Formal Models of Knowledge, Belief, & Uncertainty

• Formal models of knowledge, belief, and uncertainty are mathematical frameworks that precisely represent agents’ epistemic states using logic, probability, and order theory.
• They integrate classical modal logics, graded probabilistic approaches, and plausibility models to support multi-agent systems, decision theory, and AI.
• These models emphasize dynamic updates and global coherence in uncertainty management, making them essential for robust knowledge-based protocols.

A formal model of knowledge, belief, and uncertainty is a mathematical specification, typically using logic, probability theory, or order-theoretic structures, that precisely characterizes the representation, reasoning, and dynamics of agents’ epistemic and doxastic states, as well as their corresponding uncertainty management. These frameworks underpin the theoretical foundations of multi-agent systems, decision theory, dynamic epistemics, knowledge-based protocols, probabilistic AI, argumentation, and uncertainty quantification.

1. Classical Epistemic and Doxastic Logics

The foundational models for knowledge and belief employ modal logic, specifically multi-agent Kripke semantics. Let $\text{At}$ denote a set of propositional variables and $\text{Ag}$ the finite set of agents. The language is generated by modal operators $K_i$ (“$i$ knows φ”) and $B_i$ (“$i$ believes φ”):

$$\varphi ::= p \mid \neg\varphi \mid (\varphi\wedge\varphi) \mid K_i\varphi \mid B_i\varphi \quad\text{for } p\in\text{At},\ i\in\text{Ag}$$

A Kripke model is a triple $M=(W,\{R_i\}_{i\in\text{Ag}},V)$ with nonempty $W$ (worlds), accessibility relations $R_i$ encoding agents’ indistinguishability or plausibility, and $V$ assigning true atoms to each world.

• $K_i\varphi$ is true at $w$ iff $\forall v,\ wR_iv\implies M,v\models\varphi$.
• Common knowledge and distributed knowledge use fixed-point and intersection constructions on $R_i$.

Frame classes (e.g., K, T, S4, S5 for knowledge; KD45 for belief) correspond to combinations of reflexivity, transitivity, Euclideaness, and seriality, aligning with philosophical properties:

Frame property	Modal axiom	Epistemic significance
Reflexivity	T: $K_i\varphi\to\varphi$	Knowledge is true
Seriality	D: $\neg B_i\bot$	Belief is consistent
Transitivity	4: $K_i\varphi\to K_iK_i\varphi$	Positive introspection
Euclidean	5: $\neg K_i\varphi\to K_i\neg K_i\varphi$	Negative introspection

Knowledge (S5) is factive and fully introspective; belief (KD45) drops factivity, retaining consistency and introspection (Ditmarsch et al., 2015).

2. Quantitative and Graded Models

Probabilistic Modal Logics

Quantitative uncertainty is captured by equipping worlds (or accessibility classes) with probability distributions $\mu_{i,w}$. Assertions $P_i(\varphi)\ge q$ express that agent $i$’s subjective probability for $\varphi$ at $w$ is at least $q$:

$$M,w\models P_i(\varphi)\ge q \iff \mu_{i,w}(\{v: M,v\models\varphi\})\ge q$$

Sound and complete axiomatizations integrate modal and arithmetic axioms, with graded modalities $B_i^r\varphi$ denoting degree-of-belief thresholds (Ditmarsch et al., 2015, 1304.1508).

Certainty and Probabilistic Belief

Certainty is usually identified with probability one. In single measure models, the logic KD45 characterizes certainty, while S5 results if all worlds have positive measure. Support conditions on probability functions yield additional frame correspondences. Miller’s principle is captured by uniformity conditions on world-dependent probabilities (1304.1508).

3. Qualitative and Plausibility-Based Logics

Source-Agnostic and Plausibility Models

Plausibility models generalize Kripke frames to encode graded or conditional plausibility among worlds. Agents’ beliefs are then formulated using minimality in plausibility pre-orders or functions. Three major modalities emerge (Andersen et al., 2015):

• Conditional belief $B_i^\psi\varphi$ (“in context $\psi$, $i$ believes $\varphi$”)
• Degree-of-belief $B_i^{\geq n}\varphi$
• Safe belief $S_i\varphi$ (belief robust to adverse revisions)

Automated bisimulation notions for plausibility models exhibit that distinct belief logics—conditional, graded, safe—have pairwise-incomparable expressivity even when their bisimulation relations align.

Neighbourhood and Conditional Models

Conditional neighbourhood models formalize beliefs via agent-, world-, and condition-indexed families of “seriously possible” sets without encoding full numerical probabilities, capturing beliefs as willingness-to-bet (Eijck et al., 2017). Their logic is strictly weaker than full probabilistic logic but supports robust update/restriction operators and captures ambiguity aversion.

Model	Core object	Conditioned?	Quantitative?
Kripke	Accessibility relation $R$	No	No
Plausibility	Preorder $\ge$	Yes	Ordinal only
Probabilistic	Family of $\mu_{i,w}$	Yes	Yes
Neighbourhood	Family $N_{i,w}(X)$	Yes	Optional

4. Uncertainty in Topological, Metric, and Argumentative Settings

Logic over Metric Spaces

Expected-distance logic $\mathcal{L}_{ED}$ formalizes uncertainty when metric structure on the space of possibilities is relevant (Lee, 2012). The operator $ED_\mu^d(U)$ measures the average “distance-to-evidence,” unifying metric vagueness and probabilistic uncertainty; sound and complete axiomatics extend probability logic via inclusion-exclusion and metric compatibility. Product metrics enable multi-factor independence in spaces such as product event models.

Probabilistic Argumentation and Belief Maintenance

In environments requiring argumentation or evolving belief support (e.g., knowledge bases with conflicting, uncertain information), formal frameworks such as probabilistic PreDeLP (Shakarian et al., 2014) or general-purpose Belief Maintenance Systems (BMS) (Falkenhainer, 2013) instantiate dual-layer representations. These systems connect argument structures/dialectical trees (symbolic, defeasible reasoning) to probabilistic annotations, and employ operations like Dempster–Shafer combinations, meta-support revisions, and non-prioritized belief update.

5. Global Belief Models and Coherent Uncertainty Aggregation

To achieve coherent aggregation of local uncertainty assessments, the theory of upper expectations formalizes global belief via conservative, axiomatic extension principles (T'Joens et al., 2020). Given a sequence of local upper expectations (imprecise or set-valued), the unique most conservative global extension is characterized by monotonicity, local compatibility, conditionality, iterative law, and continuity. This “natural extension” coincides with Shafer–Vovk’s game-theoretic supermartingale upper expectations, and strictly dominates measure-theoretic suprema unless all locals are precise. This axiomatization unifies behavioural (Walley), game-theoretic (Shafer–Vovk), and measure-theoretic approaches in a provably unique, operationally computable model.

6. Dynamic, Stratified, and Category-Theoretic Approaches

Dynamic Logics and Stratification

Dynamic epistemic logics address temporal or informational evolution, such as public announcements or evidence updates, via reduction axioms and model restriction (Eijck et al., 2017, Ditmarsch et al., 2015). Stratified epistemic models, as in Ordered Epistemic Logic (OEL), employ acyclic layerings of theories where higher-level decisions are made with modal assertions about lower-level knowledge states, supporting robust, maintainable encoding of decision rules under incomplete information (Marković et al., 2023).

Category-Theoretic Foundations

In settings such as stochastic finance, categories of $\sigma$-complete Boolean algebras, filtration functors, and categorical morphisms integrate stochastic processes, agent beliefs, and group knowledge into unified categorical structures (Adachi, 2015). Modal semantics is then expressed via conditional expectations over generalized event algebras, supporting belief updates over time, group belief fixed-points, and harmonized probabilistic/logical interpretations.

---

These formal models create a layered and cross-linked toolkit for representing, reasoning, and quantifying knowledge, belief, and uncertainty in AI, distributed systems, decision support, epistemic game theory, and cognitive modeling. Current research focuses on unifying qualitative and quantitative modes, integrating metric, topological, and algebraic methods, enforcing global coherence and robustness, and extending expressive and computational capacities while preserving logic-based transparency and control (Ditmarsch et al., 2015, Lee, 2012, Andersen et al., 2015, Shakarian et al., 2014, Adachi, 2015, Marković et al., 2023, T'Joens et al., 2020, Craddock et al., 2013, Falkenhainer, 2013, Eijck et al., 2017, 1304.1508, Páez, 2020, Bjorndahl et al., 2019).