TL;DR
This paper proposes a framework for belief revision in "deductively open belief spaces" (DOBS), which are more realistic for implemented AI systems than traditional "deductively closed" spaces. The authors aim to adapt existing belief revision postulates to this more flexible setting, enabling better evaluation and comparison of AI systems.
What Problem Does This Paper Address?
Traditional belief revision research has largely focused on "deductively closed belief spaces" (DCBS). In a DCBS, if a set of beliefs is known, then all logical consequences of those beliefs are also considered known and part of the belief space. This is a strong assumption that is often difficult to meet in practice for real-world AI systems due to computational limitations and the sheer scale of potential inferences.
Implemented AI systems often operate in "deductively open belief spaces" (DOBS). In a DOBS, the system maintains a base set of assertions, and other beliefs are derived as needed. This means that not all logical consequences of the base assertions are necessarily explicit or readily available. This "openness" makes it challenging to directly apply existing belief revision postulates, which are designed for the more rigid structure of DCBS. The paper addresses the gap of how to formally define and evaluate belief revision in these more practical, deductively open environments, which is crucial for applications like data fusion where conflicting information is common.
Methods
The paper's core contribution is the proposal of a formalism for describing DOBS and their integrity constraints (ICs) and postulates. This formalism is intended to guide the development of belief revision techniques specifically for implemented systems. The authors plan to adapt existing, well-established postulates for belief revision, such as the AGM postulates, Hansson's base contraction postulates, and postulates for ranked beliefs, to this DOBS framework. This adaptation will allow for a more direct comparison of how well implemented systems adhere to theoretical standards.
A key aspect of their approach involves addressing the open discussion around how to properly weight and combine integrity constraints IC3 (minimizing damage to the belief space) and IC4 (removing less entrenched beliefs) during contraction or revision. They also highlight their ongoing work on enhancing their implemented belief revision system, SNeBR, to align with these DOBS postulates. Notably, their system will use qualitative, partial orderings for belief strength (entrenchment) rather than quantitative measures, allowing for the inclusion of information whose credibility is only known relative to other beliefs.
Datasets
Not discussed in the paper.
Key Insights and Ideas
- Realism in Belief Revision: The paper emphasizes that deductively open belief spaces are a more realistic model for implemented AI systems than deductively closed ones, bridging a gap between theoretical ideals and practical application.
- Adaptability of Postulates: Existing, foundational belief revision postulates can and should be adapted to the DOBS setting, enabling a more meaningful evaluation of AI systems.
- Qualitative Entrenchment: Using qualitative, partial orderings for belief strength (entrenchment) is a practical approach that allows for nuanced reasoning about belief credibility, especially when precise quantitative measures are unavailable.
- System Comparison Framework: The proposed formalism aims to provide a standardized way to compare different belief revision systems, aiding implementers in understanding their systems' strengths and weaknesses.
Key Results
Not discussed in the paper. The paper outlines a research direction and proposed methodology rather than presenting experimental results.
Limitations and Future Work
The paper acknowledges that the discussion on how to properly weight and combine IC3 and IC4 remains an open research question. Future work will involve developing the DOBS version of various postulates, providing comments on postulate adherence for paraconsistent logics and incomplete systems, and ultimately developing a theory for comparing systems. The authors also plan to continue refining their SNeBR system to adhere to the proposed DOBS postulates.
Summary updated