When you turn the key in your car, you expect the engine to start. In a standard logical model, you might define a rule: If the key is turned and the battery is charged, the car starts. But what if there is a potato in the tailpipe? What if the starter motor just fused? What if a solar flare has disrupted the ignition electronics?
In the real world, the list of potential reasons an action might fail is effectively infinite. This is the Qualification Problem. If an AI agent had to verify every possible precondition before acting, it would be paralyzed by "intractability," a challenge highlighted by Steve A. Chien in Using and Refining Simplifications: Explanation-Based Learning of Plans in Intractable Domains. To function, agents must be able to assume success by default while remaining "elaboration tolerant"—able to incorporate new failure modes without rewriting their entire knowledge base.
The Logic of Jumping to Conclusions
The foundation for solving this problem was laid by John McCarthy 0001 in Epistemological Problems of Artificial Intelligence. McCarthy argued that traditional monotonic logic—where adding new information never invalidates old conclusions—is insufficient for common sense. He introduced circumscription, a form of non-monotonic reasoning that allows an agent to "jump to conclusions" by assuming that only the explicitly mentioned obstacles exist.
Vladimir Lifschitz refined this in Formal Theories of Action (Preliminary Report) by introducing explicit predicates for causality and preconditions. By minimizing these predicates through circumscription, Lifschitz showed that an agent could reason about the "Yale Shooting Problem" and other classic paradoxes without needing to exhaustively list every "what-if" scenario, a goal also shared by Matthew L. Ginsberg and David E. Smith 0001 in Possible Worlds and the Qualification Problem.
Defeasibility and Normalcy
If we cannot list all failure modes, how do we represent "normal" behavior? Chitta Baral and Jorge Lobo addressed this by introducing defeasibility in Defeasible Specifications in Action Theories. They proposed that actions should have effects that "normally" happen unless overridden by specific exceptions.
This concept extends to how we communicate. Nancy L. Green explored this in Normal State Implicature, noting that when a human asks for a "car," they implicitly assume the car is functional (has an engine, inflated tires, etc.). Encoding these "normal state" assumptions is a massive knowledge engineering burden, but it is essential for plan recognition and natural language understanding.
Deriving Qualifications from Causality
A significant breakthrough occurred when researchers realized that many "qualifications" (reasons an action can't happen) are actually hidden "ramifications" (indirect effects of an action).
In Embracing Causality in Specifying the Indirect Effects of Actions, Fangzhen Lin demonstrated that if an action's indirect effects would violate a fundamental law of the domain, the action itself must be impossible. For example, if "walking through a door" would result in "being in two places at once" (a violation of domain constraints), the system can infer that the door must be locked. Norman McCain and Hudson Turner unified this in A Causal Theory of Ramifications and Qualifications, using causal laws to automatically derive the conditions that prevent actions from occurring.
From Logic to the Real World
As the field matured, the focus shifted from pure logic to practical application in robotics and uncertainty. Mehul Bhatt applied these ideas to cognitive robotics in Narrative based Postdictive Reasoning for Cognitive Robotics, using "postdictive reasoning" to help robots explain why an action failed (e.g., "the door didn't open, so someone must have bolted it").
To handle the inherent messiness of the world, Fahiem Bacchus and Daphne Koller moved beyond binary logic in Forming Beliefs about a Changing World. By combining the situation calculus with probabilistic "random-worlds," they allowed agents to reason about the likelihood of various qualifications, such as mechanical faults or external interference.
Summary of the Evolution
The research direction across these papers shows a clear progression:
- Identification: Recognizing that exhaustive lists of preconditions are impossible (McCarthy).
- Formalization: Using non-monotonic tools like circumscription to assume success by default (Lifschitz, Ginsberg).
- Unification: Linking qualifications to causal laws and indirect effects (Lin, McCain).
- Application: Integrating these theories into multi-agent systems (Tran Cao Son et al. in Exploring the KD45 Property of a Kripke Model After the Execution of an Action Sequence) and robotics to handle real-world failures.
By treating "failure" not as a missing rule, but as an exceptional state that can be inferred or observed, these researchers enabled AI to act in an open, unpredictable world.