Is preferential entailment probabilistic prediction?

Not necessarily. A preference order need not assign probabilities or calibrated confidence values.

Does it permit arbitrary contradiction?

No. Conclusions are determined by the selected models and the chosen formal semantics.

Is rational closure another name for all preferential entailment?

No. Rational closure is a particular construction related to stronger rational reasoning properties.

AI Glossary / Data engineering and architecture / Advanced

Preferential entailment

Definition

Preferential entailment is a way of drawing conclusions by considering the preferred, or most normal, possibilities compatible with the current information. A conclusion is accepted when it holds throughout those selected possibilities, allowing default conclusions to change when additional facts reveal an exception.

What it means

Classical entailment asks whether a conclusion holds in every model satisfying the premises. Preferential reasoning adds a comparison between possibilities and focuses on those that are minimal according to the chosen preference relation.

Preference here concerns normality or plausibility in a formal model, not necessarily a person's favorite outcome. The ordering is part of the reasoning system and must be justified for the intended use. It does not automatically supply numerical probabilities.

The terminology has related technical uses. One can describe consequences inside a particular preferential model, or ask which conditional assertions follow across all models satisfying a conditional knowledge base. Keeping those levels separate prevents a conclusion from one chosen ranking being mistaken for a consequence forced by every acceptable model.

How it works

The model represents possible states and specifies which are preferred to others. Given premises, it considers the states satisfying them and selects their minimal states: those with no strictly preferred eligible state below them. The conclusion must hold in all selected states.

Adding information can remove previously preferred states from consideration. Different states then become minimal, and an earlier default conclusion may no longer follow. This is the relevant nonmonotonic behavior.

In the Kraus - Lehmann - Magidor framework, preferential models and inference rules are connected through representation results. Technical conditions, including smoothness in the standard presentation, support the intended reasoning properties. A casual collection of preference scores is not automatically an implementation of that formal system.

Ranked models impose additional structure and are associated with rational consequence relations. Preferential reasoning in general should not be treated as identical to rational closure, which is a more specific construction.

Three practical examples

A rule-based assistant provisionally expects a routine delivery process, while an explicit exception changes which possibilities it treats as normal.
A knowledge-representation researcher checks whether a proposed default follows from a conditional knowledge base under preferential semantics.
A diagnostic reasoning model compares explanations compatible with observations, using an explicitly defined normality relation rather than treating every logically possible explanation equally.

Think of it this way

Think of planning from the ordinary operating timetable until a confirmed special arrangement rules it out. The new fact changes which plan is applicable, so retaining the earlier expectation would be unreasonable.

The analogy illustrates revisable conclusions. It does not authorize ignoring inconvenient facts or choosing whichever possibility is preferred emotionally. Formal preferential reasoning needs an explicit model and evaluates every selected minimal possibility, not just one attractive example.

Worked example

Revising a delivery default

Consider an invented finite model with three states for a dispatched parcel. State A is a standard parcel arriving next day. State B is an oversized parcel arriving later. State C is an oversized parcel arriving next day. All three satisfy “dispatched.”

The model ranks A as most normal, B next and C last. From “dispatched” alone, A is the unique preferred eligible state, so next-day arrival follows defeasibly in this model.

Now add the fact “oversized.” A no longer satisfies the premises. B is preferred among the remaining states, so next-day arrival no longer follows; in this toy model, later arrival does.

The ranking was stipulated for teaching. It is not evidence about real delivery services and is not claimed to be forced by a separate knowledge base. Changing the preference relation can change the conclusions, which is exactly why that relation must be visible.

Illustrative case study

A fictional support-system team encodes ordinary equipment behavior as defaults. Initially, the assistant presents an expected status as certain because it finds a normal explanation compatible with the first report.

The team revises the design so that explanations identify which statements are observations and which depend on defaults. When a new exception fact arrives, the reasoning step is repeated using the updated premises. Tests include cases with several equally preferred possibilities that disagree on the proposed conclusion.

The lesson is that a default is useful only when its basis and revision conditions remain clear. Finding one preferred-looking explanation is insufficient if another selected possibility contradicts the answer.

Common misconceptions and mistakes

Preferential means choosing the outcome a user wants. The relation represents formally specified plausibility or normality, not wishful thinking.
One preferred possibility supporting a claim is enough. The claim must hold in all selected minimal possibilities for the stated consequence relation.
More facts can never remove conclusions. Defeasible conclusions may change when the eligible preferred possibilities change.

Business use

Preferential entailment matters most in knowledge representation, formal AI and systems that reason explicitly with defaults and exceptions. Its practical value is making a particular style of revisable reasoning precise.

A business application needs more than a normal-looking rule. Identify the evidence supporting the default, how exceptions are represented and what happens when preferred possibilities disagree. Separate mandatory requirements from defeasible expectations; a default should not silently weaken a rule that must always be enforced.

For language-model applications, fluent explanations of what usually happens do not establish that the model implements preferential logic. If a product claims formal default reasoning, its represented states, preference assumptions and inference behavior need a testable account.

Check your understanding

Why does the next-day conclusion disappear after “oversized” is added in the worked example?

Answer: The previously preferred state A no longer satisfies the premises. State B becomes the preferred eligible state, and it describes later arrival.

Related terms

Default logic
Knowledge representation
Nonmonotonic reasoning
defeasible reasoning
model theory
conditional knowledge base
rational closure
plausibility

Learn next

Read classical entailment and default reasoning. Construct a finite set of possibilities with an explicit preference order, then check how the minimal eligible states change when an exception fact is added.

Sources

Kraus, Lehmann and Magidor, Nonmonotonic Reasoning, Preferential Models and Cumulative Logics: https://u.cs.biu.ac.il/~sarit/data/articles/0202021.pdf. Primary paper; relevant minimal-state definitions and preferential-reasoning passages reviewed.

Lehmann and Magidor, What does a conditional knowledge base entail?: https://arxiv.org/abs/cs/0202022. Primary research distinguishing rational relations and rational closure.

Moinard, Linking Makinson and Kraus-Lehmann-Magidor preferential entailments: https://arxiv.org/abs/cs/0208017. Primary research on related formulations. The parcel model and support case are original illustrations.

Previous: Preferential Bayesian optimization Next: Prefilter (anti-aliasing)