Decision Making with Partially Consonant Belief Functions

Phan Hong Giang, Prakash P. ShenoyUAI 2003

TL;DR

This paper introduces a novel decision-making framework for "partially consonant belief functions" (pcb), a generalization of probability and possibility theories. By developing a new axiomatic system and representation theorem, the authors provide a way to make decisions under uncertainty that is more expressive than traditional methods while still being consistent with statistical principles like the likelihood principle.

What Problem Does This Paper Address?

The paper tackles the challenge of making decisions when our uncertainty is not perfectly described by standard probability. Dempster-Shafer's belief function theory offers a more general way to represent uncertainty, linking it to evidence. However, applying this to decision-making has been difficult because it means moving away from the well-established Bayesian decision theory, which relies on maximizing expected utility. Existing approaches often try to convert belief functions back into probabilities, which can lose information or be ad hoc. This paper focuses on a specific, statistically consistent subclass of belief functions called "partially consonant belief functions" (pcb) and proposes a decision theory tailored to them.

Methods

The core of the paper is the development of a new axiomatic system for decision-making with pcb lotteries. This system is inspired by the von Neumann-Morgenstern (vNM) axioms for linear utility theory in probabilistic lotteries and similar axioms for qualitative utility in possibilistic lotteries. The authors then prove a representation theorem for this new axiomatic system. This theorem shows that a preference relation on pcb lotteries can be represented by a utility function that is a "mixed construct" of linear utility (from probabilistic lotteries) and binary utility (from possibilistic lotteries). They also review Walley's derivation of pcb and discuss its properties, including its relationship to statistical inference and the likelihood principle.

Datasets

Not discussed in the paper.

Key Insights and Ideas

  • Pcb as a Bridge: Partially consonant belief functions (pcb) are identified as a crucial subclass of belief functions that bridge the gap between probability and possibility, and importantly, are consistent with statistical principles like the likelihood principle.
  • Mixed Utility for Decision Making: The paper proposes a novel decision-making framework that uses a "mixed utility" function. This utility combines aspects of linear utility (from probability theory) and binary utility (from possibility theory), reflecting the dual nature of uncertainty captured by pcb.
  • Axiomatic Foundation: A new set of axioms is introduced for preference relations on pcb lotteries, providing a rigorous theoretical foundation for the proposed decision-making approach.
  • Consistency with Statistics: The pcb framework and the proposed decision theory are shown to be consistent with established statistical principles, addressing a key concern in generalizing probabilistic methods.
  • Addressing Ellsberg's Paradox: The paper demonstrates how their framework can explain observed behavior in scenarios like Ellsberg's paradox, where traditional expected utility theory struggles.

Key Results

The paper presents a representation theorem that establishes a utility function for pcb lotteries. This utility function is a combination of linear utility for probabilistic components and binary utility for possibilistic components. The authors also provide an example illustrating how their framework can explain behavior in Ellsberg's paradox, showing that their mixed utility calculations align with observed preferences, unlike some other approaches.

Limitations and Future Work

The paper acknowledges that Walley himself expressed reservations about pcb due to potential "sure-loss" arguments. The authors counter this by proposing a "likelihood gamble" framework, arguing that Walley's assumption about interpreting belief functions as betting rates is not always justified. Future work could explore the implications of this framework more broadly and investigate how to select the parameters within the mixed utility function. The paper also touches upon the idea of implicit priors in their likelihood gamble framework, suggesting potential avenues for further research into decision-maker attitudes towards ambiguity.

Summary updated

BibTeX
@inproceedings{giang2003decision,
  author = {Phan Hong Giang and Prakash P. Shenoy},
  title = {{Decision Making with Partially Consonant Belief Functions}},
  booktitle = {Conference on Uncertainty in Artificial Intelligence},
  year = {2003},
}
Figure