Qualitative Decision Theory with Sugeno Integrals

Didier Dubois, Henri Prade, Régis SabbadinUAI 1998

TL;DR

This paper introduces a qualitative decision theory framework using Sugeno integrals, offering a way to make decisions under uncertainty without precise numerical probabilities or utilities. It provides an axiomatic foundation for this approach, showing how it generalizes existing qualitative decision criteria and offers a more nuanced perspective on decision-making than traditional expected utility theory.

What Problem Does This Paper Address?

Traditional decision theory, like Savage's expected utility framework, relies on precise numerical representations of uncertainty (probabilities) and preferences (utilities). However, in many real-world scenarios, especially in Artificial Intelligence, decision-makers may only have qualitative information about their beliefs and desires. This paper addresses the challenge of building a decision-making framework that can operate with this limited, qualitative information. It seeks to provide an axiomatic justification for decision rules that don't require precise quantification, moving beyond simple averaging to capture more nuanced decision-making attitudes, such as optimism and pessimism.

Methods

The core of the paper is the development of a qualitative decision theory based on Sugeno integrals. A Sugeno integral is presented as a qualitative counterpart to the Choquet integral, which itself generalizes expected utility. Unlike expected utility, which relies on averaging, the Sugeno integral uses a "sup-min" (or "max-min" in the finite case) operation, making it akin to a median. The authors provide an axiomatic justification for this Sugeno integral-based utility, drawing parallels to Savage's axiomatic approach for expected utility. They define axioms that capture preference relations between acts (choices) and then show how these axioms lead to a Sugeno integral representation. The paper also explores how specific types of qualitative measures, like possibility and necessity measures, can be used within the Sugeno integral framework to represent optimistic and pessimistic decision-making attitudes.

Datasets

Not discussed in the paper.

Key Insights and Ideas

  • Sugeno Integral as a Qualitative Median: The paper highlights that the Sugeno integral, unlike expected utility (an average), acts like a median. This provides a qualitative way to aggregate uncertain outcomes, offering a different perspective on decision evaluation.
  • Axiomatic Foundation for Qualitative Decisions: By adapting Savage's axiomatic approach, the authors provide a rigorous justification for decision-making based on qualitative information, moving beyond ad-hoc rules.
  • Generalizing Pessimism and Optimism: The framework naturally accommodates optimistic and pessimistic decision-making styles by using possibility and necessity measures within the Sugeno integral. This allows for a more flexible representation of decision-maker attitudes than standard expected utility.
  • Non-Compensatory Nature: The Sugeno integral's "max-min" operation means it's non-compensatory. This is a significant departure from expected utility, where a bad outcome in one state can be offset by a good outcome in another. This aligns better with situations where decision-makers might not want to trade off certain types of risks.
  • Weakening Savage's Axioms: The paper demonstrates that the Sugeno integral does not satisfy all of Savage's original axioms (specifically the "sure-thing principle"). This is a crucial insight, showing the fundamental differences between qualitative and quantitative decision-making.

Key Results

The paper establishes an axiomatic framework for qualitative decision theory using Sugeno integrals. It shows that:

  • The Sugeno integral provides a qualitative utility function that can be axiomatized.
  • Specific qualitative measures (possibility and necessity) within the Sugeno integral framework correspond to optimistic and pessimistic decision-making criteria, respectively.
  • The framework is non-compensatory and generalizes certain qualitative decision rules.
  • The paper presents specific axioms (WS 3, RCD, RDD, DD, CD) that characterize monotonic qualitative utility, optimistic utility, and pessimistic utility.

Limitations and Future Work

  • Finite Setting: The current framework is primarily developed for finite sets of states and consequences. Extending it to infinite settings would be a significant next step.
  • Conditional Acts: The paper acknowledges that handling conditional acts in a qualitative setting is non-trivial, especially due to the failure of the "sure-thing principle." This is identified as a key area for future research.
  • Commensurability of Certainty and Priority Levels: The authors note that a strong assumption made is the commensurability of certainty levels and priority levels. Relaxing this assumption, as explored in other work, leads to more complex decision methods.
  • Observing Decision-Maker Behavior: While the framework provides theoretical foundations, practical methods for observing and eliciting these qualitative preferences from decision-makers would be valuable.

Summary updated

BibTeX
@inproceedings{dubois1998qualitative,
  author = {Didier Dubois and Henri Prade and Régis Sabbadin},
  title = {{Qualitative Decision Theory with Sugeno Integrals}},
  booktitle = {Conference on Uncertainty in Artificial Intelligence},
  year = {1998},
}
Figure