In traditional artificial intelligence, knowledge is often treated like a database: a collection of facts that an agent "knows" to be true. However, this approach struggles when an agent is uncertain or needs to reason about what it doesn't know. To solve this, researchers developed a framework based on possible-worlds semantics, where knowledge is represented not as a list of truths, but as a set of candidate realities.
Knowledge as a Set of Possibilities
The core intuition, pioneered by Moore in Reasoning About Knowledge and Action, is that an agent’s epistemic state—its state of knowledge—is defined by the collection of all "possible worlds" consistent with its observations.
If an agent is standing in front of a safe and doesn't know the combination, its epistemic state contains thousands of possible worlds, each featuring a different combination. A fact is considered "known" only if it holds true in every world the agent considers possible. As Levesque argues in The Interaction with Incomplete Knowledge Bases: A Formal Treatment, this allows a system to formally distinguish between a fact being false and a fact simply being unknown. In the latter case, the agent maintains some worlds where the fact is true and others where it is false.
Sensing as Uncertainty Reduction
This set-theoretic view transforms how we think about actions. While physical actions (like moving a chair) change the state of the world, sensing actions (like reading a thermometer) change the agent's internal state of knowledge.
Son explains in Reasoning about Sensing Actions and Reactivity that sensing acts as a filter. When an agent performs a sensing action, it observes a property of the actual world and discards all possible worlds in its mental set that do not match that observation. Consequently, the set of possible worlds "shrinks," representing a reduction in uncertainty.
Scherl et al. formalized this within the Situation Calculus in The Frame Problem and Knowledge-Producing Actions. By treating the "accessibility relation"—the mapping of which worlds are reachable or plausible—as a fluent that changes over time, the authors provided a way for agents to reason about how their knowledge will evolve after future sensing. This is crucial for planning: a robot can plan to "look in the drawer" because it knows that action will eliminate the worlds where the key is missing, thereby enabling the subsequent action of "unlocking the door."
Nested Beliefs and Multi-Agent Systems
The power of possible-worlds semantics becomes most apparent in multi-agent environments. To collaborate or compete, an agent must reason about what other agents know. This leads to nested epistemic states: "I know that you don't know the combination."
Appelt demonstrated in A Planner for Reasoning about Knowledge and Action that a robot can use this to plan communicative acts. If a robot knows a human needs a tool but doesn't know where it is, the robot can plan to provide that specific information. This requires modeling the human's epistemic state as a set of worlds from the human's perspective.
More recently, Shekhar et al. applied this to human-robot collaboration in Human-Aware Epistemic Task Planning for Human-Robot Collaboration. By tracking the set of worlds the human partner considers valid, a robot can detect when the human is confused (i.e., when the human's set of possible worlds no longer includes the actual state) and take actions to realign their mental models.
Formalizing the Logic of Knowledge
Mathematically, this direction relies on Epistemic Modal Logic. As detailed by Halpern et al. in A Guide to the Modal Logics of Knowledge and Belief, these logics use operators like $K$ (for "knows") to navigate the accessibility relations between worlds.
While standard logic deals with truth, epistemic logic deals with validity across reachable states. Ma et al. show in W-JS: A Modal Logic of Knowledge that this framework can solve complex puzzles involving "knowing what" (e.g., knowing a specific value) versus "knowing that" (e.g., knowing a fact exists). By treating knowledge as a structural relationship between possible states, these researchers have provided a rigorous foundation for agents that can reason about their own ignorance and the mental lives of others.
Key Foundational Work
Reasoning About Knowledge and Action (1977) — contributed 1 concept elements to this direction.
On the Role of Mathematical Language Concept in the Theory of Intelligent Systems (1975) — contributed 1 concept elements to this direction.