Ordinal Plausibility Measures for Qualitative Uncertainty

68 concepts52 papers2 proposals

When modeling how an artificial agent should reason under uncertainty, classical probability theory is often the default choice. However, standard probability demands exact numerical inputs—requiring an agent to state, for instance, that a given proposition has a probability of exactly $0.732$. In many real-world scenarios, such precision is unavailable or completely unnatural. An expert might easily state that one outcome is "more likely" than another, or that a certain event would be "highly surprising," without being able to assign a meaningful scalar value to that confidence.

Furthermore, forcing qualitative beliefs into a quantitative probabilistic mold can introduce logical paradoxes (such as the lottery paradox) and make common-sense default reasoning ("birds typically fly") computationally brittle. To resolve this tension, a powerful family of non-probabilistic frameworks has emerged. By shifting from continuous numerical intervals to partially ordered sets and discrete integer ranks, these frameworks provide a rigorous foundation for qualitative uncertainty, belief revision, and default reasoning.

Generalizing Uncertainty via Plausibility Measures

To understand why qualitative frameworks work, we must first look at what makes standard probability too restrictive. Probability spaces map events to the real-valued interval $[0, 1]$ and enforce additivity: the probability of two disjoint events occurring is the sum of their individual probabilities ($P(A \cup B) = P(A) + P(B)$).

To build a more flexible framework, Halpern et al. introduced plausibility measures in Plausibility measures and default reasoning and Plausibility Measures and Default Reasoning. Instead of mapping events to numbers, a plausibility measure maps sets of possible worlds to a partially ordered set (poset). The core mathematical requirement of additivity is dropped entirely, replaced by a much simpler property: monotonicity. If a set of worlds $A$ is a subset of $B$, then $B$ must be at least as plausible as $A$ ($A \subseteq B \implies Pl(A) \le Pl(B)$).

This simple abstraction is remarkably powerful. Because the target space is a poset rather than a strict line, it can gracefully accommodate scenarios where different certainties are simply incomparable, a concept early highlighted by Niblett in Judgmental Reasoning for Expert Systems. More importantly, the authors demonstrated that plausibility measures serve as a grand unification framework: standard probability distributions, fuzzy possibility measures, and Dempster-Shafer belief functions are all just special cases of this underlying qualitative structure.

By operating at this level of abstraction, researchers can design reasoning systems that do not rely on precise statistical frequencies. For example, Halpern et al. utilized plausibility structures in First-Order Conditional Logic Revisited to formalize default rules without falling into the mathematical traps that plague threshold-based probabilistic models. Similarly, Chu et al. developed Generalized Expected Utility (GEU) in Great Expectations. Part II: Generalized Expected Utility as a Universal Decision Rule, proving that substituting standard probabilities with plausibility measures allows a single decision rule to mathematically recover diverse frameworks ranging from Maximin to Minimax Regret.

The Intuition of Ordinal Rankings and Infinitesimals

While general posets offer excellent theoretical flexibility, practical AI applications often require a more concrete way to compute updates. This is where Ordinal Conditional Functions (OCFs)—originally introduced by Spohn in A general non-probabilistic theory of inductive reasoning—provide an elegant solution.

Instead of measuring how likely an event is, an OCF (often called a $\kappa$-ranking) measures how disbelieved or surprising a world state is. It maps possible worlds to non-negative integers:

  • A rank of $0$ represents maximum plausibility (the world is completely normal, expected, or consistent with current beliefs).
  • A rank of $1$ means the world is unexpected or exceptional.
  • Higher integer ranks ($2, 3, \dots$) indicate increasingly severe levels of exceptionality or disbelief.

Why does this integer-based ranking system work so perfectly as an analogue to probability? The core intuition is that these ranks behave exactly like the exponents of infinitesimals.

Imagine a standard probability model where the probability of an exceptional event is parameterized by a tiny fractional variable $\epsilon$ (where $\epsilon \to 0$). A normal event has a probability of order $\epsilon^0 = 1$. An unlikely event has a probability of order $\epsilon^1$. A highly catastrophic, doubly unexpected event has a probability of order $\epsilon^2$.

When we manipulate standard probabilities, we multiply them ($P(A \cap B) = P(A) \cdot P(B)$). But when we multiply powers of $\epsilon$, their exponents add up: $$\epsilon^a \cdot \epsilon^b = \epsilon^{a+b}$$

As formalized by Goldszmidt et al. in System-Z+: A Formalism for Reasoning with Variable-Strength Defaults, integer ranking functions are simply an order-of-magnitude abstraction of these exponents. Because of this relationship, complex probabilistic conditioning simplifies into basic integer arithmetic: multiplication becomes addition, and addition becomes a minimization operation (since the most plausible world dominates the sum of infinitesimals).

Pearl heavily exploited this order-of-magnitude abstraction in From Conditional Oughts to Qualitative Decision Theory and Qualitative MDPs and POMDPs: An Order-Of-Magnitude Approximation, demonstrating that agents can make robust, qualitative decisions using these intuitive integer inputs instead of demanding precise, hard-to-verify probability distributions.

Shifting States: Belief Revision and Update

An agent operating in a dynamic environment must constantly change its mind when faced with new information. Ordinal plausibility rankings provide a natural mechanism for tracking these cognitive shifts.

In the classical view of knowledge representation, changing one's mind takes two distinct forms:

  1. Revision: Correcting a mistaken belief about a static world (e.g., "I thought the keys were in the car, but I just found them in my pocket").
  2. Update: Tracking a changing world (e.g., "The keys were in my pocket, but I just dropped them into the river").

Friedman et al. demonstrated in Modeling Belief in Dynamic Systems, Part II: Revision and Update that by using plausibility measures over temporal histories, revision and update can be unified into a single mathematical operation: Bayesian-style conditioning applied to different types of prior structures.

Prior Ranking State (Normal world has rank 0)
   |
   +---> [Observation Received]
   |
   V
Shift Ranks via Integer Arithmetic
   |
   +---> If observation is reliable: New state has rank 0
   +---> If observation is fallible: Balance prior rank vs. observation rank

When an agent receives an observation, it shifts the integer ranks of its possible worlds. In early logical frameworks, observations were treated as absolute truths, forcing the rank of any contradictory world to infinity. However, real-world sensors and actions are fallible. Boutilier et al. addressed this in Belief Revision with Unreliable Observations by treating observations as noisy evidence that merely alters the ranking of worlds rather than eliminating them entirely.

To handle even more complex environments, Hunter et al. extended this in Belief Change in the Context of Fallible Actions and Observations by assigning integer ranks to both states and actions. If an agent attempts to open a door but still perceives it as locked, it can arithmetically balance the implausibility of the action failing against the implausibility of its sensor malfunctioning to retrospectively determine what most likely occurred.

Crucially, these updates must be controlled so that learning something new does not cause an agent to randomly discard unrelated knowledge. Jin et al. introduced a critical "Independence" postulate in Iterated Belief Revision, Revised, ensuring that the relative ordinal plausibility of independent beliefs remains perfectly preserved during successive updates.

Integrating with Logic, Modality, and Learning

Because ordinal rankings rely on discrete steps rather than continuous numbers, they bridge the historical divide between quantitative uncertainty and symbolic logic.

In traditional modal logic, a statement is either possible or impossible. By mapping worlds to qualitative rankings, Boutilier showed in Modal Logics for Qualitative Possibility and Beliefs that modal logic can completely capture the nuances of numerical possibility measures without needing the numbers themselves. This qualitative approach was further enriched by Hájek et al. in On Modal Logics for Qualitative Possibility in a Fuzzy Setting, who introduced comparative modalities ($\le$) directly into the logic, allowing systems to syntactically reason about statements like "Event A is at least as plausible as Event B."

This seamless integration with logic makes ordinal rankings highly practical for high-level agent programming and common-sense reasoning:

  • Logical Programming: Laverny et al. utilized OCFs in From knowledge-based programs to graded belief-based programs, part II: off-line reasoning to allow agents to simulate future belief states and plan actions in complex environments without the computational overhead of standard Partially Observable Markov Decision Processes (POMDPs).
  • Epistemic Reasoning: Delgrande et al. developed the Epistemic Logic of Likelihood (ELL) in Epistemic Logic of Likelihood and Belief, combining qualitative comparisons with deductive operators to let agents hold contingent beliefs without triggering threshold paradoxes.
  • Machine Learning: The utility of this framework extends all the way to modern online learning. Inoue et al. demonstrated in Iterated Belief Change as Learning that belief revision operators based on OCFs can be used to construct highly efficient online classifiers. Because these classifiers rank hypotheses using discrete consistency scores rather than probabilities, they are completely invariant to the order in which training data arrives and excel at maximizing recall in binary classification tasks.

Ultimately, ordinal plausibility measures work because they strip away the unnecessary precision of standard probability while retaining its core structural rigor. By replacing continuous multiplication with discrete integer addition, they offer a computationally elegant, logically compatible, and deeply intuitive way for artificial minds to navigate an uncertain world.

Key Foundational Work

Judgmental Reasoning for Expert Systems (1985) — contributed 1 concept elements to this direction.
A general non-probabilistic theory of inductive reasoning (1988) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_plausibility_measures_a196bc29,
  title = {plausibility measures},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/plausibility-measures/rdr_a196bc298b9a4ed3b7cde5c6dfbf6576},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_plausibility_measures_a196bc29_v1,
  title = {plausibility measures},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/plausibility-measures/rdr_a196bc298b9a4ed3b7cde5c6dfbf6576/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake