In classical epistemic logic, agents are often depicted as "logically omniscient" deities. If an agent knows the axioms of arithmetic, they are assumed to instantly know every possible theorem that could ever be derived from them—including complex proofs like Fermat’s Last Theorem. This assumption, baked into the "possible-worlds" semantics pioneered by Robert C. Moore in Reasoning About Knowledge and Action, treats knowledge as a static set of states. If a fact is true in all worlds an agent considers possible, the agent "knows" it.
However, real-world agents—whether humans or computers—have finite memory, limited time, and restricted processing power. They do not instantly see every logical consequence of their beliefs. Research in this field seeks to break the "omniscience" trap by formalizing how agents reason under constraints.
The Problem of Logical Omniscience
The core issue with standard models is referential opacity and logical closure. As Joseph Halpern and Yoram Moses explain in A Guide to the Modal Logics of Knowledge and Belief, possible-worlds semantics imply that if an agent knows $P$, and $P$ logically implies $Q$, the agent must also know $Q$.
This creates a paradox for security and mathematics. In An Epistemic Foundation for Authentication Logics, Joseph Halpern and Riccardo Pucella point out that an omniscient agent could theoretically decrypt any ciphertext without a key, simply because the plaintext is a "logical consequence" of the encrypted string and the algorithm. To build realistic models, we must decouple "what is true" from "what the agent has the power to compute."
Mechanism 1: Impossible Worlds
One of the most robust ways to model imperfect reasoning is to expand the state space to include impossible worlds. These are states where the laws of logic are suspended—where $P$ and $\neg P$ might both be true, or where a mathematical identity fails to hold.
By allowing an agent to consider these "logically impossible" configurations, researchers like Joseph Halpern and Rafael Pass in Decision Theory with Resource-Bounded Agents can represent an agent who is uncertain about a mathematical truth. Until the agent performs the necessary computation to "rule out" the impossible worlds, they remain, from the agent's perspective, possible. This effectively treats "thinking" as a process of pruning away impossible states.
Mechanism 2: Awareness and Two-Dimensional Logic
Another approach focuses on awareness. An agent might possess the information required to reach a conclusion but fail to do so because they aren't "aware" of the relevant concepts.
Shier Ju proposes a two-dimensional logic in Belief, Awareness, and Two-Dimensional Logic, where every statement has both a truth value and an awareness value. Similarly, Joseph Halpern and Leandro Chaves Rêgo explore "unknown unknowns" in Reasoning about Knowledge of Unawareness, modeling agents who don't even have the vocabulary to conceive of certain facts. This "awareness filter" ensures that even if a fact is logically implied, it doesn't enter the agent's belief set unless they are actively attending to it.
Mechanism 3: Belief Bases and Syntactic Knowledge
While "semantic" models look at sets of worlds, "syntactic" models look at sets of formulas. In In Praise of Belief Bases, Emiliano Lorini argues for belief bases—explicit lists of facts stored in an agent's memory.
The intuition here is simple: an agent "knows" what is written in its mental notebook. While the agent might implicitly know the consequences of those notes, they only explicitly know the formulas themselves. This distinction allows for the modeling of agents with finite memory who must perform explicit "deduction steps" to move a fact from implicit consequence to explicit belief. Natasha Alechina and Brian Logan utilize this in Verifying Existence of Resource-Bounded Coalition Uniform Strategies to prove that groups of agents can achieve goals even when their reasoning power is strictly capped.
Mechanism 4: Computation as a Costly Action
Perhaps the most radical shift is treating "thinking" not as a logical property, but as an economic action. In I Don't Want to Think About it Now: Decision Theory With Costly Computation, Joseph Halpern and Rafael Pass suggest that agents choose algorithms (Turing machines) rather than just actions.
In this framework, an agent might choose not to calculate the 100th digit of Pi because the "cost" of the computation (time, energy) outweighs the benefit of knowing that digit. This explains seemingly "irrational" human behavior as a rational response to resource constraints. As Lawrence E. Blume and Joseph Y. Halpern argue in Constructive Decision Theory, our very "states of the world" are subjectively constructed based on the language we use and the computational effort we are willing to expend.
Why It Works: From Ideal to Actual
By moving away from logical omniscience, these researchers have transformed epistemic logic from a branch of idealized philosophy into a tool for engineering. Whether through impossible worlds that allow for mathematical error, awareness functions that model limited attention, or belief bases that track finite memory, the goal is the same: to create a formal language that describes how agents actually think, rather than how they ought to think if they had infinite power.
This shift allows AI systems to reason about their own ignorance (Gerhard Lakemeyer, A Model of Decidable Introspective Reasoning with Quantifying-In), plan to persuade humans (Jorge Luis Fernandez Davila et al., A Simple Framework for Cognitive Planning), and maintain commitments in a messy, unpredictable world (Philip R. Cohen and Hector J. Levesque, Intention = Choice + Commitment).