AI Glossary / Graphs and symbolic AI / Intermediate
Circumscription
Definition
Circumscription is a formal approach to nonmonotonic reasoning that prefers models in which selected properties hold for as few entities as possible, subject to the stated facts and rules. It can express defaults such as treating cases as normal unless an exception is required. New information may change the preferred models and withdraw an earlier conclusion.
What it means
Ordinary logical deduction preserves earlier consequences when consistent new premises are added. Everyday defaults behave differently: a planner may initially assume a device is available, then withdraw that assumption when a fault is reported. Circumscription gives a mathematical account of certain kinds of such reasoning.
The crucial choice is what to minimize. A model might minimize an abnormality predicate while keeping the known device identities and categories fixed and allowing other properties to vary. Minimizing every predicate indiscriminately is not the method. Different choices of minimized, fixed, and varying predicates can produce different conclusions.
How it works
Begin with a logical theory containing facts and rules. Identify a predicate whose extension - the collection of objects satisfying it - should be minimized. Compare models that satisfy the theory under the chosen rules for what stays fixed and what may vary. Prefer models with no strictly smaller eligible extension of the minimized predicate.
For default reasoning, an abnormality predicate often marks exceptions. A rule can say that normal eligible objects have a certain property; circumscription discourages unsupported exceptions. This is usually subset minimization, not simply picking the smallest numerical count in every possible interpretation. There can be several incomparable minimal models, so a cautious conclusion must hold in all preferred models.
Three practical examples
- A symbolic planning model assumes no unreported event blocks an action, while allowing explicit blocking facts to defeat that default.
- A knowledge representation exercise treats ordinary members of a category as having a typical property unless the theory requires an exception.
- A configuration explanation distinguishes facts explicitly recorded about an item from conclusions obtained by minimizing a designated abnormality relation.
Think of it this way
Think of maintaining an exceptions register and adding entries only when the available information requires them. The analogy captures a preference for fewer unsupported exceptions. It is limited because formal circumscription compares logical models, not a clerk's confidence, and missing evidence in the real world does not automatically justify assuming normality.
Worked example
Consider a deliberately small theory about one named device, D. The facts say that D is a device. The rule says: if D is a device and D is not abnormal, then D is usable. Keep the device fact fixed, minimize the abnormality predicate, and allow usability to vary.
Initially, a model with D not abnormal satisfies the theory, provided D is usable. That model is preferred to an otherwise eligible model that unnecessarily marks D abnormal. The circumscribed theory therefore supports usability in this simple setting.
Now add the fact that D is not usable. Any consistent model must mark D abnormal, because leaving it normal would force usability and conflict with the new fact. The preferred models change, and the earlier default conclusion that D is usable is no longer retained. This is revision through recomputing the circumscribed theory, not a claim that classical logic can keep both usability and non-usability without inconsistency. The example assumes only the stated vocabulary and comparison choices.
Illustrative case study
At fictional Alder Lab, a teaching team models whether a demonstration device can be used. A first version treats every unreported problem as impossible. Reviewers point out that this hides the distinction between a default and an observed condition.
The revised explanation identifies which conclusion depends on minimizing exceptions and shows how an explicit fault changes that conclusion. The team does not use the teaching default as authorization to operate real equipment. This fictional case illustrates transparent assumptions in a symbolic model, rather than measured operational improvements or a recommended safety procedure.
Common misconceptions and mistakes
Circumscription is not the assertion that anything unknown is false. Its effect depends on a deliberately selected minimization policy and the surrounding theory. It is also not a probability estimate: preferring normal models does not attach a numerical likelihood to normality.
Fewer exceptions by set inclusion should not be confused with optimizing an arbitrary cost function. Multiple minimal models can remain, and they may disagree about a proposed conclusion. Another mistake is to treat a default as an unchangeable fact. The purpose of nonmonotonic reasoning is precisely to allow some conclusions to change when relevant evidence arrives. Poorly chosen fixed and varying predicates can create unintended results.
Business use
Circumscription is useful for understanding the assumptions hidden inside rule-based reasoning and planning. Before using a related default in a business workflow, document what is minimized, which facts are authoritative, and what evidence defeats the default. Ask whether absence of a record is a reasonable basis for the particular conclusion. The strongest practical benefit may be exposing a questionable assumption before it becomes an invisible rule in an automated process.
People also ask
Does circumscription mean minimizing every property?
No. The theory specifies which predicates or formulas are minimized and how others are treated.
Can it retract a conclusion?
Yes. Adding information can change the preferred models when circumscription is applied again.
Is there always one preferred model?
No. Several minimal models may remain; cautious reasoning considers what they share.
Check your understanding
Why must D become abnormal after non-usability is added in the example?
Answer: The fixed device fact and the normal-device rule would otherwise imply usability. Marking D abnormal permits the updated theory to remain consistent without that conclusion.
Related terms
- Closed-world assumption
- Knowledge representation
- Frame problem
- Nonmonotonic reasoning
- default reasoning
- minimal model
- abnormality predicate
Learn next
Learn the closed-world assumption next and compare its treatment of absent facts with selective predicate minimization. Then examine a theory with two incomparable exception sets to see why minimal does not necessarily mean unique.
More in Graphs and symbolic AI
- Choreography - Choreography is a description or coordination pattern for how multiple participants interact through messages and agreed behavior.
- Chronological backtracking - Chronological backtracking is a search method that responds to a dead end by undoing the most recent decision that still has an untried alternative.
- Circuit verification - Circuit verification checks whether a circuit design satisfies a stated specification.
- Classical planning - Classical planning is the task of finding actions that transform a known initial state into a state satisfying a goal, under a simplified model with...
- Clause - A clause, in the logic used by many symbolic AI systems, is a disjunction of literals: statements joined by logical OR, where each statement is an atom or...
- Closed-loop - Closed-loop describes a system in which information about the system's behavior feeds back into decisions that influence its subsequent behavior.
Sources
Primary sources: John McCarthy, Circumscription overview and historical qualification, https://www-formal.stanford.edu/jmc/circumscription.html ; Applications of Circumscription to Formalizing Common Sense Knowledge, https://www-formal.stanford.edu/jmc/applications.pdf ; original paper contents, https://www-formal.stanford.edu/jmc/applications/applications.html . The device example and fictional case are original illustrations.
Previous: Circular convolution Next: Citation precision