Belief Revision and Update in Dynamic Environments

18 concepts14 papers1 proposal

When an intelligent agent receives new information that contradicts its current internal model, it must change its mind. However, as Hirofumi Katsuno and Ken Satoh argue in A Unified View of Consequence Relation, Belief Revision and Conditional Logic, not all changes are created equal. The fundamental challenge in formal logic is distinguishing between revision—correcting an error about a static world—and update—tracking a world that has physically changed.

The Static World vs. The Dynamic World

To understand why this distinction matters, consider a simple scenario: an agent believes a room's light is off. Suddenly, the agent observes that the light is on. There are two ways to interpret this:

  1. Revision (Static World): The agent was simply mistaken. The light was always on, and the agent is now refining its incomplete knowledge of a fixed reality.
  2. Update (Dynamic World): The agent was correct initially, but an event occurred—someone flipped the switch. The world itself has transitioned to a new state.

As Didier Dubois and Henri Prade explain in Belief Revision and Updates in Numerical Formalisms, these two processes require different mathematical treatments. Revision is a process of selection: the agent eliminates possible worlds that are inconsistent with the new evidence. Update is a process of transition: the agent moves each possible world it previously considered to a new, successor state.

Why Traditional Logic Struggles

Most early frameworks for belief change, such as the AGM postulates, were designed for static environments. Daniel Lehmann 0001 notes in Belief Revision, Revised that these models often fail when information arrives in sequences. In a static world, if you learn $A$ and then learn $\neg A$, you conclude the second source is more reliable or you were wrong. In a dynamic world, $A$ followed by $\neg A$ is perfectly natural—it simply means the state of the world changed over time.

Katsumi Inoue and Nicolas Schwind point out in Characterization of Logic Program Revision as an Extension of Propositional Revision that this gap is particularly visible in logic programming. When previous truths are superseded by time rather than error, semantic models must account for "temporal changes and robust equivalence" rather than just resolving logical contradictions.

Unifying the Dichotomy

The separation between revision and update creates a dilemma for agents: how do they know which operator to apply? In Generalized Update: Belief Change in Dynamic Settings, Craig Boutilier argues that agents often face scenarios where a new observation implies both that the world has changed and that previous assumptions were incorrect.

To bridge this gap, researchers have proposed unified semantic frameworks:

  • Plausibility over Histories: Joseph Halpern et al. suggest in Modeling Belief in Dynamic Systems, Part II: Revision and Update that revision and update are two sides of the same coin. By applying Bayesian-style conditioning to "temporal histories" (sequences of states) rather than single states, they show that the distinction depends on whether the uncertainty lies in the initial state or the transitions that followed.
  • State Transition Models: Yasuo Kudo and Tetsuya Murai propose an "Acceptance" operator in A Unified Framework of Propositional Knowledge Base Revision and Update Based on State Transition Models. This operator uses a mathematical bridge to automatically decide whether to treat new information as a correction or a world-event.
  • Update as Action: Jérôme Lang reframes the problem in Belief Update Revisited by treating update as a "feedback-free action progression." In this view, updating is not just observing a new fact, but modeling the physical effect of an action that changes the state of the world.

Practical Implications for AI

The distinction is not merely theoretical; it affects how agents communicate and plan. Thomas Eiter et al. demonstrate in How Hard to Tell? Complexity of Belief Manipulation Through Propositional Announcements that the computational complexity of influencing an agent's beliefs changes drastically depending on whether the agent uses revision or update logic.

Ultimately, as Joseph Halpern et al. argue in An Operational Semantics for Knowledge Bases, moving from a static collection of formulas to an "operational" model of runs and states is essential. For an AI to function in the real world, it must be able to distinguish between learning more about its environment and tracking the inevitable changes that occur within it.

Key Foundational Work

A Unified View of Consequence Relation, Belief Revision and Conditional Logic (1991) — contributed 2 concept elements to this direction.
A Logic for Revision and Subjunctive Queries (1992) — 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_belief_change_dichotomy_356630dc,
  title = {belief change dichotomy},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/belief-change-dichotomy/rdr_356630dc60e941b981f634808fe2b794},
  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_belief_change_dichotomy_356630dc_v1,
  title = {belief change dichotomy},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/belief-change-dichotomy/rdr_356630dc60e941b981f634808fe2b794/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