A Qualitative Linear Utility Theory for Spohn's Theory of Epistemic Beliefs

Phan Hong Giang, Prakash P. ShenoyUAI 2000

TL;DR

This paper introduces a "qualitative linear utility theory" for decision-making under uncertainty, specifically tailored for Spohn's theory of epistemic beliefs. It provides a formal framework for rational decision-making when beliefs are expressed qualitatively (as disbelief functions) rather than probabilistically, offering an analogue to the well-established von Neumann-Morgenstern utility theory.

What Problem Does This Paper Address?

Traditional decision theory, like the von Neumann-Morgenstern (vNM) utility theory, relies on probabilistic representations of uncertainty and quantitative utility functions. However, in many real-world scenarios, beliefs are not easily quantifiable into precise probabilities. Instead, they might be expressed more qualitatively, perhaps as "I strongly disbelieve X" or "I mildly believe Y." Spohn's theory of epistemic beliefs (also known as kappa calculus) offers a powerful framework for representing these qualitative beliefs, but a corresponding normative decision theory has been lacking. This paper aims to bridge that gap by developing a qualitative utility theory that aligns with Spohn's framework, allowing for rational decision-making when only qualitative beliefs are available.

Methods

The authors propose a set of axioms for a qualitative preference relation on lotteries, analogous to the axioms used in the vNM theory. These axioms include:

  • Ordering of prizes: Prizes are completely and transitively ordered.
  • Reduction of compound lotteries: Compound lotteries can be reduced to equivalent simple lotteries using Spohn's calculus for combining disbeliefs.
  • Substitutability: Indifferent lotteries are interchangeable.
  • Quasi-continuity: For each prize, there exists a standard lottery indifferent to it.
  • Transitivity: The preference relation is complete and transitive.
  • Qualitative monotonicity: This axiom formalizes how preferences change when the likelihood of a better or worse prize is altered in a standard lottery.

Based on these axioms, they prove a representation theorem. This theorem states that if a decision-maker's preferences satisfy these axioms, then there exists a unique "qualitative linear utility function" that assigns a value from a specific set (Bo) to each lottery. The utility of a compound lottery is then defined as the "qualitative expected utility," which involves minimization and addition operations, mirroring the multiplication and addition in the quantitative vNM expected utility formula.

Datasets

Not discussed in the paper. The paper focuses on theoretical development and uses hypothetical examples (like building houses in an earthquake zone) for illustration, rather than empirical datasets.

Key Insights and Ideas

  • Qualitative Beliefs, Qualitative Decisions: The core idea is that if your beliefs are qualitative, your decision-making framework should also be qualitative, rather than forcing a quantitative interpretation where it might not be natural.
  • Analogy to vNM: The paper successfully draws a strong parallel between their qualitative utility theory and the established quantitative vNM theory. This makes the concepts more accessible and highlights the theoretical soundness of their approach.
  • Spohn's Calculus as a Foundation: The paper demonstrates the utility of Spohn's disbelief functions not just for representing beliefs but also as a basis for a normative decision theory.
  • "Qualitative Expected Utility": The introduction of a new concept, "qualitative expected utility," which uses minimization and addition, provides a concrete way to calculate the desirability of lotteries under qualitative beliefs.
  • Discrete vs. Continuous Utility: A key difference noted is that qualitative utility uses a discrete scale, unlike the continuous scale of quantitative utility. This can lead to situations where fine distinctions in preference might not be captured as smoothly as in the probabilistic case.

Key Results

The main result is the representation theorem which establishes the existence and uniqueness of a qualitative linear utility function for lotteries based on Spohn's epistemic beliefs, provided the decision-maker's preferences satisfy a set of plausible axioms. This function allows for the calculation of "qualitative expected utility." The paper also provides an example illustrating how this theory can be applied to a hypothetical earthquake scenario, ranking the situation as "slightly unacceptable."

Limitations and Future Work

  • Discrete Scale: The paper acknowledges that the discrete nature of the qualitative utility scale might not always capture fine distinctions in preferences as well as continuous quantitative utility.
  • Axiom Justification: While the axioms are presented as plausible, their empirical validation or deeper theoretical justification in various contexts could be a direction for future work.
  • Comparison with Other Qualitative Theories: The paper briefly compares its approach to Dubois and Prade's possibility theory, highlighting differences in how compound lotteries are handled and the implications of certain axioms. Further exploration of how this theory relates to other qualitative decision-making frameworks could be beneficial.
  • Practical Application: While the earthquake example is illustrative, demonstrating the theory's application in more complex, real-world decision problems would be valuable.

Summary updated

BibTeX
@inproceedings{giang2000qualitative,
  author = {Phan Hong Giang and Prakash P. Shenoy},
  title = {{A Qualitative Linear Utility Theory for Spohn's Theory of Epistemic Beliefs}},
  booktitle = {Conference on Uncertainty in Artificial Intelligence},
  year = {2000},
}
Figure