TL;DR
This paper introduces "Generalized Update" (GU), a novel model for belief change that unifies belief revision (handling static worlds) and belief update (handling dynamic worlds). GU allows agents to simultaneously correct prior beliefs and adapt to a changing world, offering a more realistic approach to how intelligent agents reason.
What Problem Does This Paper Address?
Intelligent agents need to maintain beliefs about their environment. Traditionally, two main mechanisms handle belief changes: belief revision (correcting mistaken beliefs in a static world, like the AGM theory) and belief update (adapting beliefs to a changing world, like the KM theory). However, these are often too restrictive. In reality, an agent's reasoning often involves both correcting prior misconceptions and adapting to new information about a dynamic world. For instance, if a truck arrives earlier than expected, an agent might infer that its prior belief about traffic conditions (e.g., snow) was wrong, rather than just accepting the truck's arrival as a change in a static world. Current models struggle to naturally handle situations where an observation about a change in the world also provides information about the accuracy of prior beliefs.
Methods
The paper proposes a Generalized Update (GU) model built upon a ranked belief structure. Instead of just a set of beliefs (K), an agent's state is represented by a ranking function (κ) that assigns a plausibility score to possible "worlds" (states of affairs). This ranking captures both the agent's core beliefs (most plausible worlds) and the degree of entrenchment of those beliefs.
GU incorporates events (e.g., rain, a truck arriving) that map worlds to rankings of possible outcomes. The plausibility of a transition from one world to another via an event is a combination of the initial world's plausibility, the event's plausibility, and the outcome's plausibility.
When an observation (A) occurs, GU identifies the most plausible transitions that could lead to A. The agent's new beliefs are then derived from the worlds resulting from these most plausible transitions. This process naturally allows for:
- Revision: If an observation contradicts prior beliefs, GU can select transitions that implicitly revise those prior beliefs.
- Update: GU naturally handles changes in the world by considering plausible event occurrences and their outcomes.
The paper formalizes this by showing that GU satisfies many of the standard KM update postulates under certain assumptions (completeness and centeredness of the model). Crucially, it demonstrates that plain KM update and AGM revision are special cases of GU.
Datasets
Not discussed in the paper.
Key Insights and Ideas
- Unification of Revision and Update: The core insight is that real-world belief change isn't strictly revision or update, but a blend. GU provides a single framework to capture this.
- Ranked Beliefs as the Foundation: Using ranked belief structures (epistemic states) is key. It allows for nuanced reasoning about plausibility, which is essential for both correcting prior beliefs and adapting to new information.
- Events as the Drivers of Change: The model emphasizes that events (actions, occurrences) are the impetus for change. The plausibility of these events and their outcomes dictates how beliefs are updated.
- Abductive Explanation: GU can be viewed as an abductive process where observations are explained by postulating the most plausible sequences of events and prior states.
- Revision as a Static Update: The paper formally shows that AGM belief revision can be modeled as a special case of GU where only a "null event" (no change) is considered, reinforcing the intuition that revision applies to static worlds.
Key Results
- The paper presents a formal model for Generalized Update (GU) based on ranked belief structures and events.
- It demonstrates that GU satisfies many standard KM update postulates under specific conditions (completeness and centeredness).
- It formally proves that both classical belief revision (AGM) and classical belief update (KM) are special cases of GU.
- Illustrative examples (e.g., wet grass, litmus test) show how GU can handle scenarios where observations simultaneously inform about world changes and correct prior beliefs.
Limitations and Future Work
- Representation Languages: The authors note that developing concise representation languages for events (with defeasible effects), defeasible beliefs, and other aspects of the model is an ongoing task.
- Beliefs about Dynamics: The current model assumes events and outcomes are fixed. Future work could explore agents having beliefs about these dynamics themselves, which could also be subject to revision.
- Iterated Updates: Handling sequences of events and observations (iterated updates) is a significant challenge. The paper suggests that GU's framework might be extended to reason about "histories" or sequences of events, potentially unifying explanation with belief revision.
- Concurrent Events: Reasoning about concurrent events and actions is noted as being beyond the scope of the current paper.
Summary updated