Why use probability rather than logic for uncertain knowledge?

Strict logic forces statements to be true or false, which is awkward when knowledge is incomplete or evidence is partial. Probability represents graded degrees of belief and provides principled rules, such as Bayes' rule, for updating them with evidence, making it well suited to reasoning under uncertainty.

What makes Bayesian networks important for this area?

A full joint distribution over many variables is astronomically large, but Bayesian networks use conditional independence to represent it compactly as a graph with local conditional distributions. This makes both storing the model and computing probabilistic queries feasible, which is why they are a cornerstone of reasoning under uncertainty.

Reasoning Under Uncertainty

Reasoning under uncertainty is the part of artificial intelligence that uses probability and decision theory to draw conclusions and make decisions when knowledge is incomplete, noisy, or only partially observable.

Definition

Reasoning under uncertainty represents an agent's incomplete knowledge as probability distributions and computes, via the laws of probability and the maximization of expected utility, what to believe and how to act.

Scope

This area covers the representation of uncertain knowledge with probability and the methods for reasoning and deciding under it: probabilistic graphical models such as Bayesian networks, exact and approximate probabilistic inference, decision theory combining probabilities with utilities, and sequential decision making via Markov decision processes. It treats how degrees of belief are updated with evidence and how rational choices are computed. The data-driven estimation of these models, and reinforcement learning of policies, belong to the machine-learning subfield; this area emphasizes the representations and inference and decision principles.

Sub-topics

Core questions

How are degrees of belief represented and updated as new evidence arrives?
How can large joint distributions be represented compactly using conditional independence?
How is the probability of a query computed exactly or approximately in a probabilistic model?
How are probabilities combined with preferences to choose actions that maximize expected utility?

Key concepts

probability as degree of belief
Bayes' rule
conditional independence
Bayesian networks
exact and approximate inference
utility and expected utility
decision theory
Markov decision processes

Key theories

Bayesian updating: Bayes' rule prescribes how a prior degree of belief is revised into a posterior given evidence, providing the normative basis for probabilistic reasoning and for combining background knowledge with observations.
Graphical models and conditional independence: Bayesian and Markov networks exploit conditional independence to factor a joint distribution into local components, making both representation and inference tractable for problems that would otherwise be exponentially large.
Maximum expected utility: Decision theory holds that a rational agent should choose the action maximizing expected utility, unifying probabilistic belief with preferences over outcomes and extending to sequential decisions through Markov decision processes.

Clinical relevance

Probabilistic reasoning underlies medical diagnosis systems, fault diagnosis and sensor fusion, speech and language processing, robotics and localization, risk analysis, and recommendation and decision-support systems, wherever conclusions and choices must be made from incomplete or noisy information.

History

Early AI was skeptical of probability, favoring ad hoc certainty factors, but Pearl's work in the 1980s, culminating in his 1988 book, showed that Bayesian networks made probabilistic reasoning both well-founded and computationally feasible. Decision-theoretic and graphical-model methods, consolidated in texts such as Koller and Friedman (2009), became central to modern AI.

Debates

Probability vs. alternative uncertainty formalisms: Historically, AI debated whether to model uncertainty with probability or with alternatives such as certainty factors, fuzzy logic, or Dempster-Shafer belief functions; the probabilistic, decision-theoretic view became dominant largely because of its sound foundations and the tractability afforded by graphical models.

Key figures

Judea Pearl
Daphne Koller
Nir Friedman
Ross D. Shachter
Thomas Bayes

Seminal works

pearl1986
pearl1988
koller2009

Frequently asked questions

Why use probability rather than logic for uncertain knowledge?: Strict logic forces statements to be true or false, which is awkward when knowledge is incomplete or evidence is partial. Probability represents graded degrees of belief and provides principled rules, such as Bayes' rule, for updating them with evidence, making it well suited to reasoning under uncertainty.
What makes Bayesian networks important for this area?: A full joint distribution over many variables is astronomically large, but Bayesian networks use conditional independence to represent it compactly as a graph with local conditional distributions. This makes both storing the model and computing probabilistic queries feasible, which is why they are a cornerstone of reasoning under uncertainty.