Default Reasoning

The defining feature of default reasoning is its non-monotonic nature. In classical logic, adding new premises can only increase the set of provable conclusions (monotonicity). In default reasoning, a conclusion drawn under incomplete information can be retracted when new, contradictory facts are learned. For example, the default 'Birds typically fly' allows inferring that Tweety flies. If you later learn Tweety is a penguin, the conclusion is withdrawn, demonstrating the non-monotonic update of beliefs.

Defaults encode rules about typical members of a category, explicitly acknowledging the existence of exceptions. The structure is: 'If X is true, and it is consistent to believe Y, then conclude Y.' The 'consistent to believe' clause is the exception mechanism. Key components are:

• Prerequisite: The fact that must be true (e.g., 'Tweety is a bird').
• Justification: The condition that must not be proven false (e.g., 'It is consistent that Tweety can fly').
• Consequent: The default conclusion (e.g., 'Therefore, Tweety flies'). This allows reasoning efficiently about the ordinary case without enumerating all possible exceptions.

The most influential formalization is Reiter's Default Logic (1980). It extends first-order logic with default rules of the form: (Prerequisite : Justification) / Consequent. A default theory is a pair (W, D), where W is a set of ordinary logical sentences (facts) and D is a set of default rules. An extension is a maximally consistent set of beliefs that can be generated by applying the defaults in D to W. A key complexity is that a default theory can have zero, one, or multiple extensions, leading to different possible sets of reasonable beliefs.

Determining the consequences of a default theory is computationally challenging. The existence of multiple extensions necessitates defining inference semantics:

• Skeptical Inference: A conclusion is accepted only if it is contained in every extension of the theory. This is a conservative, risk-averse stance.
• Credulous Inference: A conclusion is accepted if it is contained in at least one extension. This is a more adventurous, choice-requiring stance. For example, in the famous 'Nixon Diamond' (Quaker Republicans are typically pacifists; Republicans are typically non-pacifists), skeptical reasoners conclude nothing about Nixon's pacifism, while credulous reasoners choose one consistent view.

A foundational application is formalizing the Closed-World Assumption (CWA), a database principle where facts not explicitly known to be true are assumed false. In default logic, this is encoded as a set of defaults: For every atomic fact P, include the rule : ¬P / ¬P. This reads: 'If it is consistent to assume not P, then conclude not P.' This provides a rigorous, non-monotonic basis for negation-as-failure in logic programming (e.g., Prolog), allowing systems to make assumptions about the completeness of their knowledge.

Default reasoning is a distinct mode of inference within the family of non-monotonic logics.

• vs. Abductive Reasoning: Abduction seeks the best explanation for observations (inference to the best explanation). Default reasoning applies general rules to specific cases in the absence of contradiction. They are complementary: defaults can provide the background knowledge used to generate abductive hypotheses.
• vs. Inductive Reasoning: Induction generalizes from specific examples to general rules. Default reasoning starts with general default rules and applies them deductively, albeit with a non-monotonic proviso. Induction could be used to learn the default rules themselves from data.

Non-monotonic reasoning is a family of formal logics where conclusions are provisionally drawn and can be retracted when new, contradictory evidence arrives. This contrasts with classical monotonic logic, where adding premises never invalidates previous conclusions.

• Core Property: The set of derivable conclusions does not monotonically increase with new information.
• Key Use: Modeling commonsense reasoning, where agents operate with incomplete knowledge.
• Formalisms: Includes default logic, autoepistemic logic, and circumscription.

Default reasoning is the most widely studied and implemented type of non-monotonic logic.

The Closed-World Assumption (CWA) is a rule of inference stating that any statement not explicitly known to be true is assumed to be false. It is a simplifying assumption for reasoning with complete knowledge bases.

• Mechanism: If a proposition P cannot be proven from the knowledge base, then not P is inferred.
• Contrast with Defaults: CWA is a stronger, categorical assumption. Default reasoning uses weaker, defeasible rules (e.g., 'typically, birds fly').
• Application: Foundational to database query answering and logic programming (e.g., Prolog).

While both handle missing information, CWA leads to categorical falsehoods, whereas default reasoning yields tentative conclusions.

Defeasible reasoning is inference that leads to tentative conclusions that hold in typical circumstances but may be overridden by more specific information. It is the general philosophical and cognitive category that includes default reasoning.

• Key Characteristic: Conclusions are defeasible—they can be defeated without logical contradiction.
• Structure: Often uses rules with exceptions (e.g., 'Birds fly, unless they are penguins').
• Formalization: Default logic provides one precise, computable formalism for defeasible reasoning.

All default reasoning is defeasible, but not all defeasible reasoning is captured by the specific rules of formal default logic.

Autoepistemic logic is a non-monotonic logic designed to model an agent's reasoning about its own knowledge and beliefs. It introduces a modal operator L (for 'believed' or 'known') to represent introspection.

• Core Idea: An agent can reason with statements like 'If I do not believe P, then conclude Q.'
• Difference from Defaults: Focuses on belief introspection rather than typicality. A default rule like 'Birds typically fly' is represented differently, as 'If x is a bird and it is consistent to believe that x flies, then believe x flies.'
• Application: Foundational for modeling self-aware or reflective agents in knowledge representation.

Circumscription is a non-monotonic logic formalism developed by John McCarthy that minimizes the extension of predicates to assume as few objects as possible have a given property.

• Mechanism: It's a model preference technique. Among all possible models of a theory, circumscription selects those where the set of objects satisfying certain predicates (e.g., 'Abnormal') is minimal.
• Example: To formalize 'Birds typically fly,' one introduces an Abnormal predicate and writes Bird(x) ∧ ¬Abnormal₁(x) → Flies(x). Circumscription then minimizes Abnormal objects.
• Comparison: Provides a different semantic approach to default reasoning than Reiter's default logic, often yielding similar intuitive results.

Belief revision is the process of rationally integrating new information into an existing set of beliefs (a knowledge base) when the new information conflicts with the old. It governs how to retract old beliefs to maintain consistency.

• Relation to Defaults: When a default conclusion is defeated by new evidence, a belief revision operation must occur. The AGM postulates (Alchourrón, Gärdenfors, Makinson) provide the standard axioms for this process.
• Key Difference: Default reasoning focuses on jumping to conclusions from incomplete data. Belief revision focuses on managing the fallout when those jumps are proven wrong.
• System Design: Robust agentic systems require both a default inference engine and a belief revision policy.