TL;DR
This paper explores decision-making under uncertainty using only ordinal preferences and uncertainty relations, avoiding numerical representations. They propose a method that, while preserving some core decision theory axioms, reveals fundamental inconsistencies with probabilistic uncertainty, leading to potentially risky or indecisive decision-making.
What Problem Does This Paper Address?
Classical decision theory, notably Savage's framework, assumes decision-makers can quantify their uncertainty (via probabilities) and preferences (via utilities). This paper tackles the more realistic scenario where decision-makers can only express "at least as likely as" or "at least as preferred as." The challenge is to derive a coherent decision-making strategy from these qualitative, ordinal statements, without resorting to numerical scales. This is particularly relevant because eliciting precise numerical values can be difficult or impossible for individuals.
Methods
The authors propose a method to derive a preference relation on acts (sequences of consequences given states of the world) from an ordering on events (representing uncertainty) and an ordering on consequences (representing preferences). Their core technique involves a "lifting" procedure: an act $f$ is considered at least as good as act $g$ if the set of states where $f$ is preferred to $g$ is at least as likely as the set of states where $g$ is preferred to $f$. They explore this using qualitative probability and, more successfully, with possibility theory (specifically, necessity orderings). They also investigate how this approach aligns with or deviates from Savage's axioms, particularly the "sure thing principle."
Datasets
Not discussed in the paper. The paper is theoretical and uses examples to illustrate concepts rather than empirical datasets.
Key Insights and Ideas
- Ordinality is Tricky: While intuitively appealing, building a decision theory purely on ordinal preferences and uncertainty is fraught with difficulties. It doesn't naturally align with probabilistic representations.
- The "Sure Thing Principle" is Preserved (Mostly): The paper shows that a key axiom of Savage's theory, the "sure thing principle," can be maintained even without numerical representations. This is a significant finding for qualitative decision theory.
- Indifference is Not Transitive: A major consequence of their approach is that the indifference relation between acts (where one act is neither strictly preferred nor strictly dispreferred to another) is not necessarily transitive. This is analogous to the Condorcet paradox in social choice theory.
- Connection to Non-Monotonic Reasoning: The derived uncertainty relations bear a strong resemblance to those found in non-monotonic reasoning (like preferential inference), suggesting a cautious, default-based approach to uncertainty.
- The "Lifting Axiom" is Problematic: The core mechanism for combining uncertainty and preference (the lifting axiom) leads to either overly risky decisions (assuming the most favorable state of the world) or indecisiveness, making it difficult to recommend a prudent course of action.
Key Results
- The proposed method can derive a preference relation on acts from ordinal uncertainty and preference relations.
- The approach is compatible with Savage's "sure thing principle" and can be formulated using possibility theory.
- A key finding is that the indifference relation between acts is not transitive under this ordinal framework, a departure from standard decision theory.
- The decision guidelines derived can be either overly risky (e.g., always assuming the best-case scenario) or too indecisive, making them impractical for many real-world scenarios.
Limitations and Future Work
The authors acknowledge that their approach, particularly due to the "lifting axiom," leads to decision guidelines that are often "very risky" or "very little decisive." They suggest that this might be an inherent limitation of purely symbolic approaches to decision-making under uncertainty.
Future work could explore alternative ways to combine uncertainty and preference, perhaps by reversing the roles of states and consequences or by explicitly modeling attitudes towards risk (e.g., pessimism or optimism) through different aggregation rules (min/max over consequences). The paper poses an open question about finding an alternative to the lifting axiom that would yield both cautious and decisive decision-making without relying on numerical representations.
Summary updated