When an artificial intelligence encounters information that contradicts its current knowledge, it cannot simply "add" the new fact to its database. Doing so in standard logic leads to a "logical explosion" where every possible statement becomes simultaneously true and false. To reason rationally, the agent must decide which of its previous beliefs to discard and which to keep.
Traditional Bayesian approaches solve this using precise numerical probabilities. However, as Tim Niblett argued in Judgmental Reasoning for Expert Systems, forcing human-like uncertainty into a 0-to-1 scale often fails to capture the qualitative nature of judgment. Instead, a robust research direction has emerged that models an agent’s mental state as a plausibility ranking over possible worlds.
The Intuition: Layers of Surprisingness
Imagine the set of all possible ways the world could be. In this framework, an agent doesn't just hold a list of facts; it organizes these "possible worlds" into layers, like the skins of an onion.
The innermost layer contains the worlds the agent currently believes are most likely (the "most plausible"). As you move outward, the worlds become increasingly "surprising" or "implausible." This structure is formally known as an Ordinal Conditional Function (OCF) or a Ranking Function.
In System-Z+: A Formalism for Reasoning with Variable-Strength Defaults, Moisés Goldszmidt and Judea Pearl demonstrated that these integer-based ranks act as qualitative abstractions of probabilities. A rank of 0 means "not surprising at all," while a rank of 10 means "extremely unlikely." This allows an agent to say "birds typically fly" by ranking worlds where birds fly at 0 and worlds where they don't at 1.
Revision vs. Update: Why the World Matters
A critical distinction in this field, synthesized by Joseph Halpern and Nir Friedman in Modeling Belief in Dynamic Systems, Part II: Revision and Update, is how an agent handles new information based on its source:
- Belief Revision: The agent realizes its initial information was wrong. It is correcting a mistake about a static world.
- Belief Update: The agent realizes the world itself has changed.
In a plausibility framework, Revision involves shifting the "spotlight" to a different layer of the onion. If I believe "the car is in the garage" (Rank 0) but see it is empty, I look to the next most plausible layer—perhaps "the car was stolen" (Rank 1)—and make that my new reality.
Update, however, requires a more complex transformation. As explored by Quanlong Guan and colleagues in An Investigation of Darwiche and Pearl's Postulates for Iterated Belief Update, updating involves tracking how the relative plausibility of scenarios evolves over time as events occur, which requires mapping internal states to partial preorders of possible histories.
Iterated Revision: Keeping a Consistent Policy
The hardest problem in belief change is iteration: what happens when you change your mind, and then change it again?
If an agent only tracks its current beliefs, it loses the "meta-information" about why it held those beliefs. Gabriele Kern-Isberner addresses this in Postulates for Conditional Belief Revision by treating an epistemic state not as a set of facts, but as the entire ranking function itself.
When new information arrives, the agent doesn't just pick a new "most plausible" world; it reorders the entire hierarchy. This ensures that if the new information is later retracted, the agent can "remember" its previous convictions. Yi Jin 0006 and Michael Thielscher further refined this in Iterated Belief Revision, Revised by introducing an "Independence" postulate, ensuring that learning something new about the weather doesn't accidentally cause the agent to change its mind about what it had for breakfast.
Beyond Total Orders: Handling Incomparability
Most early models assumed a "total preorder," meaning an agent could always say whether World A was more, less, or equally plausible than World B. However, real-world information is often incomparable.
Hua Meng and others in Belief Revision with General Epistemic States have pushed the field toward General Epistemic States. By relaxing the requirement for a total ordering, agents can maintain agnosticism when faced with incomplete or conflicting evidence. This is mathematically supported by Belief Algebras, which rank subsets of worlds rather than individual states, allowing for a more nuanced representation of "I don't know enough to compare these two scenarios."
Why It Works
The power of plausibility rankings lies in their semi-quantitative nature. They provide the mathematical rigor of probability—allowing for Bayesian-style updates as shown by Craig Boutilier in Belief Revision with Unreliable Observations—without requiring the agent to assign arbitrary decimal points to every possibility. By organizing the world into layers of "surprisingness," AI systems can navigate contradictions and changing environments with a stability that mimics human common sense.