Circumscription—A Form of Nonmonotonic Reasoning

John McCarthy proposes circumscription as a formal rule of conjecture supporting Common-Sense Reasoning. The qualification problem motivates it: specifying every condition for an action’s success can require an impractical list, with further exceptions always possible.^[1]

Approach

Predicate circumscription minimizes a selected predicate’s extension subject to the stated facts. A planner can conjecture that no additional obstacle prevents using a boat. Unlike ordinary monotonic deduction, this conclusion may be withdrawn as information grows. The paper gives logical schemas and illustrative examples, rather than an empirical safety evaluation.^[1]

The river-crossing puzzle illustrates the distinction: adding a bridge changes the appropriate representation. A puzzle convention excludes unmentioned alternatives; outside puzzles, that exclusion remains a hypothesis.^[1]

Reliability and limits

McCarthy says a reasoning program needs domain-specific heuristics for choosing circumscriptions and retracting them. Different predicates representing the same domain can produce different results. The formalism therefore leaves consequential choices about representation and revision to the system designer.^[1]

For Outer Alignment, the editorial connection is that task descriptions depend on background assumptions about actions and obstacles. Revisable beliefs concern what an agent takes to be true; they do not by themselves specify whose objectives it should pursue. No direct influence on modern alignment methods is established here.

Historical context

Published in Artificial Intelligence 13 (1980), pages 27–39. The reviewed author-hosted reprint has a 1986 header and a later addendum. The author’s archive says the distinct 1986 Applications of Circumscription to Formalizing Common Sense Knowledge substantially supersedes this formalism, particularly its treatment of variables not minimized. These dates describe different versions, not a 1986 first publication.^[1]