In standard logic, knowledge is additive. If you know that "it is raining," and then you learn that "it is Tuesday," your knowledge of the rain doesn't vanish. This property, called monotonicity, is the bedrock of classical mathematics. However, it fails to capture a fundamental aspect of human intelligence: the ability to reason about our own ignorance.
When you look at a flight board and don't see a flight to Chicago, you don't just "not know" if there is a flight; you conclude there is no flight. You are reasoning from the absence of information. To do this formally, AI researchers developed the logic of Only Knowing, a framework that allows an agent to draw a boundary around its information and treat everything outside that boundary as unknown.
The Problem of the Open World
Standard epistemic logic uses a "knowledge operator" ($K$). To say an agent knows $\alpha$, we write $K\alpha$. But as Hector J. Levesque pointed out in All I Know: An Abridged Report, $K\alpha$ only provides a lower bound. It says the agent knows at least $\alpha$, but it leaves open the possibility that the agent also knows $\beta, \gamma$, and $\delta$.
This "Open World Assumption" makes it impossible to answer questions about ignorance. If a system only knows $K(\text{Bird})$, it cannot logically conclude that it doesn't know $K(\text{Penguin})$. There might be a hidden fact about penguins tucked away in some unexamined corner of the logic. To solve this, researchers introduced the "Only Knowing" operator ($O$). To say an agent only knows $\alpha$ ($O\alpha$) is to say that $\alpha$ is the entirety of its knowledge.
The Two Fences of Knowledge
Intuition for "Only Knowing" can be built by imagining two symmetrical bounds, a concept formalized by Hector J. Levesque and later expanded by Gerhard Lakemeyer in Only-Knowing: Taking It Beyond Autoepistemic Reasoning:
- The Lower Bound ($K$): This is what the agent is committed to. If the agent knows "Birds fly," this fact is inside the fence.
- The Upper Bound: This is the limit of the agent's belief. To "only know" something is to assert that the upper bound and the lower bound meet exactly at the same set of facts.
By collapsing these bounds, the agent gains negative introspection: the ability to look at its own knowledge base, see that a fact is missing, and confidently state, "I do not know that." This mechanism, which Kurt Konolige explored as Circumscriptive Ignorance, allows the absence of information to function as a positive piece of evidence.
Bridging Monotonic and Nonmonotonic Logic
The "Only Knowing" framework serves as a bridge between two worlds. Classical logic is monotonic (adding facts never changes old conclusions), while common-sense reasoning is often nonmonotonic (new information can cancel old assumptions).
As Gerhard Lakemeyer and Hector J. Levesque demonstrated, "Only Knowing" allows nonmonotonic conclusions to emerge from a monotonic framework. If an agent's knowledge is strictly bounded to "Birds fly," and it is told "Tweety is a bird," it can conclude "Tweety flies." If it later learns "Tweety is a penguin," the boundary of its knowledge changes. The old conclusion is not "proven false" in the classical sense; rather, the premise $O(\text{Birds fly})$ is replaced by a new premise $O(\text{Birds fly} \land \text{Tweety is a penguin})$, which no longer supports the flight conclusion.
Scaling to Multiple Agents
The logic becomes significantly more complex when multiple agents are involved. In Reasoning about only Knowing with Many Agents, Joseph Y. Halpern and Gerhard Lakemeyer explored how agents reason about each other's ignorance.
In a multi-agent system, saying "Alice only knows $P$" is a powerful statement for Bob. It allows Bob to conclude that Alice doesn't know $Q$, which might change Bob's own plan. However, this creates a "Planet Kripke" problem: how do you represent a world where everyone's knowledge is perfectly bounded without falling into infinite loops? Guillaume Aucher and Vaishak Belle addressed this in Multi-Agent Only Knowing on Planet Kripke by using "cover modalities" to precisely characterize the set of possible worlds an agent considers, effectively "covering" only the necessary possibilities and no more.
Practical Applications: From Robots to Databases
While it sounds abstract, this research has direct implications for how AI systems interact with the world:
- Robotics: In ESP: A Logic of Only-Knowing, Noisy Sensing and Acting, Alfredo Gabaldon and Gerhard Lakemeyer showed how robots can use only-knowing to handle noisy sensors. A robot needs to know exactly what its sensors have told it—and what they haven't—to avoid making dangerous assumptions about its environment.
- Databases: Systems like EQL-Lite, described by Diego Calvanese and colleagues in EQL-Lite: Effective First-Order Query Processing in Description Logics, use epistemic operators to perform "closed-world" queries over "open-world" data. This allows users to ask, "What does the database know about this employee?" rather than just "What is true about this employee?"
- Planning: Planners like the XII planner by Keith Golden and Oren Etzioni use "Local Closed World" reasoning to determine when they have sensed enough of a specific area to stop looking and start acting.
By formalizing the "Only Knowing" operator, these researchers have given AI a way to respect its own limits. It transforms ignorance from a simple lack of data into a structured tool for reasoning.