TL;DR
This paper explores how different ways of numerically representing beliefs (quantitative measures like probability) align with more abstract, comparative ways of expressing beliefs (qualitative relations). They show that certain qualitative structures are compatible with specific quantitative belief measures, providing a theoretical foundation for combining these approaches and potentially leading to generalized utility theories.
What Problem Does This Paper Address?
In modeling uncertain situations, we often need to represent beliefs. Traditionally, this has been done in two main ways: quantitatively, by assigning a numerical value (like a probability) to a proposition, and qualitatively, by expressing preferences between propositions (e.g., "I believe A more than B"). While probability theory has extensively explored both quantitative and qualitative frameworks, the compatibility between various quantitative belief measures and qualitative preference structures, particularly within the context of belief functions, has been less understood. This paper aims to bridge this gap by investigating which qualitative preference structures are compatible with different quantitative belief measures, specifically focusing on probability functions, monotonic belief functions, Shafer's belief functions, and Smets' generalized belief functions. Understanding this compatibility is crucial for developing consistent and interpretable models of belief and uncertainty.
Methods
The authors analyze four classes of quantitative belief measures:
- Probability Function: Adheres to Kolmogorov's axioms, including additivity.
- Monotonic Belief Function: Extends belief functions by requiring an additional monotonicity axiom.
- Shafer's Belief Function: Based on Dempster-Shafer theory, it allows for belief to be assigned to sets of propositions and is characterized by a "basic probability assignment."
- Smets' Generalized Belief Function: A generalization of Shafer's belief function that can handle open-world assumptions (where the true answer might not be in the defined frame).
They then define and analyze various qualitative preference relations, characterized by axioms such as asymmetry, transitivity, and monotonicity. The core of their method involves establishing formal compatibility criteria: a quantitative measure is compatible with a qualitative relation if the relation's ordering of propositions is preserved by the quantitative measure's numerical assignments. They prove theorems that link specific qualitative preference structures to each of the four quantitative belief measures.
Datasets
Not discussed in the paper. The paper is theoretical and focuses on formal axiomatic relationships rather than empirical evaluation on datasets.
Key Insights and Ideas
- Hierarchical Compatibility: The paper reveals a hierarchy of compatibility. Qualitative probability relations are compatible with monotonic belief functions, which are a subset of Shafer's belief functions. A slightly weaker qualitative structure is compatible with Smets' generalized belief functions.
- Foundation for Generalized Utility: The compatibility findings suggest a theoretical basis for developing generalized utility theories that can handle uncertainty represented by belief functions, moving beyond traditional expected utility theory.
- Axiomatic Characterization: The work provides axiomatic characterizations for preference relations that are compatible with specific belief measures, offering a rigorous way to define and understand these relationships.
- Bridging Quantitative and Qualitative: The research demonstrates that quantitative and qualitative representations of belief are not mutually exclusive but can be integrated and understood in relation to each other.
Key Results
- Qualitative probability relations are compatible with monotonic belief functions.
- A preference structure slightly weaker than qualitative belief is compatible with Smets' generalized belief functions.
- The paper provides theorems establishing necessary and sufficient conditions for a preference relation to be represented by a probability function, a belief function, and a monotonic belief function.
- Figure 1 and Figure 2 visually summarize the relationships between the quantitative belief measures and their corresponding qualitative preference structures, illustrating inclusion and implication relationships.
Limitations and Future Work
- Generalization of Belief Functions: The authors acknowledge that while they characterize compatibility for several key belief functions, a comprehensive set of axioms for preference relations that can only be represented by belief functions (and not necessarily probability functions) might be a difficult task.
- Consonant Belief Functions: They mention that Dubois (1986) has proposed axioms for consonant belief functions, suggesting this as an area for more detailed analysis.
- Broader Analysis: The paper focuses on four specific classes of belief measures. Future work could explore compatibility with other types of belief functions or uncertainty measures.
- Practical Applications: While the theoretical implications for generalized utility theory are highlighted, further research is needed to translate these findings into practical decision-making frameworks.
Summary updated