Multi-Attribute Decision Making using Weighted Description Logics

Erman Acar, Christian MeilickePRUV 2014 · OpenReview

Introduction

Classical decision theory, particularly Multi-Attribute Utility Theory (MAUT), provides a robust framework for choosing among alternatives by evaluating their various attributes. However, a significant limitation of traditional models is their inability to account for the semantic relationships between these attributes. In many real-world scenarios, attributes are not independent; they are linked by background knowledge and terminological definitions. For instance, in a vehicle purchase scenario, the classification of a car as a "Roadster" inherently implies certain characteristics like "High Horsepower" or "Prestigious" status.

This paper introduces a formal framework that bridges the gap between raw data and subjective value by integrating Description Logics (DL) with weighted utility functions. By leveraging the reasoning capabilities of DL, the proposed system can automatically infer the utility of an alternative based on its logical implications within a comprehensive knowledge base.

The Decision Base Framework

The core of the proposed approach is the Decision Base ($D$), defined as a triple $(K, \mathcal{C}, \mathcal{U})$:

  • Background Knowledge ($K$): This consists of a DL knowledge base, including a TBox (terminology) defining the hierarchy of concepts and an ABox containing general facts about the domain.
  • Choice Box ($\mathcal{C}$): This represents the set of available alternatives. Each alternative is modeled as an individual ABox, containing specific assertions that describe the choice.
  • Utility Box ($\mathcal{U}$): The UBox contains "utility assertions"—pairs consisting of a logical formula and a numerical weight representing the utility gained if that formula is satisfied.

Reasoning and Utility Computation

The utility of a specific choice $C \in \mathcal{C}$ is determined by the sum of the weights of all assertions in the Utility Box that are logically entailed by the combination of the choice and the background knowledge ($K \cup C \models a_i$).

This mechanism relies on standard DL reasoning services, such as instance checking and consistency checking. A critical feature of this framework is its handling of constraints: if a choice $C$ is logically inconsistent with the background knowledge $K$, its utility is considered undefined. This effectively filters out impossible or contradictory alternatives (e.g., a car that is simultaneously "Inexpensive" and "Luxury" if the TBox defines these as disjoint). To handle numerical attributes like price or fuel efficiency, the authors incorporate concrete domains into the logic.

Illustrative Example: The Car Buyer

To demonstrate the framework, the authors present a qualitative use case involving a car buyer choosing between a Mazda Mx-5 and a BMW 335i.

In this scenario, the TBox defines categories (e.g., a "Roadster" is a "Sporty Car") while the UBox reflects the buyer's preferences, such as high weights for "Inexpensive," "Economic," and "Roadster" status. The results of the logical inference were as follows:

  • The Mazda Mx-5 achieved a utility of 220, benefiting from being logically classified as "Inexpensive," "Economic," and a "Roadster."
  • The BMW 335i achieved a utility of 170. While it gained points for being "Prestigious" and "Very Strong," it failed to meet the "Inexpensive" and "Economic" criteria.

Despite the BMW's superior performance metrics, the framework identifies the Mazda as the rational choice based on the buyer's specific semantic preferences and the logical consequences of the cars' attributes.

Theoretical Implications and Insights

The framework offers several key insights into logic-based decision making:

  • Implicit Preference Matching: Users do not need to explicitly map every attribute to a utility value. If a user values "Safety" and the ontology defines "Side Airbags" as a safety feature, the system automatically assigns utility to any car equipped with them.
  • Expressive Flexibility: The authors prove that the framework is highly flexible and can represent non-monotonic, non-concave, and non-additive utility functions depending on how the UBox is structured.
  • Value of Information: The paper formalizes the "Value of Information" as the change in total utility resulting from the addition of a new logical axiom. This allows agents to quantify the benefit of performing additional reasoning or data acquisition.

Conclusion and Future Work

The integration of Description Logics into multi-attribute decision making provides a powerful tool for handling complex, semantically-rich environments. While the current model is deterministic, it lays the groundwork for more advanced systems.

Future research directions include extending the framework to handle uncertainty via Probabilistic Description Logics, exploring multi-agent collaborative decision-making, and developing practical tools like a plugin for the Protégé ontology editor to facilitate real-world implementation.

Summary updated

BibTeX
@inproceedings{acar2014multi,
  author = {Erman Acar and Christian Meilicke},
  title = {{Multi-Attribute Decision Making using Weighted Description Logics}},
  booktitle = {PRUV},
  year = {2014},
  url = {https://openreview.net/forum?id=toTVlc5dLse},
}
Figure