Does model preference mean selecting the best predictive model?

Not in this entry. It refers to ordering logical interpretations for nonmonotonic reasoning.

Must there be one preferred model?

No. Several models can remain tied or incomparable, and skeptical inference uses what all preferred models share.

Can a conclusion disappear after adding information?

Yes. New facts can change the preferred models and withdraw a conclusion that previously depended on defaults.

AI Glossary / Statistics and probability / Advanced

Model preference

Definition

Model preference is a way of ranking or partially ordering the interpretations that satisfy a logical knowledge base so that conclusions can be drawn from the preferred interpretations. In the knowledge-representation sense used here, a model is a possible world that makes the stated facts and rules true. Preference adds a criterion beyond mere consistency, such as minimizing unsupported exceptions. A conclusion may be accepted when it holds in every preferred model. This meaning differs from choosing a machine-learning algorithm by validation accuracy or measuring a person's preference between generated answers.

What it means

Ordinary logical entailment asks what is true in every model of the knowledge base. That can leave many questions unresolved when information is incomplete. A preference relation narrows attention to models considered more normal or otherwise appropriate under an explicit policy. The additional conclusions depend on that policy, not just the facts. They can be withdrawn when new facts change which models are preferred. This provides a formal route to defeasible reasoning: reasoning whose conclusions remain revisable in the light of further information.

How it works

Define the language and knowledge base, then specify which interpretations satisfy its constraints. Add a preference relation, such as preferring a smaller extension of a selected abnormality predicate while holding designated information fixed. Determine the preferred or minimal models under that relation. For skeptical inference, accept a proposed conclusion only if it holds in all those models. Multiple incomparable preferred models may remain. Circumscription is an important example, but its exact behavior depends on what is minimized, fixed or allowed to vary. Existence of suitable minimal models and computational feasibility require attention in general settings.

Three practical examples

Default reasoning: treat a device as operating normally unless facts force an exception under the chosen representation.
Diagnostic explanation: compare consistent explanations using an explicit preference for fewer unsupported abnormalities.
Knowledge-base completion: reason about a selected relation as containing only the instances required by the available facts. These are modeling policies, not universal permissions to treat every unrecorded fact as false.

Think of it this way

Think of several stories that all fit the evidence. A stated preference might favor stories requiring fewer extra exceptions. That can support a tentative conclusion, but two equally preferred stories may still disagree. The appropriate result is then unresolved under skeptical reasoning, rather than an arbitrary choice presented as certainty.

Worked example

Consider two propositions, A and B, meaning that device A or device B is abnormal. The knowledge base states A or B. Its three satisfying assignments are (true,false), (false,true) and (true,true). Prefer assignments whose set of abnormal devices is smaller by strict set inclusion. The first two are minimal and incomparable; the third has an unnecessary additional abnormality relative to either. Both preferred assignments satisfy not(A and B), so the preference-based reasoning accepts that conclusion. Neither A alone nor B alone holds in every preferred assignment, so neither is skeptically concluded. Now add the explicit facts A and B. Only (true,true) satisfies the expanded knowledge base, so the previous conclusion not(A and B) is withdrawn. This original exhaustive example demonstrates nonmonotonicity without contradicting a stated fact. Minimizing by set inclusion is not generally identical to minimizing the numerical count of abnormalities.

Illustrative case study

In a fictional case, Cedar Diagnostics designs a rule-based assistant for equipment maintenance. Its engineers specify which faults may be treated as exceptional and which observations must remain fixed. When two minimal explanations fit the evidence, the assistant reports the unresolved alternatives and requests another measurement. A later observation can overturn a default conclusion while preserving the observation itself. Reviewers test such updates explicitly, because a system that silently keeps an obsolete default could direct technicians toward an inappropriate repair.

Common misconceptions and mistakes

A preferred model is not necessarily the most probable world. A preference relation need not assign probabilities or rank every pair. Minimal does not always mean uniquely smallest, and incomparable minimal models must not be collapsed without justification. Adding facts can remove an earlier defeasible conclusion even though ordinary classical entailment is monotonic. Preference rules cannot repair an inconsistent knowledge base merely by choosing a favorite interpretation when none satisfies it. The meaning of absent information depends on the specified formalism and application, rather than a blanket closed-world assumption.

Business use

Businesses using rule systems should document both the explicit rules and the preferences that generate defaults. Test how conclusions change when exceptions arrive, and retain explanations showing which assumptions supported a decision. Ambiguity among preferred models may justify escalation or further data collection. Distinguish logical consistency from empirical accuracy and operational suitability. This approach can make default assumptions more visible, but building and maintaining the preference structure requires domain expertise. It should not be represented as a substitute for evidence when the decision requires direct verification.

Check your understanding

Two preferred models disagree about A. Can skeptical preference-based reasoning conclude A?

Answer: No. A must hold in every preferred model; disagreement leaves it unestablished under that inference rule.

Related terms

Circumscription
Default logic
Entailment
Closed-world assumption
Knowledge representation
Nonmonotonic reasoning
logical model
abnormality predicate

Learn next

Review ordinary model-theoretic entailment, then compare it with minimal entailment. Practice enumerating small propositional knowledge bases and checking which conclusions survive after new facts are added.

Entailment

Sources

Russell and Norvig, Artificial Intelligence: A Modern Approach, index locating model preference under logic: https://aima.cs.berkeley.edu/aima-index.html John McCarthy, Circumscription: A Form of Nonmonotonic Reasoning, model theory of predicate circumscription: https://www-formal.stanford.edu/jmc/circumscription/node6.html John McCarthy, formalism and examples of circumscription: https://www-formal.stanford.edu/jmc/circumscription/node4.html

Previous: Model poisoning Next: Model Quantization