Belief Revision¶

Incorporate newly accepted information into a logically represented belief state while restoring consistency and preserving as much of the prior commitment structure as a declared minimal-change policy warrants.

Version: v3 · 2026-09-19 · History
Domain-specific #: 1366
Domain group: Applied Sciences & Engineering
Origin domain: Computer Science & Software Engineering
Subdomain: Knowledge Representation → Computer Science & Software Engineering
Aliases: AGM belief revision, Belief change by revision

Core Idea¶

Belief Revision is the formal knowledge-representation operation that incorporates a newly accepted proposition into an existing belief set or epistemic state while resolving contradiction and preserving prior commitments according to a rational minimal-change policy. If a deductively closed belief set K is revised by proposition phi, the result is written K * phi. Success normally requires the result to include phi; consistency requires a consistent result when phi is consistent; preservation principles constrain which old consequences are lost. Alchourrón, Gärdenfors, and Makinson's 1985 paper establishes the canonical AGM theory through revision and partial-meet contraction functions.

How would you explain it like I'm…

Changing Your Mind Just a Little

Say you believe all birds can fly, and penguins are birds, so penguins can fly. Then you learn penguins can't fly. You have to change your mind, but you only give up as little as you need to, like deciding "not all birds can fly," instead of throwing away everything you knew about birds.

Fixing Beliefs That Clash

Belief Revision is the study of how a computer, or a person, should change what it believes when it learns something new that doesn't fit. You can't just add the new fact, because then your beliefs would contradict each other. You have to give up some old beliefs, and usually there's more than one choice of which ones to drop. The rule is to change as little as possible, using some ranking of which old beliefs you trust more. This is different from adding a fact that doesn't conflict, or from updating because the world itself changed, like a light that was off and is now on.

Minimal-Change Logical Revision

Belief Revision is a formal operation in knowledge representation for adding a newly accepted statement to a set of beliefs while keeping the set consistent and changing it as little as possible. If K is the belief set and φ the new statement, the result is written K * φ; it must include φ and must be consistent whenever φ is. Suppose K contains A, B and 'A and B imply C', and we learn 'not C': we can't just add it, so at least one of A, B or the rule must go, and a ranking of plausibility decides which. The standard theory is AGM, from Alchourrón, Gärdenfors and Makinson (1985). Revision differs from expansion (adding without conflict), contraction (giving up a belief without accepting its opposite) and update (the world itself changed). It is also different from Bayesian updating, which shifts numerical probabilities rather than choosing which full beliefs to keep.

Belief revision, in knowledge representation, is the operation that incorporates a newly accepted proposition φ into a deductively closed belief set K (or richer epistemic state), yielding K*φ, while restoring consistency and preserving prior commitments under a minimal-change policy. The AGM theory of Alchourrón, Gärdenfors, and Makinson gives rationality postulates, including success (φ ∈ K*φ), consistency (K*φ is consistent if φ is), closure, extensionality, and preservation constraints, and constructs revision via partial-meet contraction. The problem is nontrivial because classical inconsistency is explosive and multiple repairs are possible: if K derives C from A, B, and A∧B→C, revising by ¬C requires relinquishing at least one support, chosen by a selection or plausibility ordering. Katsuno and Mendelzon characterize AGM revision model-theoretically as selecting the minimal models of φ under a faithful preorder tied to K. They also distinguish revision, where old and new information describe the same situation, from update, where the world has changed. Unlike Bayesian updating, which conditions graded probabilities, AGM revision changes categorical acceptance and prioritizes consistency and minimal loss; the two can be related through ranking functions but are not the same operation.

Scope of Application¶

Belief Revision is literal when a represented epistemic state incorporates accepted information, repairs conflict, and selects retained commitments through an explicit qualitative rationality or minimal-change policy.

Knowledge representation. Maintaining logically coherent agent theories after new inputs.
Artificial intelligence. Designing rational qualitative belief-change operators.
Database integrity. Distinguishing correction of beliefs from update of a changing world.
Nonmonotonic reasoning. Retracting consequences when premises lose standing.
Epistemic logic. Representing contraction, revision, and iterated change.
Ontology maintenance. Repairing inconsistent axiom sets under priorities.
Theory change. Comparing entrenchment and remainder-selection policies.
Hybrid uncertainty models. Relating qualitative rankings to probabilistic representations without collapse.

Clarity¶

A clear revision problem states the base logic, whether the state is a deductively closed belief set or finite base, the new formula, input reliability, operator notation, rationality postulates, minimal-change representation, and treatment of inconsistency. It distinguishes same-world correction from changed-world update and qualitative acceptance from probability revision. An implementation gives the ordering, distance, entrenchment, or selection function rather than appealing to similarity informally. Iterated revision requires an epistemic state richer than a bare consequence set if future priorities matter.

Manages Complexity¶

Belief revision compresses a combinatorial repair problem into an operator constrained by postulates and a preference representation. Instead of enumerating every arbitrary deletion, it preserves theories closest to the prior state while enforcing success and consistency. The AGM framework also exposes impossibility and underdetermination: postulates characterize rational families but do not uniquely choose one policy. Complexity returns through large model spaces, inconsistent bases, nonclassical logics, iterated inputs, source reliability, and relevance. The abstraction manages this by separating logical requirements from the extra ordering that carries domain priorities.

Abstract Reasoning¶

Represent the current belief set or richer epistemic state in a declared logic.
Receive the new proposition and determine whether it is to be accepted categorically.
Test whether simple expansion preserves consistency and other integrity conditions.
If conflict occurs, generate or characterize the maximal coherent ways to relinquish old commitments.
Apply an entrenchment, remainder-selection, distance, or plausibility policy.
Construct the successor state that includes the input and preserves preferred prior content.

Knowledge Transfer¶

Belief Revision transfers a general constrained-state-repair pattern: accept a binding new condition, locate conflict, preserve a maximal or preferred part of the prior state, and expose the policy that decides among repairs. The pattern applies to configuration management, ontology repair, legal-rule maintenance, and requirements change. The logical accent remains important: equivalence, consequence closure, categorical acceptance, and postulates differ from numerical Bayesian learning or ordinary document versioning. Minimal change is not self-defining; every transferred use must declare its distance or priority structure.

Relationships to Other Abstractions¶

Current abstraction Belief Revision Domain-specific

Parents (1) — outgoing typed relations

Belief Revision is a kind of State and State Transition Prime

State and State Transition is the strict parent by specialization.

Hierarchy path (1) — routes to 1 parentless root

Belief Revision → is a kind of State and State Transition → includes Phase Space

Neighborhood in Abstraction Space¶

Belief Revision sits in a sparse region of the domain-specific corpus (93^rd percentile for distinctiveness): few abstractions share its structure, so a faithful description tends to retrieve it precisely.

Family — Unclustered & Miscellaneous (2769 abstractions)

Nearest neighbors

Computed from structural-signature embeddings · 2026-10-25