Introduction
In the field of Multiple Criteria Decision Aid (MCDA), the goal is to assist decision-makers (DMs) in choosing between alternatives based on multiple, often conflicting, criteria. A significant challenge in this process is the inherent logical inconsistency of human preferences. When a DM provides rankings that violate the mathematical assumptions of a chosen model, the system often fails. This paper, presented at EUSFLAT-LFA 2011, introduces an algorithmic framework to identify and resolve these inconsistencies specifically for the 2-additive Choquet integral model.
The 2-additive Choquet Integral and Consistency
The Choquet integral is a sophisticated tool for modeling interactions between criteria, such as redundancy or synergy. To maintain computational feasibility, researchers often use the 2-additive version, which restricts interactions to pairs of criteria. However, for a 2-additive Choquet integral to represent a DM's preferences, two fundamental conditions must be met:
- Absence of Strict Cycles: The DM cannot prefer A over B, B over C, and C over A.
- The MOPI Property: The Monotonicity of Preferential Information (MOPI) ensures that the criteria weights remain logically consistent when interactions are factored in.
In practice, DMs frequently provide ordinal information (rankings) that violates these rules. Rather than simply rejecting the input, the authors propose a "recommendation engine" to guide the DM toward a consistent model.
The Proposed Algorithm
The authors' algorithm utilizes a graph-based representation of preferences. By defining "binary actions"—fictitious alternatives where criteria are either "satisfactory" (1) or "neutral" (0)—the elicitation process is simplified.
The algorithm follows these key steps:
- Graph Construction: An oriented graph is built where vertices represent binary actions and edges represent the DM's preferences (strict preference, indifference, or natural monotonicity).
- Cycle Detection: Using Strongly Connected Components (SCC) analysis, the algorithm identifies cycles. If a strict preference exists within a cycle, the model is flagged as inconsistent.
- MOPI Validation: The system checks for violations of the MOPI property, which indicates a logic contradicting the 2-additive structure.
- Interactive Resolution: Instead of automated overrides, the algorithm suggests minimal adjustments. It typically recommends changing a problematic strict preference to "indifference" or declaring items "incomparable."
A critical theoretical contribution is the proof that these recommendations are "safe." The authors demonstrate that by following the suggested fix of replacing strict preference with indifference, the DM will not inadvertently trigger new MOPI violations elsewhere in the model—a common "whack-a-mole" problem in decision support systems.
Case Study: Digital Camera Selection
To demonstrate the algorithm's utility, the paper employs a toy example involving the purchase of a digital camera. The DM evaluates six cameras based on three criteria: Resolution, Price, and Zoom.
In this scenario, the DM's initial preferences might lead to a cycle or a MOPI violation. For instance, a DM might value high resolution and low price but provide rankings that imply an interaction effect impossible to represent with a 2-additive model. The algorithm identifies the specific pair of preferences causing the conflict and presents the DM with the minimal changes required to achieve a valid Choquet representation.
Implementation and Results
The algorithm was implemented in a JAVA-based software tool called MYRIAD, developed at THALES. This implementation proves the practical utility of the approach in industrial decision-support contexts. The results show that the number of inconsistencies reduces gradually as the DM accepts the suggested fixes, eventually converging on a representable model that respects the DM's core intent while satisfying mathematical constraints.
Limitations and Future Work
While effective, the current approach is limited to 2-additive models. Future research could extend this to $k$-additive models involving more complex interactions. Additionally, while the current "indifference" recommendations are safe, more sophisticated "repairs"—such as suggesting a reversal of preference direction—could be explored. Finally, formal user studies are needed to evaluate how DMs interact with these logical "corrections" in real-world scenarios.
Summary updated