Imagine an artificial intelligence with a knowledge base containing exactly one fact: Bird(Tweety). If you ask this system whether Tweety is a bird, it will confidently answer yes. But what happens if you ask, "Is Tweety capable of flight?"
In standard classical or modal logic, the system cannot answer no. It can only state that it does not know. Standard logical operators establish a lower bound on knowledge: asserting that an agent knows $X$ means it knows at least $X$, leaving open the possibility that it secretly knows a thousand other facts.
To build autonomous agents that can reason about common-sense scenarios, plan under uncertainty, or recognize their own limitations, we need a way to establish an upper bound. We need to be able to state that an agent knows $X$, and nothing else. This problem—formalizing the concept of "only knowing" or epistemic closure—is a cornerstone of knowledge representation.
The Lower-Bound Problem
As John McCarthy 0001 argued in his foundational work on the epistemology of artificial intelligence (Epistemological Problems of Artificial Intelligence), an intelligent agent must be able to "jump to conclusions" when information is missing. However, standard monotonic logics are poorly suited for this. If you learn that a bird cannot fly, a monotonic system cannot retract its previous conclusions.
The root of the issue lies in how standard modal logic models knowledge. We typically define an agent's knowledge using a set of "possible worlds." If an agent knows a proposition, that proposition must be true in every world the agent considers possible. If we simply add facts to a knowledge base, we restrict the set of possible worlds.
The challenge, as Moshe Y. Vardi demonstrated in A Model-Theoretic Analysis of Monotonic Knowledge, is that standard frameworks lack a precise mechanism to draw a boundary around a knowledge base. Without this circumscription, unintended higher-order information leaks into the agent's belief state, making it impossible to state that the agent possesses minimal information.
The Mechanics of Epistemic Closure
To solve this, researchers introduced the "only knowing" (or "all I know") operator. The core intuition relies on a beautiful symmetry: to minimize an agent's knowledge, we must maximize its uncertainty.
If an agent only knows a specific knowledge base $\alpha$, it should consider possible every single world where $\alpha$ is true. By opening the floor to every logically consistent alternative, we ensure that the agent remains completely ignorant of any fact not strictly necessitated by $\alpha$. If a proposition $\beta$ does not logically follow from $\alpha$, there will be at least one maximized possible world where $\beta$ is false. Because the agent considers that world possible, it cannot claim to know $\beta$.
Standard Knowledge (Lower Bound):
[Possible Worlds] ──> Restricted to a small subset where α is true (and potentially other facts).
Only Knowing (Exact Boundary):
[Possible Worlds] ──> Maximized to include EVERY world where α holds.
Any fact β not implied by α varies across these worlds, proving ignorance.
Kurt Konolige formalized this in Circumscriptive Ignorance by extending modal logic with a circumscriptive operator. This allowed the absence of a proof to be treated as a positive piece of information, enabling agents to solve classic logic puzzles that require reasoning about what they do not know.
Expanding on this semantic foundation, Craig Boutilier showed in A Logic for Revision and Subjunctive Queries how an "only knowing" operator anchors belief revision. By forcing the most plausible worlds in a model to be exactly those where the knowledge base holds, it prevents external, unstated assumptions from corrupting counterfactual "what if" reasoning.
Achieving Introspection and Decidability
When an agent can formally bound its own knowledge, it gains the capacity for perfect introspection—the ability to look inward and accurately declare its own ignorance.
Gerhard Lakemeyer spent significant effort making these introspective systems computationally viable. In Decidable Reasoning in First-Order Knowledge Bases with Perfect Introspection and A Model of Decidable Introspective Reasoning with Quantifying-In, Lakemeyer proved that an epistemic closure or "only-believing" operator allows first-order knowledge bases to answer queries about their own lack of information without falling into infinite loops or undecidable states.
This boundary-setting mechanism serves as the underlying engine for non-monotonic reasoning. In Only-Knowing: Taking It Beyond Autoepistemic Reasoning, Lakemeyer and Hector J. Levesque demonstrated that the only-knowing operator acts as a grand unifier. It allows complex, non-monotonic frameworks (where conclusions can be revised as new evidence appears) to be cleanly embedded within a standard, monotonic modal logic simply by explicitly bounding the agent's axioms.
Multi-Agent Dynamics and Communication
The problem becomes significantly more intricate when moving from a solitary agent to a multi-agent system. In a shared world, Agent A must reason not only about its own boundaries, but also about what Agent B only knows.
Halpern et al. addressed this challenge by extending the single-agent boundary to multi-agent contexts, creating a framework where the total extent of each individual's knowledge could be isolated. Concurrently, Gerhard Lakemeyer analyzed this in All They Know: A Study in Multi-Agent Autoepistemic Reasoning, demonstrating how agents can draw sharp conclusions about a system by exploiting the explicit absence of information in another agent's belief state.
This capacity to model bounded knowledge shapes how we understand communication and coordination:
- Conversational Shortcuts: In linguistics, traditional models often assume that dialogue requires an infinite chain of mutual knowledge (I know that you know that I know...). However, Nadathur et al. argued in Mutual Beliefs in Conversational Systems: Their Role in Referring Expressions that human speakers rely on a formalization of ignorance. They use referring expressions based on a "conversational conjecture"—assuming a reference is understood simply because they have no explicit reason to doubt it.
- Operational History: In distributed systems, Halpern et al. proved in An Operational Semantics for Knowledge Bases that an operational system can track "only knowing" dynamically. By identifying the maximal set of world states consistent with an agent's exact input history, the system ensures it never assumes certainty beyond what has been explicitly communicated.
- True Comprehension: Bounding knowledge is also vital for resolving ambiguity. Naumov et al. demonstrated in Comprehension and Knowledge that formal knowledge of a proposition across every possible interpretation is a prerequisite for true comprehension, showing how strict epistemic boundaries prevent misalignments in high-stakes communication.
Probability, Game Theory, and Optimization
While the foundations of only knowing are rooted in symbolic logic, the mechanism has expanded into probabilistic reasoning, game theory, and computational optimization.
Probabilistic and Infinite Domains
Real-world environments require handling both ignorance and numerical uncertainty. To bridge this gap, Belle et al. combined only knowing with subjective probability in infinite domains (A First-Order Logic of Probability and Only Knowing in Unbounded Domains). This allows an agent's beliefs and non-beliefs to emerge naturally as logical consequences of a probabilistic knowledge base. Furthermore, when an environment changes, Belle et al. utilized an epistemic closure operator to efficiently progress a knowledge base over time, keeping updates strictly within first-order logic constraints and avoiding second-order complexity.
Game Theory Paradoxes
In game theory, the "all I know" operator has been used to resolve long-standing conceptual paradoxes regarding how rational players treat eliminated strategies. In A Conceptually Well-Founded Characterization of Iterated Admissibility Using an "All I Know" Operator, Halpern et al. applied the operator to mandate that all logically consistent states remain possible. This mechanism forces a full-support probability distribution, ensuring that no valid strategy is erroneously assigned a zero probability.
Computational Efficiency
Representing the exact upper bound of an agent's knowledge is traditionally complex and computationally expensive. To make these systems practical for contingent planning, Niveau et al. introduced Epistemic Splitting Diagrams (ESDs) in Efficient Representations for the Modal Logic S5. ESDs provide a compact data structure explicitly designed to encode the exact extent of an agent's knowledge rather than just a lower bound, making only-knowing operations tractable in complex environments.
Key Foundational Work
Circumscriptive Ignorance (1982) — contributed 6 concept elements to this direction.
Epistemological Problems of Artificial Intelligence (1977) — contributed 1 concept elements to this direction.