Non-monotonic reasoning
Classical formal logic is monotonic: if a conclusion follows from a set of premises, it continues to follow no matter how many new premises are added. The ...
Revising Beliefs When New Information Arrives
Classical formal logic is monotonic: if a conclusion follows from a set of premises, it continues to follow no matter how many new premises are added. The set of conclusions can only grow. This property is mathematically clean but practically inadequate for representing commonsense knowledge, where new information regularly forces us to revise earlier conclusions.
Non-monotonic reasoning allows conclusions to be retracted when new information contradicts them. A previously valid conclusion may become invalid after learning an additional fact. This mirrors how human reasoning actually works — we draw tentative conclusions based on available information and update them as we learn more.
The canonical example is the Tweety problem. We know that birds fly. We are told Tweety is a bird. By default, we conclude Tweety flies. If we then learn that Tweety is a penguin, we retract this conclusion because penguins are birds that do not fly. Classical logic cannot handle this gracefully because the additional fact about penguins does not contradict the general rule — birds fly — it just reveals an exception.
Several formal systems implement non-monotonic reasoning. Default logic, developed by Raymond Reiter, adds default rules that apply unless their conclusions are explicitly defeated. Circumscription minimizes the extension of predicates — essentially assuming that only the exceptions explicitly stated are exceptions, and everything else follows the rule. Autoepistemic logic reasons about what an agent does and does not know.
Truth maintenance systems (TMS) are practical implementations that track justifications for beliefs and automatically retract conclusions when their justifications are undercut by new information.
Non-monotonic reasoning is essential in legal AI, medical diagnosis, and any domain with incomplete information and exceptions to general rules.
Exam tip: The Tweety problem is the go-to exam example. Know default logic syntax: a default rule has a prerequisite, a justification (what must be consistent), and a consequent.