🤖 AI

Epistemic Skills: Reasoning about Knowledge and Oblivion

This paper introduces a weighted model-based epistemic logic framework that utilizes an "epistemic skills" metric to dynamically model knowledge acquisition as upskilling and oblivion as downskilling, while also analyzing group knowledge, knowability, forgettability, and the computational complexity of the system.

Original authors: Xiaolong Liang, Yì N. Wáng

Published 2026-04-22

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Xiaolong Liang, Yì N. Wáng

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are playing a massive, complex board game where the goal is to know the truth about the world. In this game, your ability to see the truth depends entirely on your skills.

This paper introduces a new way of thinking about knowledge, not as a static "on/off" switch, but as a dynamic process of leveling up and leveling down. The authors, Xiaolong Liang and Yì N. Wáng, call this "Epistemic Skills."

Here is the breakdown of their ideas using simple analogies:

1. The Core Concept: Knowledge is a Skill Set

In traditional logic, we usually ask, "Does Agent A know the answer?" The answer is just Yes or No.

In this paper, the authors say: "It depends on what tools Agent A has."

The Analogy: Imagine you are trying to identify a bird.
- If you have no skills, you just see a blurry shape. You don't know what it is.
- If you have the skill of "Ornithology 101," you can tell it's a sparrow.
- If you have "Ornithology 101" AND "Binoculars," you can tell it's a female sparrow.
- If you lose the "Binoculars" skill (downskilling), you go back to just seeing a sparrow, not knowing the gender.

The paper builds a mathematical language to track these skills. If your skill set is strong enough to distinguish between two possibilities, you "know" the difference. If your skills are weak, those possibilities blur together, and you are in a state of "oblivion" (forgetting or not knowing).

2. The Four Main Moves (The Update Modalities)

The authors created a language to describe how agents change their knowledge by changing their skills. Think of this like a video game character inventory:

Upskilling (+S): You pick up a new tool.
- Example: You learn to read maps. Suddenly, you know where you are. You gained knowledge.
Downskilling (-S): You drop a tool.
- Example: You lose your map. Now you are lost again. You entered a state of oblivion.
Reskilling (=S): You swap your entire toolkit for a new one.
- Example: You trade your map and compass for a GPS. Your way of knowing the world changes completely.
Learning (≡b): You copy someone else's toolkit.
- Example: You decide to think exactly like your expert friend. You adopt their skills and see the world exactly as they do.

3. Group Knowledge: The Team Effort

The paper also looks at how groups know things.

Mutual Knowledge: Everyone in the room knows the secret.
Common Knowledge: Everyone knows the secret, and everyone knows that everyone else knows it, and so on (like a rumor that everyone is sure everyone else has heard).
Distributed Knowledge: No single person knows the whole truth, but if you put all their skills together in a bucket, the group could figure it out.
- Analogy: Alice knows the password, Bob knows the username. Neither knows the login, but together they have the "Distributed Knowledge" to log in.
Field Knowledge: The group only knows what everyone in the group shares in common. If Alice knows a secret Bob doesn't, the group doesn't "Field Know" it.

4. The Big Questions: "Can I Know?" and "Can I Forget?"

The authors introduce two fascinating concepts:

Knowability: Is it possible for me to learn this if I just get the right skills? (e.g., "If I study hard enough, can I pass the test?")
Forgettability: Is it possible for me to forget this if I lose the right skills? (e.g., "If I stop practicing, will I forget how to play the piano?")

They use special mathematical symbols to ask these questions: "Is there some version of my future self (with different skills) who knows this?"

5. The "De Re" vs. "De Dicto" Distinction

This is a fancy philosophical term, but the paper explains it simply:

De Dicto (Knowing That): "I know that there is a spy in the room." (I know the fact exists, but I don't know who it is).
De Re (Knowing How/Who): "There is a specific person in the room whom I know is a spy." (I have the skill to identify the specific individual).

The paper shows how our "skill-based" logic can perfectly distinguish between these two types of knowing, especially when we talk about how we might gain that knowledge in the future.

6. The Computer Science Part: Is it Hard to Solve?

The authors didn't just write a theory; they checked how hard it is for a computer to solve problems using this logic.

Checking a specific scenario (Model Checking): If you give a computer a specific situation and a specific question, it can solve it very quickly (in "Polynomial time"). It's like checking a map to see if a route exists.
Finding if a scenario is possible (Satisfiability): If you ask, "Is there any possible world where this happens?" it gets much harder.
- Without "Common Knowledge," it's hard but solvable (PSPACE).
- With "Common Knowledge," it becomes very hard (EXPTIME), taking a long time for computers to figure out.

Summary

This paper is like a rulebook for a dynamic knowledge game. It moves away from the idea that knowledge is a fixed state. Instead, it treats knowledge as a fluid resource that you can gain, lose, swap, or copy.

It gives us the tools to mathematically describe:

How learning works (upskilling).
How forgetting works (downskilling).
How groups solve problems together.
Whether a piece of knowledge is even reachable for us, given our current or potential skills.

It bridges the gap between philosophy (how we know things) and computer science (how we calculate if we can know things), offering a fresh, flexible way to model human intelligence and learning.

1. Problem Statement

The paper addresses the limitations of classical epistemic logic in modeling the dynamics of knowledge acquisition and loss (oblivion). Traditional epistemic logics typically treat knowledge as static or rely on syntactic belief revision (e.g., AGM theory) or semantic erasure of propositions. However, these approaches often fail to capture the capacity or skill required to distinguish between possible worlds.

The authors identify a gap in modeling:

Upskilling: The process of acquiring new capabilities to distinguish previously indistinguishable worlds (gaining knowledge).
Downskilling: The process of losing capabilities, leading to the inability to distinguish worlds (descending into oblivion/forgetting).
Group Dynamics: How these individual skill changes affect mutual, common, distributed, and field knowledge.
Knowability vs. Forgettability: The potential for an agent to gain or lose knowledge through skill modification.

2. Methodology

The authors propose a unified logical framework based on Weighted Epistemic Logic with Implicit Skill Sets.

A. Semantic Framework: Weighted Kripke Models

The core semantic structure is a weighted Kripke model $M = (W, E, C, \beta)$ :

$W$ : A set of possible worlds.
$E$ (Edge Function): $E: W \times W \to \wp(S)$ assigns a set of "skills" to each pair of worlds. If an agent's skill set does not contain a skill that distinguishes $w$ and $u$ , the worlds are indistinguishable to that agent.
$C$ (Capability Function): $C: A \to \wp(S)$ assigns a set of skills to each agent.
$\beta$ (Valuation): Maps worlds to true atomic propositions.
Constraints: The model enforces Positivity (if $E(w,u)=S$ , then $w=u$ ) and Symmetry.

Knowledge Definition:
An agent $a$ knows $\phi$ ( $K_a\phi$ ) at world $w$ if $\phi$ holds in all worlds $u$ where the agent's current skill set $C(a)$ is a subset of the edge set $E(w,u)$ (i.e., the agent lacks the skills to distinguish $w$ from $u$ ).

B. Logical Operators

The language extends classical epistemic logic with three categories of modalities:

Group Knowledge Modalities:
- $E_G$ (Mutual), $C_G$ (Common), $D_G$ (Distributed), and $F_G$ (Field) knowledge, defined based on the union or intersection of agents' skill sets.
Update Modalities (Epistemic Actions):
These operators modify the capability function $C$ of an agent $a$ :
- Upskilling $(+S)_a$ : Adds skills $S$ to agent $a$ .
- Downskilling $(-S)_a$ : Removes skills $S$ from agent $a$ .
- Reskilling $(=S)_a$ : Resets agent $a$ 's skills to set $S$ .
- Learning $(\equiv b)_a$ : Agent $a$ adopts the exact skill set of agent $b$ .
Quantifying Modalities:
These quantify over all possible skill updates, defining Knowability and Forgettability:
- $\boxplus_a$ (Arbitrary Upskilling): "There exists an upskilling such that..." (Knowability).
- $\boxminus_a$ (Arbitrary Downskilling): "There exists a downskilling such that..." (Forgettability).
- $\square_a$ (Arbitrary Reskilling): "For all reskillings..."

C. Variants and Extensions

Fuzzy Skills: Skills can be represented as fuzzy sets with membership degrees.
Lattice Structures: Skills can form a lattice, allowing for generalized join (union) and meet (intersection) operations, generalizing the subset relation.
De Re vs. De Dicto: The framework distinguishes between "knowing that" (de dicto) and "knowing how" (de re) by analyzing the scope of quantifiers over skill updates.

3. Key Contributions

Unified Framework for Knowledge Dynamics: The paper introduces a novel semantic approach where knowledge updates are modeled as changes in the agent's capability set rather than changes in the model's accessibility relation or the agent's belief set. This naturally captures "learning" and "forgetting" as upskilling and downskilling.
Formalization of Knowability and Forgettability: By introducing quantifiers over skill updates, the authors formally define what it means for a proposition to be "knowable" (can be known via some upskilling) or "forgettable" (can be forgotten via some downskilling).
Refinement of De Re/De Dicto Distinction: The paper provides a granular analysis of epistemic expressions, distinguishing between:
- De Dicto: Knowing that a skill set exists.
- Implicit De Re: There exists a skill set such that the agent knows.
- Explicit De Re: The agent knows that a specific skill set exists.
Computational Complexity Analysis: The paper provides a comprehensive complexity analysis for model checking and satisfiability across 2,048 possible logic variants (combinations of operators).

4. Results

A. Model Checking Complexity

Logics without Quantifiers: The model checking problem is in P (Polynomial Time). This holds for logics containing individual knowledge, group knowledge, and update modalities.
Logics with Quantifiers: The model checking problem becomes PSPACE-complete. This is proven via a reduction from the Undirected Edge Geography (UEG) game.

B. Satisfiability Complexity

Without Common Knowledge, Updates, or Quantifiers: PSPACE-complete. (Reduction from classical mono-modal logic $KB_1$ ).
With Common Knowledge (but no updates/quantifiers): EXPTIME-complete. (Reduction from $S5^C_2$ and CPDL).
With Updates or Quantifiers: The complexity for satisfiability in logics containing update modalities or quantifiers remains an open problem. The paper notes that while model checking is tractable (P or PSPACE), the satisfiability problem for these richer logics is significantly harder and currently undecidable in the general case (analogous to issues in Arbitrary Public Announcement Logic).

C. Case Studies

The authors demonstrate the framework's utility through:

Solar System Dataset: Modeling planetary attributes (size, distance, moons) as skills to show how different agents (with different attribute access) derive different knowledge.
Sophisticated Model: A complex 5-world model illustrating the interplay between upskilling, downskilling, and the distinction between implicit and explicit de re knowledge.

5. Significance

Theoretical Advancement: The paper bridges the gap between rough set theory (where attributes define indiscernibility) and dynamic epistemic logic. It offers a semantic alternative to syntactic belief revision, treating "forgetting" not as deleting information but as losing the capacity to distinguish.
Practical Applicability: The framework is highly relevant for:
- Data Science: Modeling knowledge in datasets where attributes (skills) can be added or removed.
- AI and Multi-Agent Systems: Modeling agents that learn new capabilities or suffer from "cognitive degradation" (loss of skills).
- Security: Analyzing what information becomes accessible or inaccessible when an agent's clearance (skill set) changes.
Complexity Boundaries: By establishing that model checking remains in P or PSPACE even with complex update mechanisms, the paper suggests that verification of such systems is feasible, even if full satisfiability checking is harder.
Future Directions: The work opens new avenues for studying "deskilling" (simplifying skills to gain knowledge) and axiomatizing the full logic, which remains a significant challenge due to the undecidability of similar quantified announcement logics.

In summary, this paper provides a robust, mathematically rigorous framework for reasoning about how the capacity to know (skills) evolves, offering new tools to analyze learning, forgetting, and the potential for knowledge in multi-agent systems.