Frame Problem (and Qualification + Ramification)
The “three problems of planning” — McCarthy & Hayes (1969) and Reiter (1991). Whenever your AI exam mentions planning, at least one of these three appears. They’re easy to mix up — so memorize the distinction.
The trio at a glance
| Problem | Question it asks | Concrete example |
|---|---|---|
| Frame Problem | What stays the same after an action? | I move block A onto block B. Block C is still on the table. How does the system know C didn’t move? |
| Qualification Problem | What are all the preconditions for the action to succeed? | ”Pick up block A” — but the gripper might be broken, the room might be dark, gravity might be off, etc. You can never list all preconditions. |
| Ramification Problem | What are the indirect effects (side effects) of the action? | I move a box onto a shelf. The items inside the box also move (without being explicitly mentioned). |
One-line memory aids:
- Frame = “what doesn’t change”
- Qualification = “what must be true before”
- Ramification = “what changes as well”
Why each is hard
Frame Problem
In a logic-based planner (e.g. Situation Calculus), you’d naively need a frame axiom for every (action, fluent) pair, asserting that the action does not affect that fluent. With n actions and m fluents, that’s O(n·m) axioms — exponentially impractical in a realistic world.
STRIPS’ “solution” is the closed-world assumption + ADD/DEL lists: anything not mentioned in the ADD or DEL list is assumed unchanged. This eliminates frame axioms syntactically — but the underlying ontological problem (deciding which actions actually change which facts in the real world) is just pushed onto the designer of the action schema.
Qualification Problem
You can never write down all preconditions of a real-world action. “Make coffee” needs: kettle works, water available, coffee available, electricity available, no earthquake, no power outage, no broken cup, … The list is infinite.
Pragmatic “solution”: assume normal conditions (defaults) and use non-monotonic reasoning (default logic, circumscription) to handle exceptions when they arise. Real planners ignore it and hope.
Ramification Problem
Actions have effects you didn’t bother modeling. Tipping a glass of water → water spills → carpet gets wet → mold grows → property value drops. Most can’t be enumerated.
Pragmatic “solution”: causal rules that propagate indirect effects (Reiter 1991, Lin 1995). Or just accept that some effects are missed.
⚠️ Common exam traps
- Don’t confuse Frame with Ramification. Frame = “C stays where it was.” Ramification = “things inside the box move with the box.” They feel similar but address opposite directions: Frame says non-change, Ramification says additional change.
- STRIPS doesn’t solve the Frame Problem — it sidesteps it syntactically. This is a frequent gotcha. The closed-world assumption + ADD/DEL implicitly handles frame axioms, but the real problem (knowing what changes in the world) is still on the modeler.
- Qualification is about preconditions, Frame is about effects. PRE-list vs. (implicit) frame axioms.
Where these problems still matter today
- Classical symbolic planners (PDDL, STRIPS-style): all three still active — modern planners assume closed-world and bounded action effects.
- Robotic task planning: ramification problem is huge (a robot pushing a box affects everything in the box).
- LLM-based planning agents (ReAct, AutoGPT, Devin): the qualification problem is now most visible — these systems regularly fail because some unstated precondition wasn’t met (file doesn’t exist, API rate-limited, network down).
- Self-driving cars: ramification is everywhere — every motion of the car affects the trajectory of everything nearby.
Where they’ve been “solved” pragmatically
| Problem | Modern workaround |
|---|---|
| Frame | PDDL-style action schemas + closed-world assumption. World models in modern RL (Dreamer, MuZero) learn frame implicitly. |
| Qualification | Default reasoning + bounded preconditions in classical planners. LLM agents use re-planning when preconditions fail. |
| Ramification | Causal rules, or — in deep-learning systems — learn the dynamics from data (world models) instead of specifying effects. |
Status: none of these are fully solved at the foundational level. They’re tractably managed in practice — but the original 1969 critique still bites whenever a planner meets the real world.
See also
- Situation Calculus — the FOL framework where the Frame Problem was originally identified
- STRIPS — sidesteps the Frame Problem via CWA + ADD/DEL lists
- PDDL and STRIPS — the algorithmic side of classical planning
- Why an an AI struggles with Planning — modern angle on these problems
- Markov Decision Process (MDP) — probabilistic planning sidesteps frame problem via transition probabilities
- lernzettel_planning-i_30-04-26 — Planning I Lernzettel
- quiz_planning_30-04-26 · Questions for Methods of AI (Q30-39, Q86 cover these directly)
- Methods of AI Lecture
- MOCs: 610 🤖Artificial Intelligence, Künstliche Intelligenz
Tags: methods-of-ai planning frame-problem qualification-problem ramification-problem
Created: 18-05-26