Is it the same as assigning numerical penalties?

Not generally. Predicate inclusion and priority ordering differ from an arbitrary weighted sum.

Can a higher-priority default have an exception?

Yes, when the hard theory requires one.

Why can a conclusion disappear after adding information?

New facts can eliminate the formerly preferred models.

AI Glossary / Graphs and symbolic AI / Advanced

Prioritized circumscription

Definition

Prioritized circumscription is a form of nonmonotonic reasoning that assigns precedence to predicates being minimized. It prefers models with fewer exceptions to higher-priority defaults before considering lower-priority exceptions, according to a specified ordering. The minimization operates within the models satisfying the stated theory. It does not permit a preferred default to override a hard fact or repair an inconsistent theory by ignoring an axiom.

What it means

Defaults express what normally holds when no relevant exception is established. Two defaults can point toward conflicting conclusions. Ordinary simultaneous minimization may leave several incomparable models, each making an exception to a different default. Priorities express which exception should be avoided first. This makes the chosen default behavior explicit, but the priorities themselves remain part of the model.

How it works

Represent defaults using abnormality or exception predicates. Specify the underlying theory, the predicates to minimize, their priority ordering and which other predicates may vary or must remain fixed. Compare models using the higher-priority minimized predicates first; lower-priority minimization is considered when the relevant higher-priority extensions are equal.
Predicate minimization generally compares extensions by inclusion, not simply by counting true instances. In a small propositional example, each exception flag has only false and true as alternatives, making a lexicographic comparison easy to display. Conclusions supported by every preferred model are skeptical consequences. Adding facts can change which models are preferred and retract earlier default conclusions, which is why the reasoning is nonmonotonic.

Three practical examples

A configuration assistant has a general default favoring automatic updates and a more specific default for demonstration machines. An explicit priority can determine which default yields when both apply.
A scheduling knowledge base represents normal availability and a separate default associated with a special event. It can document precedence between the defaults while retaining confirmed appointments as hard information.
A symbolic reasoning prototype compares different exception policies. Examining its preferred models can reveal whether a priority choice produces the intended conclusion or leaves alternatives unresolved.

Think of it this way

Imagine reviewing two competing working assumptions. You try harder to preserve the first assumption, then preserve the second wherever doing so does not require abandoning the first. Confirmed facts remain constraints on both. This is an ordering of defeasible assumptions, not permission to dismiss evidence that makes a favored assumption inapplicable.

Worked example

Consider one demonstration machine. Let U mean 'automatic updates are enabled.' Let G mean 'there is an exception to the general update default,' and D mean 'there is an exception to the demonstration-machine default.'
Use two implications: not G implies U, and not D implies not U. The first says updates are normally enabled. The second says this demonstration machine normally has updates disabled. Only U, G and D vary. These are the entire toy theory; neither U nor not U is initially asserted as a hard fact.
The exception pair (G, D) = (false, false) is impossible because it would require both U and not U. The pair (false, true) permits only U = true. The pair (true, false) permits only U = false. With (true, true), either value of U is permitted. Thus four of the eight assignments to the three variables satisfy the theory.
Under simultaneous minimization of G and D, the two one-exception models are incomparable: each avoids one exception while introducing the other. Both are preferred over models with both exception flags true. Therefore neither U nor not U holds in every preferred model.
Now give minimizing D higher priority than minimizing G. Models with D false exist, so the preferred model has D = false, G = true and U = false. Updates are disabled because the demonstration default is preserved while the general default receives an exception. Reversing the priority selects D = true, G = false and U = true.
Finally, restore D's higher priority and add U as a hard fact. The second implication now requires D = true. Among the remaining models, minimizing G selects G = false. The result respects U; it does not disable updates by overriding the added fact. The earlier default conclusion not U is withdrawn.

Illustrative case study

In a fictional device-policy prototype, the team expects demonstration machines to avoid automatic updates. A model with two unprioritized defaults instead leaves the update setting unresolved. The team enumerates the four satisfying assignments and sees why either default can take the exception.
It specifies priority for minimizing D and obtains the intended default conclusion. A later test supplies an explicit instruction that updates are enabled. The system must then retain that fact and treat the demonstration default as exceptional. This test distinguishes a defensible default policy from an implementation that blindly executes whichever rule appears first.

Common misconceptions and mistakes

Priority does not mean replacing a logical contradiction with whichever statement is preferred. If hard axioms have no model, circumscription cannot manufacture one merely by minimizing predicates. Defaults must be represented so exceptions can be expressed consistently.
Another mistake is equating predicate inclusion with minimum cardinality. Two different exception sets may be incomparable even if one contains fewer elements. Nor does prioritized circumscription necessarily produce a single preferred model. The toy example becomes unique under its stated priorities, but more complex theories can retain uncertainty. Finally, minimizing the wrong predicates or fixing a predicate that should vary can change the conclusions substantially.

Business use

Prioritized circumscription provides a formal way to examine policies that combine ordinary expectations, exceptions and precedence. Before implementation, identify which statements are hard requirements and which are defeasible defaults. Explain why each priority exists and test conflicting defaults, explicit exceptions and newly added facts. Review conclusions across all preferred models when uncertainty matters.

Check your understanding

Why can both exception flags not initially be false?

Answer: The rules would require contradictory update settings.

What follows when D has higher minimization priority?

Answer: U is false in the preferred model.

What happens when U is added as a hard fact?

Answer: D becomes true, and the default conclusion not U is withdrawn.

Related terms

Circumscription
Default rules
nonmonotonic reasoning
default rule
abnormality predicate
preferred model
predicate minimization
skeptical consequence
lexicographic ordering

Learn next

Add a second machine with separate exception instances. Compare inclusion-minimal exception sets with merely counting exceptions, and check whether all preferred models agree on each machine's update setting.

Sources

John McCarthy, Applications of Circumscription to Formalizing Common Sense Knowledge, section 12: https://www-formal.stanford.edu/jmc/applications.pdf AIMA glossary, prioritized circumscription: https://github.com/aimacode/aima-glossary/blob/master/glossary.md

Previous: Prior (regularizer) Next: Prioritized experience replay