A Hybrid Approach to Reasoning with Partially Elicited Preference Models

Vu A. Ha, Peter HaddawyUAI 1999

TL;DR

This paper proposes a hybrid approach to automated decision-making by combining classical multi-attribute utility theory with qualitative preference elicitation methods. It tackles the challenge of high elicitation costs in decision theory by developing techniques to reason with partially specified utility functions, enabling the identification of suboptimal alternatives without requiring a complete preference model.

What Problem Does This Paper Address?

Automated decision-making systems often rely on utility theory to model preferences. However, a major hurdle is the "elicitation cost" – the significant effort required to fully specify a user's utility function. This is impractical for complex decisions and can be undesirable as preferences can change. Existing qualitative approaches offer expressiveness but lack the inferential power for practical systems. This paper bridges this gap by proposing a method that leverages both the quantitative rigor of utility theory and the expressive power of qualitative preference statements to make decision support more efficient and flexible.

Methods

The core of the paper is a hybrid approach that assumes a partially specified multi-linear utility function (MLUF). The authors introduce a framework that uses polyhedral cones to represent constraints on the unknown scaling coefficients of the MLUF. These constraints are derived from two sources:

  1. Utility Independence (UI): Mathematical conditions derived from the assumption that certain attributes are utility independent of others. This translates into linear, homogeneous constraints on the scaling coefficients.
  2. Pairwise Comparisons: Explicit statements from the decision maker comparing two decision consequences (or prospects). These comparisons are translated into further linear constraints on the scaling coefficients.

By combining these constraints, the authors define a "preference cone" representing the set of admissible scaling coefficients. They then develop algorithms to:

  • Test for Induced Dominance: Determine if one decision alternative is strictly preferred to another across all possible admissible utility functions.
  • Test for Potential Optimality: Identify alternatives that could be optimal under some admissible utility function.
  • Detect Inconsistency: Identify conflicting preference statements that lead to an empty preference cone.

The paper also introduces the concept of "ceteris paribus" (all else being equal) comparative statements from qualitative decision theory, showing how they can be integrated to generate a large number of pairwise preferences efficiently.

Datasets

Not discussed in the paper.

Key Insights and Ideas

  • Synergistic Integration: The paper demonstrates that quantitative (utility theory) and qualitative (preference statements) approaches can be effectively combined, leveraging the strengths of each to overcome the limitations of using either in isolation.
  • Preference Cone as a Constraint Representation: Using polyhedral cones to represent the space of admissible utility functions (defined by scaling coefficients) provides a powerful mathematical framework for reasoning about partial preferences.
  • Efficiency through Partial Elicitation: By focusing on identifying suboptimal alternatives rather than fully specifying the utility function, the approach significantly reduces the elicitation burden on the decision maker.
  • Qualitative Statements as Powerful Constraints: "Ceteris paribus" statements, which are more natural for humans to express, can implicitly encode a large number of pairwise comparisons, leading to more efficient inference.

Key Results

The paper presents theoretical results and algorithms for:

  • Induced Dominance: A method to deduce that one alternative is strictly preferred to another, even with partial preference information.
  • Potential Optimality: A method to identify alternatives that are potentially optimal.
  • Inconsistency Detection: The framework can reveal inconsistencies in the decision maker's stated preferences.

An example is provided for a three-attribute scenario, illustrating how utility independence constraints and pairwise comparisons translate into inequalities that define the preference cone and how these can be used to infer relationships between alternatives.

Limitations and Future Work

  • Computational Complexity: The authors acknowledge that the complexity of their algorithms, particularly for determining the generators of the polyhedral cone, can be exponential in the number of attributes. They suggest exploring qualitative logical constructs and exploiting further utility independencies to mitigate this.
  • Proactive Elicitation: The current framework is reactive, responding to user-provided preferences. Future work could involve developing systems that proactively ask targeted questions to gain the most informative preference data.
  • Case-Based Integration: The authors suggest integrating their approach with case-based reasoning frameworks for preference elicitation, where existing user models can inform initial preference representations.
  • Handling Inconsistency: While inconsistency detection is a result, identifying which comparative statements are problematic and how to resolve them remains an open research question.

Summary updated

BibTeX
@inproceedings{ha1999hybrid,
  author = {Vu A. Ha and Peter Haddawy},
  title = {{A Hybrid Approach to Reasoning with Partially Elicited Preference Models}},
  booktitle = {Conference on Uncertainty in Artificial Intelligence},
  year = {1999},
}
Figure