Generalized Qualitative Probability: Savage revisited

Daniel LehmannUAI 1996

TL;DR

This paper revisits Savage's foundational work on decision theory under uncertainty, proposing a weaker set of postulates that relax the requirement of complete ordering of preferences. This generalization allows for incomparabilities, better reflecting real-world indecision, and leads to a characterization of "generalized qualitative probabilities" that bridges traditional probability and logical approaches.

What Problem Does This Paper Address?

L. Savage's seminal work in "The Foundations of Statistics" established a framework where rational decision-making under uncertainty is equivalent to maximizing expected utility. This implies that individuals act as if they assign subjective probabilities to events and possess utility functions for outcomes. However, several aspects of Savage's postulates have been challenged:

  • Completeness of Preferences: Savage's first postulate requires preferences to form a "simple order," meaning any two acts must be comparable (either one is preferred to the other, or they are indifferent). This can be unrealistic, as individuals often experience genuine indecision or incomparability between choices.
  • Subjective Probabilities as Probabilities: Some approaches, like Dempster-Shafer, question whether subjective beliefs must strictly adhere to the laws of probability calculus.
  • Null Events and Conditioning: Savage's framework requires events to have positive probability for conditioning. This can be problematic for events that are "null" relative to other null events.
  • Strong Axioms of Choice: Certain postulates, like P4, imply very strong assumptions about how preferences over consequences interact with events, which may not always hold in practice.

This paper aims to address these criticisms by weakening Savage's postulates, particularly the completeness assumption, while still achieving a characterization of subjective probabilities.

Methods

The authors propose a new set of postulates (Q0-Q6) for preferences over acts, which are weaker than Savage's original set. Key methodological innovations include:

  • Partial Pre-orders: Instead of a total pre-order, preferences on any given event A are allowed to be a partial pre-order (postulates Q0 and Q1). This means some acts might be incomparable given event A.
  • Event-Conditional Preferences: The core of the new framework is the introduction of preference relations indexed by events (e.g., $f \le_A g$). This allows for a richer representation of preferences that depend on the specific circumstances (events) under consideration.
  • Weaker Axioms: Postulates like Q2 (equivalence on A implies indifference on A) and Q3/Q4 (axioms relating preferences on disjoint events and their unions) are introduced to capture more nuanced relationships between preferences and events.
  • Generalized Qualitative Probability: The paper defines a new concept, "generalized qualitative probability" (g.q.p.), based on the derived preference relation on events. This g.q.p. relation is shown to be reflexive and transitive, generalizing Savage's definition.

Datasets

Not discussed in the paper.

Key Insights and Ideas

  • Embracing Indecision: The most significant insight is that relaxing the completeness axiom (allowing for incomparable acts) doesn't necessarily lead to a "blind alley" but can provide a more realistic model of decision-making. This directly addresses the intuition that sometimes we are genuinely undecided.
  • Event-Centric Preferences: By making preferences explicitly conditional on events ($f \le_A g$), the framework captures how uncertainty about specific events can influence choices, even when the underlying consequences are the same.
  • Bridging Probability and Logic: The derived "generalized qualitative probability" relation on events offers a way to compare the plausibility of events without strictly adhering to the numerical calculus of probabilities, potentially bridging gaps between probabilistic and logical reasoning approaches to uncertainty.
  • Sure Thing Principle Derivation: The paper demonstrates that Savage's Sure Thing Principle (P2) can still be derived even with the weaker postulates, highlighting its robustness.

Key Results

The paper successfully weakens Savage's postulates and derives a characterization of "generalized qualitative probabilities." The core results include:

  • Postulates Q0-Q6: A new set of postulates is introduced that are weaker than Savage's, particularly by allowing for partial pre-orders of preferences on events.
  • Derivation of Sure Thing Principle: Savage's Sure Thing Principle (P2) is shown to be derivable from the new set of postulates.
  • Definition of Generalized Qualitative Probability (g.q.p.): A formal definition of g.q.p. is provided, which is a reflexive and transitive relation on events that generalizes Savage's qualitative probability.
  • Consistency Proof: Theorem 1 establishes that any g.q.p. can be generated by a preference structure satisfying the proposed postulates (Q1-Q6), demonstrating the consistency and completeness of the framework.

Limitations and Future Work

  • Completeness of the Framework: While the paper shows that the proposed postulates are consistent and lead to a g.q.p., the authors acknowledge that further work is needed to fully explore the utility theory for these generalized structures.
  • Conjecture on Equivalence of Acts: The authors conjecture that a specific property (Theorem 5.2.1 analogue) related to acts taking the same values on equivalent events is not implied by their postulates alone, suggesting a potential avenue for future research or the need for an additional postulate.
  • Search for Uniform Generation: An open problem is identified: finding a uniform way to generate all generalized qualitative probabilities.
  • Utility Theory on G.Q.P.: The paper suggests that developing utility theory on these generalized structures is a promising direction for future work.
  • Non-standard Models: The paper hints at the existence of non-standard probability measures that can generate these g.q.p. structures, which could be a fruitful area for further investigation.

Summary updated

BibTeX
@inproceedings{lehmann1996generalized,
  author = {Daniel Lehmann},
  title = {{Generalized Qualitative Probability: Savage revisited}},
  booktitle = {Conference on Uncertainty in Artificial Intelligence},
  year = {1996},
}
Figure