Bayesian networks
Era 3 · From the expert boom to the statistical turn (1980–2011) › Statistical learning and probabilistic models
- Lineage
- Data and learning
- Kind
- Publication
Who
UCLA
What happened
Judea Pearl of the University of California, Los Angeles, coins the term "Bayesian network" in 1985 and in 1988 publishes Probabilistic Reasoning in Intelligent Systems, which establishes the field.
Why it mattered
A directed graph shows which variables depend directly on which, and it makes efficient reasoning under uncertainty possible. AI moves from rules with certainty factors to well-founded probability.
The full story
Expert systems handled uncertainty with makeshift devices, such as MYCIN's certainty factors, because probability seemed unmanageable: with many variables, the joint distribution has an astronomical number of values. Pearl, a computer science professor at UCLA since 1970, found a way to tame it.
The key is that, in practice, each variable depends directly on only a few others. A Bayesian network is a directed graph with no cycles: each node is a variable, an arrow marks a direct dependence, and each node stores the probability of its values given those of its parents. That is enough to reconstruct the whole distribution, and the structure of the graph tells you which variables are independent of which, given certain information.
Pearl also developed algorithms for updating beliefs when evidence arrives, by passing messages between neighboring nodes, what is now called belief propagation. The 1988 book brought together the theory, the algorithms, and a philosophy of reasoning, and it helped bring about a revolution in AI, which adopted probability as the language for reasoning with incomplete information.
Pearl later went a step further, from probabilistic reasoning to causal reasoning: telling apart what happens when a variable is observed from what happens when someone intervenes on it. For this body of work Pearl received the Turing Award in 2011.
Three nodes joined by arrows: Rain points to Sprinkler and to Wet grass, and Sprinkler points to Wet grass. Each node stores only the probability of its value given those of its parents. A note says that, on seeing wet grass, you can work out whether it rained.
Under the hood
A Bayesian network is a directed graph with no cycles. Each node is a variable and stores only the table P(Xᵢ | parents(Xᵢ)). The joint distribution is the product of those tables: P(X₁, …, Xₙ) = Π P(Xᵢ | parents(Xᵢ)).
The savings are enormous. With 30 yes-or-no variables, the full joint table needs 2³⁰ − 1 numbers, more than a billion. If each variable has at most three parents, 30 · 2³ = 240 are enough.
In the diagram's example, seeing wet grass raises the probability that it rained and the probability that the sprinkler was on. If you also learn that the sprinkler was on, the probability of rain drops again: the sprinkler already explains the grass. This reasoning between rival causes is known as explaining away. Pearl's belief propagation algorithm computes it exactly when the graph has no cycles, even ignoring the direction of the arrows.
Common misconceptions
What people often say: The arrows in a Bayesian network show causes.
In fact: They show dependencies between probabilities, and the same distribution can be described with arrows pointing the other way. Reading them as causes takes additional assumptions, and that is what Pearl later formalized with causal models.
Videos
- Judea Pearl, 2012 ACM A.M. Turing Award Lecture "The Mechanization of Causal Inference" (opens in a new tab) (Association for Computing Machinery (ACM))
Builds on
MYCIN (1972)
Inherited, with no proven direct link
Bayesian networks replace MYCIN's certainty factors as a way of reasoning under uncertainty
Sources
Scholarship and context
Further reading
How to cite
“Bayesian networks.” The History of AI, revised November 13, 2026. https://thehistoryofai.com/en/milestone/pearl-bayesian-networks
This entry was last reviewed on November 13, 2026.