Is this about visual frames in video?

No. Here frame refers to reasoning about change and persistence in a represented world.

Do practical planning languages address it?

Many provide conventions for specifying effects while treating other facts as persistent, within the assumptions of their action model.

Has one method solved every related common-sense problem?

No. Technical approaches address defined representations, while richer domains introduce additional issues such as indirect effects, uncertainty and incomplete action knowledge.

AI Glossary / Graphs and symbolic AI / Intermediate

Frame problem

Definition

The frame problem is the challenge of representing what stays unchanged when an action occurs without explicitly listing every unaffected property for every action. In classical AI, it arises when a reasoning system must infer a new state from an action and a description of the previous state.

What it means

Saying what an action changes is often easy. Moving a box changes its location. But a reasoning system also needs to know that the move does not normally change the box's color, label or the location of an unrelated chair.

Ordinary logic does not automatically infer these non-effects from a missing statement about change. If a model specifies only the move's direct effect, it may leave many facts about the resulting situation undetermined.

The technical frame problem concerns an adequate, manageable representation of persistence. Broader philosophical discussions connect it to relevance and common sense, but those broader questions should not be confused with the narrower action-representation problem.

How it works

One direct approach writes frame axioms stating which properties persist through which actions. This becomes cumbersome when a domain has many actions and properties, especially as new ones are introduced.

Other approaches encode inertia more compactly. Action representations may state the conditions under which a property becomes true or false and otherwise preserve its previous value, subject to explicit assumptions about completeness of the action model. Nonmonotonic formalisms can also represent persistence defaults with exceptions.

The challenge is not solved by blindly retaining every old fact. Direct effects can invalidate old facts, and indirect effects may follow through other rules. The model must distinguish legitimate persistence from consequences that require an update.

Three practical examples

Robot planning: Moving one object should update its position while preserving unrelated modeled properties.
Workflow simulation: Approving a request changes its status while leaving its identifier and recorded submission date intact.
Game-state reasoning: A move changes selected board relationships while most other represented facts remain unchanged.

Think of it this way

Think of editing one cell in a spreadsheet while expecting the other stored cells to keep their values. You should not have to re-enter every unaffected cell after each edit. The analogy has limits: formulas may create indirect changes, just as action models must account for consequences beyond the directly edited property.

Worked example

Consider a small world containing boxes A and B and rooms 1 and 2. Initially, A is in room 1, B is in room 2, and A is red. The action Move(A,2) is intended to place A in room 2.

An effect rule establishes that A is in room 2 after the action. That alone does not establish that B remains in room 2 or that A remains red. Those statements need persistence support in the chosen formalism.

The model also needs to remove or exclude the old claim that A is in room 1. If locations are exclusive, retaining both locations would be wrong. Thus, “keep all facts” is not an adequate solution.

In an explicit add-and-delete representation for this simplified domain, the action adds At(A,2), deletes At(A,1), and leaves other stored facts unchanged. The result contains At(A,2), At(B,2) and Red(A).

Now add the rule that a box moved into a paint booth becomes blue, and suppose room 2 is that booth. Preserving Red(A) without considering that indirect effect may no longer match the intended model. The original persistence policy depended on the assumed action semantics. Extending the domain requires revisiting those assumptions rather than treating unchanged information as universally guaranteed.

Illustrative case study

Fictional case: Cedar Simulation models warehouse movements for a planning demonstration. Its first logical model states where each moved container arrives but omits rules about properties unaffected by movement.

Queries about container labels after a move become unanswerable even though no label-changing action occurred. The team adds a systematic persistence treatment and tests both affected and unaffected properties.

A later extension introduces a washing station that changes a container's cleanliness when it enters. The team updates the action model to represent that consequence instead of letting the old persistence assumption preserve the wrong cleanliness state. The case shows why a compact model of non-effects must remain aligned with the domain's actual modeled effects.

Common misconceptions and mistakes

The frame problem is simply storing the previous state. Storage alone does not specify which facts remain valid after an action.
Unmentioned effects are automatically impossible in classical logic. That requires an additional representational commitment.
Assuming everything persists solves the problem. It conflicts with properties the action actually changes.
It is identical to the qualification problem. Qualifications concern conditions that might prevent an action from having its expected effect; persistence is a distinct issue.

Business use

The frame problem matters when AI systems plan, simulate or reason about changing states. An application can make incorrect decisions if it forgets unaffected information or preserves facts that should have changed.

Review action definitions for direct effects, indirect effects and non-effects. Make the completeness assumptions of the representation explicit. When a workflow or environment changes, inspect whether new consequences invalidate old persistence rules. Useful tests include questions about unrelated properties after an action, not just checks that its intended main effect occurred.

Check your understanding

A rule says that painting box A makes it blue. Does that rule alone establish that box B keeps its previous color?

Answer: No. Persistence of B's color needs support from the model's frame axioms, inertia convention or equivalent representation. The painting effect alone does not state it.

Related terms

Learn next

List five properties in the worked world and mark which Move(A,2) directly changes. Add an indirect effect and explain why the list must be revised. Compare explicit non-effect statements with a compact persistence convention.

Sources

McCarthy and Hayes introduced the frame problem in their foundational discussion of AI reasoning: https://www-formal.stanford.edu/jmc/mcchay69/mcchay69.html

McCarthy explains its representation and elaboration difficulties with action examples: https://www-formal.stanford.edu/jmc/generality.pdf

His discussion of event effects illustrates effect and frame axioms: https://www-formal.stanford.edu/jmc/narrative2/node7.html

The box-world calculation and Cedar Simulation case are original illustrations of explicitly stated modeling assumptions.

Previous: Frame interpolation Next: Framing effect