must sum to one for each parent configuration. Continuous or mixed-variable networks need appropriate conditional densities or other valid local models.

To answer a query, incorporate the observed evidence, sum or integrate over irrelevant unobserved variables, and normalize. Algorithms such as variable elimination exploit the factorization, while sampling and other approximations can help when exact inference is costly. Graph structure strongly influences computational difficulty; a compact-looking graph need not make every query easy.

Conditional independencies can be read using the appropriate graphical rules, including d-separation. Chains, common causes, and common effects behave differently under conditioning. Learning the structure from observational data may leave several compatible graphs, and a causal interpretation requires assumptions beyond an ordinary probabilistic factorization.

Three practical examples

Troubleshooting: combine uncertain symptoms and component states to update beliefs about possible faults.
Operational forecasting: represent relationships among demand, capacity, delays, and observed indicators, then query likely outcomes under observed conditions.
Data completion: infer distributions for missing variables using observed information when the network and missingness assumptions support that analysis.

Think of it this way

Think of a set of local probability instructions that fit together into one joint picture. Each instruction describes one variable using its parents. The graph tells you how the instructions connect, while inference combines them to answer questions that no single local instruction can answer alone.

Worked example

Consider three binary variables: fault F, warning W, and shutdown S. Specify the chain F → W → S, meaning shutdown is conditionally independent of fault once warning status is known in this model.

Let P(F) = 0.10; P(W given F) = 0.80; P(W given no F) = 0.05; P(S given W) = 0.60; and P(S given no W) = 0.02. Complements supply the corresponding false-state probabilities.

The joint probability of fault, warning, and shutdown is 0.10 × 0.80 × 0.60 = 0.048. The marginal warning probability is 0.10 × 0.80 + 0.90 × 0.05 = 0.125.

The shutdown probability is therefore 0.125 × 0.60 + 0.875 × 0.02 = 0.0925. Observing a warning changes the fault probability to (0.10 × 0.80) / 0.125 = 0.64.

Under the specified chain, also observing shutdown after warning does not further change that fault probability: P(F given W,S) remains 0.64. This is a consequence of the model's conditional independence, not a universal engineering claim. A real system with a direct fault-to-shutdown relationship would require a different model.

Illustrative case study

Fictional case: Beech Maintenance builds a diagnostic aid using equipment states and sensor reports. Engineers define variables carefully and review the proposed graph before analysts estimate local probabilities.

During testing, the team finds that two sensors share a power supply. Treating their readings as independent evidence exaggerates confidence in some diagnoses. The model is revised to represent the shared influence, and its predictions are checked on held-out incidents.

Technicians receive probability-ranked explanations with the supporting observations. The tool does not treat the highest-probability explanation as a confirmed fault, and repair decisions incorporate inspection evidence and operational consequences beyond the network's probability calculations.

Common misconceptions and mistakes

Do not interpret every arrow as a proven causal mechanism. Observationally equivalent structures can represent the same conditional-independence relationships.

Do not assume no direct edge means no association. Variables may be linked through other paths, and conditioning can either remove or create dependence depending on the structure.

Do not multiply several symptom likelihoods as independent unless the relevant conditional independence is part of a defensible model. Shared causes can make that shortcut misleading.

Do not confuse a well-formed probability table with an accurate model. Missing variables, sparse data, changing populations, and incorrect structure can all undermine predictions. Feedback over time can be represented by time-indexed variables, but a standard Bayesian network cannot contain a directed cycle within its graph.

Business use

Bayesian networks can help teams combine domain knowledge and data into an inspectable uncertainty model. They are especially useful when users need to query several different variables after different combinations of observations.

Successful use requires clear variable definitions, appropriate probability estimates, and validation of both predictions and structural assumptions. Document whether the network is intended for association, diagnosis, or causal intervention. A graph that supports observing an event does not automatically support predicting what happens when the business actively forces that event to occur.

Check your understanding

What two components specify a Bayesian network model?

Answer: An acyclic directed graph and valid local conditional distributions for its variables.

Why does shutdown add no fault information after warning in the example?

Answer: The chosen chain explicitly assumes conditional independence of fault and shutdown given warning.

Related terms

Learn next

Construct a three-variable network and enumerate its joint probabilities. Then compare chain, fork, and collider structures to see how observing different variables changes independence relationships.

Sources

Poole and Mackworth, Artificial Intelligence: Foundations of Computational Agents, third edition, Section 9.3, Belief Networks, for factorization, conditional independence, and evidence queries: https://www.cs.ubc.ca/~poole/aibook/3e/html/ArtInt3e.Ch9.S3.html The fault-warning-shutdown network is an original illustrative model with independently checked probabilities.

Previous: Bayesian multivariate linear regression Next: Bayesian neural network