A Savage-style Utility Theory for Belief Functions

Chunlai Zhou, Biao Qin, Xiaoyong DuIJCAI 2018 · OpenReview

Abstract

This paper establishes the first axiomatic foundation for decision-making using Dempster-Shafer belief functions within a finite state space. By introducing a set of six core axioms, the authors demonstrate that a rational agent's preferences can be represented as the maximization of expected utility relative to a subjective belief function, bridging the gap between classical decision theory and the treatment of ambiguity in artificial intelligence.

Introduction: Beyond Additive Probability

In classical decision theory, Savage’s Theorem (1954) serves as the foundational pillar. It posits that if an agent's choices adhere to specific "rational" axioms, their behavior is mathematically indistinguishable from maximizing expected utility based on a personal probability distribution. However, Savage’s framework relies on additive probability, which struggles to account for "ambiguity"—situations where the probabilities themselves are unknown or ill-defined, as famously highlighted by the Ellsberg Paradox.

The Dempster-Shafer theory of belief functions provides a robust alternative for modeling such uncertainty. Despite its utility in AI and evidence theory, it long lacked a rigorous "Savage-style" justification based on simple preferences over acts in a finite setting. This research addresses this gap, providing a formal bridge between qualitative preferences and the quantitative belief function framework.

The Axiomatic Framework

The authors consider a finite set of states $S$ and a continuum of consequences $Z$. To model a rational decision-maker’s preference relation ($\succ$), they propose six foundational axioms:

  1. Weak Order: Ensures preferences are consistent and transitive.
  2. Dominance: If act $f$ yields better results than act $g$ across all states, $f$ must be preferred.
  3. Comparative Belief: This axiom is applied to "ideal events"—subsets of states where the agent feels certain enough to behave in a Bayesian manner.
  4. Continuity: Ensures that infinitesimal changes in consequences result in only small changes in preference.
  5. Uniformity: A critical addition that formalizes the "principle of complete ignorance." It ensures that acts with identical best- and worst-case outcomes are treated as equivalent when the underlying states are equally ambiguous.
  6. Independence: A mixture-space axiom that facilitates the creation of a linear utility function and ensures the uniqueness of the resulting belief function.

A central mathematical innovation in this work is the concept of "Ideal Splitting." This technique allows for the decomposition of complex, ambiguous acts into simpler components that can be evaluated against "ideal" (probabilistic) events, serving as a yardstick for uncertainty.

Key Insights and Theoretical Contributions

The paper moves beyond mere formalization to offer several deep insights into the nature of belief-based decision-making:

  • The Bridge of Ideal Events: The authors do not assume global rationality in the Savage sense. Instead, they identify a subset of "ideal events" where the standard Sure-Thing Principle holds. These events provide the necessary structure to quantify uncertainty in non-ideal events.
  • The Neutrality Theorem: A major contribution is the proof that two acts are indifferent if they generate the same belief functions over consequences. While often assumed in prior literature, this "neutrality" is derived here directly from the axioms.
  • Qualitative to Quantitative Transition: The research demonstrates that a qualitative preference over events (e.g., "Event A is more likely than Event B") is sufficient to guarantee the existence of a unique quantitative belief function, provided the axioms are satisfied.
  • Interval Utility: The representation utilizes an "interval utility" $u(\min, \max)$. This captures how a decision-maker weighs the tension between worst-case and best-case scenarios within a set of possible outcomes, offering a more nuanced view than standard expected utility.

The Representation Theorem

The primary result of the paper is Theorem 3.12, the Representation Theorem. It proves that a preference relation satisfies the six axioms if and only if there exists a unique belief function ($Bel$) on the state space and an interval utility function ($u$) (unique up to a positive affine transformation).

Under this model, an agent prefers act $f$ over $g$ if and only if: $$\sum_{A \subseteq S} u(\min f(A), \max f(A)) m_f(A) > \sum_{A \subseteq S} u(\min g(A), \max g(A)) m_g(A)$$ where $m$ represents the "mass" (basic probability assignment) derived from the belief function.

Limitations and Future Directions

While the paper provides a landmark foundation, it notes several areas for expansion. The focus on finite state spaces, while practical for AI, differs from some economic models that require infinite spaces. The authors suggest that their Uniformity axiom effectively replaces the Archimedean axioms used in infinite-space models.

Additionally, the current framework relies on an "inner probability" interpretation of belief functions. Future work may explore Savage-style theorems for "pure" belief functions that do not rely on probabilistic interpretations. Finally, the computational complexity of eliciting these preferences and calculating expectations in large-scale state spaces remains an open question for practical implementation.

Summary updated

Abstract

In this paper, we provide an axiomatic justification for decision making with belief functions by studying the belief-function counterpart of Savage's Theorem where the state space is finite and the consequence set is a continuum [l, M] (l
BibTeX
@inproceedings{zhou2018savage,
  author = {Chunlai Zhou and Biao Qin and Xiaoyong Du},
  title = {{A Savage-style Utility Theory for Belief Functions}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2018},
  url = {https://openreview.net/forum?id=Sk-U6mGdZS},
}

Loading related content...

Figure