Imagine an agent—a robot or a software system—that believes a specific door is closed. Suddenly, the agent receives new information: the door is actually open. To maintain a consistent worldview, the agent must change its beliefs. But how?
In classical AI, this problem splits into two distinct scenarios. In the first, the agent was simply mistaken; the door was open all along, and the agent is correcting an error about a static world. This is Belief Revision. In the second, the world itself changed; someone walked by and opened the door. This is Belief Update.
For decades, these two processes were governed by different logical rules. Recent research, however, has focused on bridging this gap, showing that revision and update are not competing theories but two facets of a single, unified mechanism for reasoning about change.
The Semantic Divide: AGM vs. KM
The distinction between these two modes of change was famously formalized by Hirofumi Katsuno and Ken Satoh in A Unified View of Consequence Relation, Belief Revision and Conditional Logic. They provided a "Rosetta Stone" for understanding the different logical postulates involved:
- Revision (AGM Theory): When learning about a static world, the agent should treat the new information as a constraint that narrows down the set of possible worlds. It picks the "closest" or "most plausible" worlds where the new information is true.
- Update (KM Theory): When the world changes, the agent must acknowledge that every possible state it previously considered might have transitioned into a new state. Instead of narrowing down a global set of possibilities, it updates each possible world independently.
As Jérôme Lang argues in Belief Update Revisited, update is essentially a "feedback-free action progression." It treats the new information as the result of a physical event rather than a mere observation.
Unification through Plausibility
The most influential step toward unification came from Joseph Halpern and Nir Friedman. In their series of papers, including Modeling Belief in Dynamic Systems, Part II: Revision and Update, they proposed that both revision and update can be viewed as conditioning on a prior.
The difference lies in what you are conditioning on. If you assume the world is static, you condition on the current state. If you assume the world is dynamic, you condition on a "history" or a sequence of events. By using "plausibility measures"—a generalized form of probability—they showed that an agent doesn't need two different sets of logic. It only needs a single mechanism that accounts for how the world evolves over time.
This perspective allows for a "Belief Change System" where the agent can naturally decide whether a surprising observation means they were wrong (revision) or the world moved on (update), often by applying a qualitative version of the Markov assumption, as explored in A Qualitative Markov Assumption and its Implications for Belief Change.
The Mechanics of Integrated Change
To make this unification practical, researchers like James P. Delgrande and Aaron Hunter 0001 developed frameworks that handle sequences of both actions and observations. In Iterated Belief Change: A Transition System Approach, they introduced a "belief evolution operator."
This operator treats the agent's mental state as a trajectory through a state space. When an agent performs an action (an ontic change), the entire trajectory shifts. When it receives a sensor reading (an epistemic change), it prunes trajectories that are no longer plausible. This unified approach solves long-standing paradoxes where standard logic failed to reconcile how new observations could force an agent to rethink its past actions.
From Logic to Numerical Measures
The intuition behind this unification also extends to numerical reasoning. Didier Dubois and Henri Prade demonstrated in Belief Revision and Updates in Numerical Formalisms that the revision/update distinction maps directly onto two mathematical operations:
- Conditioning: Used for revision. It refines an existing distribution by zeroing out impossible states and normalizing the rest.
- Imaging: Used for update. It "shifts" the probability mass from an old state to the most similar new state where the update holds.
By recognizing that these are just different ways of redistributing "belief mass," the field has moved toward a more robust architecture for intelligent agents. Whether an agent is correcting a database error or tracking a moving target, it is now seen as performing the same fundamental task: maintaining a coherent history of a world in flux.