AI Glossary / Data engineering and architecture / Advanced

Yale shooting problem

Definition

The Yale shooting problem is a classic artificial-intelligence example showing how apparently reasonable default rules about persistence can produce unintended conclusions about change over time. Introduced by Steve Hanks and Drew McDermott, it tests whether a formal representation captures the intended effects of actions.

What it means

We normally expect a fact to remain true unless something changes it. Encoding that expectation as a default seems straightforward, but interacting defaults can admit surprising alternatives.

In the classic scenario, a gun is loaded, time passes, and a shooting action occurs under simplified assumptions about its effect. The intended reasoning carries the loaded state through the waiting period and applies the stated action effect. Some initial formalizations also permit an unintended history in which the loaded state disappears during waiting, preventing the expected conclusion.

How it works

The model contains facts whose truth can change over time, often called fluents, together with action-effect rules and persistence defaults. The defaults express that properties generally continue unless an exception applies.

The difficulty is that minimizing exceptions without sufficiently representing their causal and temporal roles may not distinguish the intended history from an unintended one. One history makes an exception to persistence when an action changes a state. Another shifts an exception earlier, allowing a prerequisite to disappear without the intended cause.

The lesson is to inspect the models or extensions allowed by the formalization, not merely read the rules and assume they match intuition. Later research proposed richer treatments of action, causality, and persistence. The problem does not prove that all logical approaches to temporal reasoning are impossible.

Three practical examples

1. Workflow modeling: A designer checks that an approval remains valid through an unrelated wait, while an explicit withdrawal changes its status.

2. Planning systems: Researchers use small action sequences to test whether a reasoner preserves unaffected facts and applies specified effects at the correct point.

3. AI state tracking: A team tests whether an assistant distinguishes a recorded change from a merely possible, unobserved change when summarizing an event history.

Think of it this way

Think of a theater prop remaining on a table between scenes. If the script later says an actor picks it up, a reasonable stage plan should not explain the action away by inventing an earlier disappearance without a cause.

The analogy illustrates persistence and causal order. Its limit is that the Yale problem concerns precise formal semantics, including which alternative histories a theory permits. Real-world uncertainty can justify considering an unobserved change; the issue here is whether a deliberately simplified model represents its intended assumptions.

Worked example

Consider a nonviolent teaching analog. At time 0, a service request is approved and not yet released. At time 1, the process waits. At time 2, a release action is attempted. The toy model states that releasing an approved request makes it released, and no withdrawal event is included.

The intended history preserves approval during the wait, then marks the request released after the action. A poorly specified default theory might also admit a history where approval stops during the wait, so the release rule's prerequisite fails and the request remains unreleased.

If both histories remain admissible, a reasoning method requiring a conclusion to hold in every admissible history cannot conclude release. The point is not to force an outcome by wishful thinking. It is to encode clearly which changes have causes and how persistence applies. This analog is explanatory, not a full formal reproduction of the original theory.

Illustrative case study

Fictional case: Temporal Orchard builds a rule-based simulator for internal approval workflows. A short test includes approval, an unrelated delay, and a publication step. The team expects the simulator to infer publication, but inspecting the allowed histories reveals that one history silently drops approval during the delay.

Rather than adding the desired conclusion directly, the designers revise their representation of change events and persistence. They also add tests for genuine withdrawal, missing initial approval, and later evidence that changes what can be inferred.

The scenario illustrates model validation. It does not claim that a particular rule patch solves every temporal-reasoning problem or that the original Yale problem is simply a software coding error.

Common misconceptions and mistakes

A common mistake is treating the example as a puzzle about real firearm reliability. Its action effects are stipulated within a simplified logical domain. Adding real-world contingencies changes the modeled problem.

Another mistake is assuming that every possible history is equally intended just because it satisfies a weak set of axioms. The challenge is that the axioms may fail to express the intended causal constraints.

It is also misleading to say that the problem permanently defeated logic-based AI. It exposed shortcomings in particular approaches and motivated further work. Solutions and their assumptions must be evaluated within their formal setting.

Business use

The business relevance is important for systems that represent state across time. Workflow rules, planning models, and event interpretations can sound sensible while admitting conclusions their designers did not intend.

Use small counterexamples to test initial states, unrelated waiting periods, explicit changes, and conflicting evidence. Keep observed events separate from default assumptions. For an AI application, an explanation that sounds causally plausible is not enough: check whether the underlying representation actually rules out unsupported changes and whether uncertainty is being represented intentionally.

Check your understanding

A toy model permits a prerequisite to disappear during an unrelated wait without a specified cause. What should a designer inspect?

Answer: The persistence rules, action effects, and permitted histories. The intended causal constraints may be missing or inadequately represented.

Related terms

Learn next

Model a harmless two-state process with a wait and one state-changing action. List both intended and unintended histories. Then explain which rule excludes each unintended change, distinguishing an explicit effect from a default assumption.

Sources

Sources: Hanks and McDermott, Default Reasoning, Nonmonotonic Logics, and the Frame Problem: https://cdn.aaai.org/AAAI/1986/AAAI86-054.pdf. Original paper read.

Baker and Ginsberg, Temporal Projection and Explanation: https://www.ijcai.org/Proceedings/89-2/Papers/009.pdf. Read for a subsequent formal treatment and explanation of competing histories. The business analogues are original and intentionally simplified.

Previous: XML database Next: Yarowsky algorithm