A Logic of Deliberation

Marvin BelzerAAAI 1986

Abstract

This paper, originally presented at AAAI-86, establishes a foundational formal framework for practical reasoning—the cognitive process of transforming values and beliefs into actionable intentions. By extending deontic logic to incorporate defeasible rules, the author provides a mathematical structure for AI agents to resolve conflicting obligations and maintain internal consistency in complex, non-monotonic environments.

Introduction

In the mid-1980s, a significant challenge in artificial intelligence was the development of "normative reasoners": systems capable of adhering to rules, policies, and strategic guidelines. Traditional logical systems struggled with the reality that real-world rules are rarely absolute. Most norms are prima facie (valid at first glance) but subject to being overridden by specific exceptions or competing duties. This paper addresses the need for a "Logic of Deliberation" capable of distinguishing between initial duties and "all-things-considered" commitments, while modeling an agent's internal rationality.

The Problem of Defeasible Rules

The core difficulty in normative reasoning lies in the non-monotonic nature of real-world information. Consider a scenario involving two rules:

  1. One should not disclose a secret to Reagan.
  2. One should disclose a secret to Reagan if Gorbachev is also informed.

In standard monotonic logic, if both conditions are met (Gorbachev is told), the system may encounter a contradiction between the prohibition and the obligation. The author argues that a robust logic must handle such exceptions by allowing new information to cancel previous conclusions without collapsing the entire logical framework.

Methodology: The 3-D Logical Framework

The author develops three interlocking logical systems to bridge the gap between knowledge and action:

1. Logic 3-D (Deontic)

This system introduces two distinct "ought" operators: $O(Q/P)$, representing a prima facie obligation (it ought to be $Q$ given $P$), and $!(Q/P)$, representing an all-things-considered commitment. A critical innovation here is the "U" operator, which functions as an "other things being equal" clause. This allows the system to "detach" a final commitment from a general rule provided no relevant exceptions are identified.

2. Logic 3-Db (Belief)

The same structural principles are applied to belief systems. Logic 3-Db distinguishes between "expectations" (defeasible rules of thumb, such as "birds fly") and "certainties" (verified information). This allows the agent to reason about the world using generalities while remaining flexible to specific evidence.

3. Logic 3-Dpr (Practical Reasoning)

This system represents the synthesis of the previous two. It embeds the belief system within the normative system, deriving an agent's final plan ($!Q$) from their values ($O$) filtered through their subjective certainties and expectations. The semantics are grounded in possible-worlds modeling, where potential histories are ranked based on their compliance with the agent's values and expectations.

Key Insights and Results

Formalization of Deliberation

The paper successfully defines a set of axioms governing the transition from desires and beliefs to settled intentions. It provides a formal definition for the "relative weight" of rules, suggesting that one rule takes precedence over another if it prevails in a direct conflict. This allows for qualitative decision-making without the necessity of assigning arbitrary numerical utility scores.

Conflict Resolution

The framework demonstrates that classic logical paradoxes, such as the "Nixon Secret" or the "Quaker Pacifist" problem, can be resolved. In the Nixon scenario (where Nixon is both a Quaker/pacifist and a Republican/non-pacifist), the logic allows the agent to hold general expectations while concluding a specific commitment once more specific conditions are met. This mirrors human-like reasoning where "all-things-considered" commitments are non-monotonic.

Internal Consistency

A primary contribution is the definition of a "well-balanced" agent. In this logic, an agent is not judged by the objective "correctness" of its values, but by whether its plans are logically consistent with its own subjective beliefs and normative priorities.

Limitations and Future Directions

While foundational, the paper identifies several areas for further research:

  • Proof Theory: At the time of publication, a complete proof theory for the 3-D logic—a set of rules to derive all possible truths—remained an open question.
  • The Omniscience Problem: The logic deals with implicit beliefs (what follows logically from knowledge), but does not account for the fact that real agents have limited computational resources and may not realize every implication of their beliefs.
  • Computational Complexity: The paper focuses on formal expression rather than the computational cost of ranking possible worlds in real-time AI systems.
  • Quantitative Integration: The author suggests that future work could explore how these qualitative logical structures might interface with Bayesian probabilities or numerical utility functions to create more hybrid reasoning systems.

Summary updated

BibTeX
@inproceedings{belzer1986logic,
  author = {Marvin Belzer},
  title = {{A Logic of Deliberation}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1986},
}
Figure