The frame problem: why describing what does not change is hard for AI
The frame problem is the difficulty of representing in logic all the things an action does not change, without listing each non-effect explicitly. Classical logic requires many frame axioms and fails to capture common sense inertia. AI researchers address it with nonmonotonic logics and default reasoning, but wider epistemological questions remain.
What is the frame problem
The frame problem originated as a technical issue in logic-based artificial intelligence, according to the Stanford Encyclopedia of Philosophy's frame problem entry. In a logical formalism such as the situation calculus, an action is described by formulae that state what holds after the action. For example, painting an object changes its color, and moving it changes its position.
The difficulty appears when those formulae do not rule out unintended consequences. A set of formulae about moving an object does not by itself imply that the object's color remains unchanged. To get the expected conclusion, one must add explicit frame axioms stating that moving does not affect color, and painting does not affect position.
In a domain with many actions and many properties, the number of such frame axioms grows almost as the product of the two counts. This burden is unwelcome, because most actions do not affect most properties. The challenge is to capture these non-effects more succinctly.
Why classical logic is monotonic
Classical logic is monotonic: the set of conclusions derivable from a set of formulae never shrinks when new formulae are added. The Stanford Encyclopedia of Philosophy's logic-AI entry explains that this makes it impossible to express a default rule with an open-ended set of exceptions, such as the common sense law of inertia.
The common sense law of inertia says that an action can be assumed not to change a property unless there is evidence to the contrary. Because monotonic logic cannot retract a conclusion when exceptions appear, it forces the reasoner to list every exception explicitly. That is what drives the explosion of frame axioms.
Automated reasoning programs, as described in the Stanford Encyclopedia of Philosophy's automated reasoning entry, work by proving conclusions from given assumptions. If the assumptions omit a non-effect, the program has no way to infer it. The frame problem is therefore a direct obstacle for such systems.
Nonmonotonic solutions and the role of knowledge representation
AI researchers developed nonmonotonic logics that allow conclusions to be drawn by default and later withdrawn in the face of counterevidence. The MIT CSAIL knowledge representation paper describes knowledge representation as a surrogate, a set of ontological commitments, and a medium for computation, among other roles. These roles shape how a representation handles non-effects.
The nonmonotonic approach formalizes the default assumption that properties persist unless an action is known to affect them. This reduces the need for frame axioms, but it also complicates the logic. The technical frame problem is largely solved in this way, as noted in the frame problem entry, though the wider epistemological questions remain.
A knowledge representation must state what stays the same as well as what changes. The SEP's computational mind entry discusses how mental processes can be seen as computation over representations, and the frame problem shows that even simple reasoning about change relies on a way to represent persistence.
The wider frame problem
Philosophers have extended the frame problem to broader epistemological issues, such as how a reasoner can limit its attention to what is relevant without considering all that is not. The Stanford Encyclopedia of Philosophy's frame problem entry connects this to the question of whether it is possible in principle to bound the scope of reasoning about the consequences of an action.
The AIMA textbook site lists a chapter on logical agents and a chapter on automated planning in its table of contents. Those chapters treat the frame problem as a central issue in reasoning about action and change, illustrating that the problem is not merely historical but continues to inform modern AI curricula.
The tension between the technical solution and the philosophical implications has shaped debates about common sense and rationality. While the narrow problem has been addressed, the wider question of why a reasoner can ignore most non-effects without explicit instruction remains open.
| Feature | Classical logic | Nonmonotonic logic |
|---|---|---|
| Monotonicity | Conclusions only increase | Conclusions can be retracted |
| Frame axioms | Required for each non-effect | Default persistence reduces them |
| Common sense inertia | Cannot express open-ended exceptions | Formalized as default rule |
| Handling exceptions | Must list all exceptions explicitly | Exceptions defeat defaults |
Frequently asked questions
What is the frame problem in simple terms?
It is the difficulty of telling a reasoning system all the things an action does not change, without having to list each one explicitly.
Why does classical logic make the frame problem hard?
Classical logic is monotonic, so any conclusion must be supported by axioms; without an explicit axiom, a non-effect is not derivable.
How do nonmonotonic logics help?
They allow a default assumption that properties persist unless contradicted, reducing the need for many frame axioms.
Is the frame problem solved?
The narrow technical problem is largely solved by nonmonotonic reasoning, but the wider epistemological questions remain open.
Why is the frame problem important for AI?
It affects any system that reasons about actions and change, such as planners and logical agents, and it highlights the need for common sense knowledge.