TL;DR
This paper proposes a unified axiomatic framework for qualitative decision-making under uncertainty, specifically within the realm of possibility theory. It reconciles two previously distinct approaches (pessimistic and optimistic) by introducing a more general system that can encompass both, offering a more flexible and potentially more applicable model for decision-making when precise probabilities are unavailable.
What Problem Does This Paper Address?
Classical decision theory relies heavily on precise probabilities and utilities to guide choices. However, in many real-world scenarios, especially within Artificial Intelligence, obtaining such precise inputs is difficult or impossible. This has led to the development of "qualitative" decision theories that work with less precise representations of uncertainty, such as possibility theory. The challenge is that different axiomatic approaches to possibility-based decision-making have emerged, each with its own set of assumptions about how a decision-maker should behave (e.g., being risk-averse or risk-seeking). This paper aims to bridge the gap between these disparate approaches by proposing a unified framework that can accommodate them, simplifying the landscape and potentially offering a more robust model for decision-making.
Methods
The core of the paper is the introduction of a new, unified axiomatic system (denoted as 'S') for qualitative decision-making using possibility theory. This system is designed to generalize and subsume two previously proposed systems: the "pessimistic" system (Sp) and the "optimistic" system (So). The unification is achieved through two key modifications:
- Replacing Uncertainty Attitude Axioms: Instead of separate axioms for "uncertainty aversion" (requiring a pessimistic outlook) and "uncertainty attraction" (requiring an optimistic outlook), the unified system uses a single axiom of "qualitative monotonicity" (B2). This axiom focuses on the structure of preferences over lotteries without pre-supposing a specific attitude towards risk.
- Generalizing Continuity Axioms: The continuity axioms from the previous systems (A4 and A4+) are generalized into a single axiom (B4) that is more broadly applicable.
The authors then prove representation theorems for their unified system, showing that it corresponds to a "binary utility scale." This scale uses pairs of numbers to represent utility, allowing for a richer representation than a single numerical utility. They also demonstrate that the previously proposed pessimistic and optimistic systems are indeed special cases of their unified system.
Datasets
Not discussed in the paper. The paper focuses on theoretical axiomatic systems and their mathematical properties, rather than empirical evaluation on datasets.
Key Insights and Ideas
- Unification is Possible: The paper demonstrates that seemingly distinct pessimistic and optimistic decision-making frameworks within possibility theory can be unified under a single, more general axiomatic system. This suggests that the differences might be less about fundamental theory and more about specific parameterizations within a broader structure.
- Focus on Structure, Not Attitude: By replacing specific "uncertainty attitude" axioms with a more general "monotonicity" axiom, the unified system shifts the focus from pre-defined psychological stances (optimism/pessimism) to the structural properties of preferences. This is argued to be more aligned with how decision theory should operate, separating the core decision framework from individual utility assessments.
- Binary Utility as a Natural Fit: The representation theorem for the unified system leads to a "binary utility scale." This suggests that representing utility as a pair of values might be a more natural and expressive way to capture decision-making under possibility theory, where uncertainty is inherently described by two related measures (like possibility and necessity, or upper and lower bounds).
- Simplification and Generalization: The unified system simplifies the axiomatic landscape and offers a more expressive power, capable of handling situations where the previous systems might struggle (e.g., when lotteries have equivalent standard lotteries in both "pessimistic" and "optimistic" halves).
Key Results
- A new axiomatic system (S) for qualitative decision-making with possibility theory is proposed.
- This system unifies the previously proposed pessimistic (Sp) and optimistic (So) axiomatic systems, showing them to be special cases of S.
- A representation theorem is proven for S, demonstrating that it corresponds to a decision-maker whose preferences can be represented by a utility function mapping to an ordered binary utility scale.
- The paper argues that the binary utility scale is a more appropriate representation for possibility-based decision-making than single numerical utilities.
Limitations and Future Work
- Counterintuitive Axioms: The authors acknowledge that some of the axioms in the original pessimistic and optimistic systems (specifically A4 and A4+) can lead to counterintuitive results when considered in isolation. Their unified system, by modifying these axioms, aims to resolve these issues.
- Application Complexity: While the unified system is theoretically more general, the practical application of the binary utility scale might still present challenges for decision-makers in terms of eliciting and interpreting these paired utility values.
- Future Work: The paper suggests that further exploration of the implications of the binary utility scale and its comparison with other uncertainty formalisms would be valuable. They also hint at the potential for their framework to address situations where neither purely pessimistic nor purely optimistic approaches are sufficient.
Summary updated