A Unified Setting for Inference and Decision: An Argumentation-based Approach

Leila AmgoudUAI 2005

TL;DR

This paper proposes a novel argumentation framework that unifies logical inference and decision-making. By extending existing argumentation systems, it allows for reasoning with inconsistent information and provides a structured way to compare decisions based on the strength of supporting arguments. This approach offers a more human-like deliberation process for complex decision problems.

What Problem Does This Paper Address?

Traditionally, artificial intelligence research has treated logical inference (drawing conclusions from information) and decision-making (choosing among options) as separate problems, leading to distinct models for each. This separation can be limiting, as real-world decision-making often involves inferring conclusions from potentially inconsistent information and then using these inferences to evaluate choices. Existing decision-making models often assume coherent information, which is not always realistic. This paper aims to bridge this gap by presenting a single, unified framework that can handle both inference from inconsistent knowledge bases and various types of decision-making, including those under uncertainty, with multiple criteria, and rule-based decisions.

Methods

The core of the paper is the development of a general argumentation framework. This framework builds upon the idea of constructing arguments for and against statements or decisions. Key components and steps include:

  • Argument Construction: Arguments are built from a knowledge base (beliefs) and goals (what an agent wants to achieve or avoid). These arguments can be:
  • Epistemic: Supporting beliefs.
  • Recommending: Deductively deriving a decision from rules.
  • Decision Arguments: Abductively supporting a decision by linking it to goal satisfaction or violation.
  • Argument Comparison: Arguments are assigned a "force" based on two factors:
  • Certainty Level (Cert): How certain the beliefs used in the argument are.
  • Importance Degree (Imp): How important the goals satisfied or violated by the argument are.
  • Conflict Resolution: The framework defines various types of conflicts between arguments (rebutting, assumption-attacking, undercutting) and a unified notion of "defeat" to resolve them.
  • Acceptability Semantics: Standard argumentation semantics (admissible, preferred, complete, stable, grounded) are used to determine which arguments are considered acceptable in the face of conflicts.
  • Inference and Decision:
    • Inference: Conclusions are inferred if they are supported by acceptable epistemic arguments.
  • Decision: Decisions are compared based on the strength of their supporting (pro) and opposing (con) acceptable arguments, taking into account both certainty and importance.

A crucial aspect is that the framework extends classical decision theory by not requiring the information about the environment to be coherent.

Datasets

Not discussed in the paper.

Key Insights and Ideas

  • Unification of Inference and Decision: The most significant insight is that inference is not a separate problem but an integral part of decision-making. By inferring justified conclusions first, we can then use these to build stronger arguments for or against decisions.
  • Argument Force as a Dual Measure: The concept of argument "force" being a combination of belief certainty and goal importance is a powerful way to capture the nuanced evaluation humans use when making choices.
  • Handling Inconsistency: The framework explicitly allows for and reasons with inconsistent information, which is a more realistic representation of many real-world scenarios.
  • Human-like Deliberation: The argumentation-based approach mimics how humans often deliberate, weighing pros and cons, which can lead to more transparent and understandable decision support systems.
  • Generalizability: The framework is designed to be general enough to encompass various decision-making scenarios, from simple rule-based choices to complex multi-criteria problems under uncertainty.

Key Results

The paper presents formal definitions for argument structures, comparison criteria, conflict relations, and acceptability semantics within its unified framework. It demonstrates that:

  • The framework can capture different types of decision problems.
  • The set of conclusions inferred from a knowledge base within this framework is consistent.
  • An argumentation framework considering only epistemic arguments is equivalent to one considering all argument types (epistemic and decision-related) in terms of inference results.
  • A principle for comparing decisions is proposed, which prioritizes decisions supported by stronger arguments (considering both certainty and importance).

Limitations and Future Work

The authors acknowledge that the proposed criteria for comparing arguments and decisions are intuitive and may be subject to discussion, suggesting that further refinement might be needed. They also mention that the framework can be applied to various decision problems, implying that future work could involve instantiating and testing the framework on specific real-world applications such as multi-criteria decision problems or rule-based decision-making. The paper also notes that the comparison of arguments is not exhaustive and can be refined.

Summary updated

BibTeX
@inproceedings{amgoud2005unified,
  author = {Leila Amgoud},
  title = {{A Unified Setting for Inference and Decision: An Argumentation-based Approach}},
  booktitle = {Conference on Uncertainty in Artificial Intelligence},
  year = {2005},
}
Figure