Abstract
Belief revision and belief update have been proposed as two types of belief change serving different purposes. Belief revision is intended to capture changes of an agent's belief state reflecting new information about a static world. Belief update is intended to capture changes of belief in response to a changing world. We argue that both belief revision and belief update are too restrictive; routine belief change involves elements of both. We present a model for generalized update that allows updates in response to external changes to inform the agent about its prior beliefs. This model of update combines aspects of revision and update, providing a more realistic characterization of belief change. We show that, under certain assumptions, the original update postulates are satisfied. We also demonstrate that plain revision and plain update are special cases of our model, in a way that formally verifies the intuition that revision is suitable for "static" belief change.
Introduction
In the field of Knowledge Representation (KR), artificial agents must constantly reconcile new information with existing knowledge. Traditionally, this process has been split into two distinct paradigms: Belief Revision (AGM theory) and Belief Update (KM theory). Revision assumes a static world where the agent corrects its own past errors, while update assumes a dynamic world where the agent tracks physical changes.
However, real-world scenarios often blur these lines. If an agent observes an unexpected outcome, it must decide whether the world changed in an unusual way or if its initial assumptions about the world were flawed. This paper introduces Generalized Update (GU), a framework that unifies these concepts by allowing observations of the present to inform beliefs about the past.
The Generalized Update Framework
The GU model is built upon the semantics of plausibility rankings, specifically Spohn’s $\kappa$-rankings. In this system, every possible world state is assigned a rank $k \ge 0$, where $k=0$ represents the most plausible states and higher values represent increasingly "surprising" or less plausible states.
Beyond static states, GU incorporates Event Models. Transitions between states are modeled as events (or actions), each with its own plausibility rank. When an agent receives a new observation, it performs abductive reasoning to find the most plausible explanation—a sequence involving an initial state and a transition event—that accounts for the current data.
Illustrative Scenarios
The power of GU is demonstrated through examples where traditional theories fail to capture the full scope of belief change.
- The Warehouse Example: An agent expects a truck to be delayed by snow. If the truck arrives on time, the agent should not only update its belief that the truck is present but also revise its prior belief about the weather conditions.
- The Litmus Test: A sensing action, like testing a liquid with litmus paper, is modeled as an event. The outcome of the test (color change) provides information that forces a revision of the agent's belief about the liquid's initial chemical properties (acid vs. base).
Formal Unification and Results
A key contribution of this work is the formal proof that both AGM revision and KM update are special cases of Generalized Update:
- Static Verification: If the world is assumed to be static (only the "null" event occurs), GU satisfies the standard AGM revision postulates ($R1-R8$).
- Update Validation: Under "centered" conditions (where the agent is certain of the current state), GU satisfies the core KM update postulates.
This unification allows agents to trade off the likelihood of rare events against the likelihood of prior misconceptions, providing a robust mechanism for routine belief change.
Conclusion
Generalized Update provides a more realistic characterization of how agents should handle information in dynamic environments. By treating revision and update as orthogonal roles within a single framework, GU enables "revision through update." Future work remains in addressing the computational complexity of managing these rankings and extending the model to handle iterated updates over long sequences of observations.
Summary updated