Qualitative Decision Theory

Sek-Wah Tan, Judea PearlAAAI 1994

Abstract

Traditional decision theory relies on precise numerical probabilities and utilities, which are often unavailable in complex human environments. This paper presents a qualitative framework for decision-making that utilizes conditional desires and normality defaults. By refining "Ceteris Paribus" semantics, the system handles specific exceptions to general rules and accounts for low-probability, high-impact events. We demonstrate the framework's efficacy through classic reasoning problems, showing it can justify risk-averse behavior and resolve conflicting preferences.

Introduction

Rational choice is typically modeled using complete probability distributions and precise utility functions. However, humans communicate desires and beliefs qualitatively: "I prefer coffee over tea in the morning" or "If it's cloudy, it usually rains." This paper, presented by Sek-Wah Tan and Judea Pearl, addresses two primary challenges in formalizing such qualitative reasoning: handling specificity (where specific rules override general ones) and the "surprise" problem (where unlikely but critical events are ignored by naive qualitative models).

Qualitative Reasoning Framework

The proposed system processes two primary inputs: Normality Defaults (beliefs) and Conditional Desires (preferences).

System architecture for qualitative decision reasoning.

Instead of continuous values, worlds are assigned integer ranks. Preference Rankings ($\pi$) represent desirability, while Belief Rankings ($\kappa$) represent "surprisingness" (where a rank of 0 indicates a normal world). The system employs a "principle of maximal indifference" to select the most compact ranking ($\pi^+$) that satisfies all given constraints, avoiding unjustified assumptions about the world's state.

Specificity and CP-Semantics

A core contribution of this work is the refinement of "Ceteris Paribus" (CP) semantics to resolve contradictions between general and specific desires. For instance, a general desire to stay alive might conflict with a specific willingness to sacrifice oneself for a noble cause.

Visualizing specificity through context subsets in CP-semantics.

By defining specificity through the "context" of a desire—the set of worlds it actually constrains—the logic ensures that the more specific context (e.g., $W \subset V$) takes precedence. Additionally, the authors introduce $\delta$-believability to solve the "surprise" problem. By evaluating preferences across all worlds up to a certain "surprisingness" threshold, the system can justify risk-averse behaviors, such as purchasing fire insurance despite the low statistical probability of a fire.

Results: The Umbrella and Insurance Problems

The framework was validated using classic AI reasoning problems. In the "Umbrella Example," the system correctly infers that an agent should carry an umbrella when cloudy if the desire to stay dry outweighs the inconvenience of the umbrella. In the "Fire Insurance Example," it successfully justifies insurance by considering the high negative utility of a fire in "surprising" worlds, which previous qualitative models often ignored.

Conclusion

Tan and Pearl successfully demonstrate a logic that can infer non-obvious choices and justify risk-averse behavior using purely qualitative inputs. While the logic is sound, future work must address the computational complexity of calculating these rankings in large-scale systems and explore how these integer-based rankings align with nuanced human linguistic expressions like "ideally" or "strongly prefer."

Summary updated

BibTeX
@inproceedings{tan1994qualitative,
  author = {Sek-Wah Tan and Judea Pearl},
  title = {{Qualitative Decision Theory}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1994},
}
Figure