TL;DR
This paper introduces a novel way to think about "independence" within possibility theory, moving beyond probabilistic definitions. They propose three new types of independence (unrelatedness, strong, and weak) that better capture intuitive notions of how information influences beliefs, particularly in exception-tolerant reasoning systems.
What Problem Does This Paper Address?
In reasoning under uncertainty, the concept of "independence" is crucial. Probabilistically, two events A and C are independent if knowing A doesn't change the probability of C. However, this definition has limitations, especially when dealing with qualitative reasoning or situations where probabilities are not precisely known. The authors argue that existing probabilistic independence doesn't always align with intuitive notions of dependence, particularly in scenarios involving logical relationships and exceptions. They aim to develop a more suitable framework for independence within possibility theory, which deals with degrees of possibility rather than strict probabilities, and apply it to plausible reasoning.
Methods
The paper proposes three new definitions of independence within possibility theory:
- Unrelatedness (≈₂): This is Zadeh's original notion, based on whether the possibility of the conjunction of two events differs from the minimum of their individual possibilities. It's a weaker form of independence.
- Strong Independence (=>): This definition is based on conditional possibility and necessity measures. It captures a stronger sense of independence where the acceptance of one proposition doesn't affect the acceptance of another. It's formally defined using necessity measures, N(CIA) = N(C) > 0.
- Weak Independence (≠>w): This is a milder form of independence, focusing on the unconditional acceptance of a proposition. It's defined as N(CIA) > 0 and N(C) > 0.
The authors analyze the formal properties of these definitions, comparing them to probabilistic independence and to each other. They also provide axiomatizations for strong independence and discuss its relationship to belief revision and exception-tolerant reasoning.
Datasets
Not discussed in the paper.
Key Insights and Ideas
- Possibilistic Independence is Different: The paper highlights that independence in possibility theory behaves differently from probabilistic independence, particularly concerning symmetry and negation.
- Multiple Flavors of Independence: The authors demonstrate that a single notion of independence is insufficient for possibility theory. They introduce "unrelatedness," "strong independence," and "weak independence," each capturing different intuitive aspects.
- Strong Independence Captures Qualitative Possibility: Strong independence is shown to be expressive enough to fully characterize qualitative possibility orderings, providing a way to axiomatize them without directly referring to the underlying possibility measure.
- Weak Independence for Belief Revision: Weak independence is presented as a natural fit for belief revision, aligning with the idea that beliefs independent of new information should be retained.
- Addressing Blocking of Property Inheritance: The paper suggests that these new notions of independence, particularly weak independence, can help solve problems like blocking of property inheritance in exception-tolerant reasoning systems.
Key Results
- The paper provides formal definitions and analyzes the properties of three new types of independence in possibility theory: unrelatedness, strong independence, and weak independence.
- A complete axiomatization for strong independence is provided.
- It's shown that strong independence can express qualitative possibility, allowing for an axiomatization of the dependence relation itself.
- Weak independence is linked to belief revision, where it signifies that a proposition remains accepted after revision if it was already accepted and is independent of the new information.
- The authors suggest that these new notions can help address issues in exception-tolerant reasoning, such as blocking of property inheritance.
Limitations and Future Work
- Completeness for Unrelatedness: The authors conjecture that unrelatedness cannot be fully axiomatized independently of the possibility measure.
- Ternary Independence: The paper focuses on binary independence relations for simplicity but acknowledges that ternary relations (e.g., "A and C are independent given B") are more general and warrant further investigation.
- Conditional Possibility Definition: The paper primarily uses a "min-operation" definition for conditional possibility. The case where a "product-operation" is used is left for future research.
- Application to Complex Scenarios: While the paper outlines applications to exception-tolerant reasoning, further development is needed to fully integrate these concepts into practical AI systems, especially concerning graph-based reasoning and aggregation of conditional distributions.
- Bayesian Method Adaptation: Adapting Bayesian methods for these possibilistic independence notions is challenging due to differences in how distributions are handled and the non-symmetric nature of weak independence.
Summary updated