A Modal Logic for Beliefs and Pro Attitudes

Kaile Su, Abdul Sattar, Han Lin, Mark ReynoldsAAAI 2007

Abstract

This paper presents a novel non-normal modal logic designed to formalize "pro attitudes"—such as goals, desires, and intentions—within rational agents. By transitioning from standard Kripke structures to linear neighborhood frames, the authors address the long-standing "side-effect problem" and establish a framework that is computationally grounded in probabilistic programming and utility theory. The resulting BPA (Beliefs and Pro Attitudes) logic is shown to be sound, complete, and philosophically aligned with established requirements for rational agency.

1. Introduction

In the development of artificial intelligence agent theory, a fundamental distinction is made between informational attitudes (what an agent believes to be true) and pro attitudes (what an agent desires or intends to achieve). While epistemic logics for knowledge and belief are well-matured, formalizing pro attitudes has remained a significant challenge. Traditional approaches often fall into logical traps that do not reflect human-like or rational reasoning, or they remain too abstract for practical computational implementation. This paper introduces a hybrid neighborhood model to bridge the gap between high-level philosophical logic and concrete utility-based computation.

2. The Side-Effect Problem

The primary hurdle in formalizing pro attitudes within "normal" modal logics (those based on Kripke structures) is the side-effect problem. In a normal modal logic, if an agent desires a state $p$, and $p$ logically entails $q$, the agent is mathematically required to desire $q$.

This leads to counter-intuitive results in agent modeling. For instance, if an agent desires to be cured of a specific disease, and being cured logically implies having had the disease in the first place, a normal logic would conclude that the agent desires to have the disease. To build realistic agents, a logic must be "non-normal," allowing an agent to desire a specific outcome without being forced to desire every trivial or necessary logical consequence of that outcome.

3. Methodology: Linear Neighborhood Frames

To solve these issues, the authors propose a hybrid neighborhood model. Unlike standard Kripke models that utilize binary accessibility relations between possible worlds, neighborhood semantics use a function that maps each world to a collection of sets of worlds, known as neighborhoods.

3.1 Linearity and Utility

A key innovation in this framework is the introduction of linearity to the neighborhood functions. This mathematical property mirrors the behavior of probability and utility expectation. Specifically, it allows the "goodness" or utility of disjoint events to be additive. This property is essential for grounding the logic in quantitative models, as it aligns the abstract logical operators with the way agents calculate expected value in probabilistic environments.

3.2 The BPA Axiom System

The authors develop the BPA (Beliefs and Pro Attitudes) axiom system and provide a formal proof of its soundness and completeness. This ensures that every formula provable within the system is true within the model, and conversely, every true property of the model can be derived through the proof theory. This mathematical solidity is a prerequisite for using the logic in formal verification and reliable agent design.

4. Computational Grounding

One of the paper's most significant contributions is the semantic derivation of the logic from a concrete model of probabilistic programs with utilities. The authors define a modality $H\phi$ (representing that the agent is "happy" or satisfied if told $\phi$). This is not treated as a vague psychological state but is defined by the expected utility an agent gains when informed that a certain set of execution runs is possible. By linking $H\phi$ to computable expectations, the authors move pro attitudes into a domain where they can be algorithmically evaluated.

5. Evaluation and Insights

As a theoretical work, the paper evaluates its logic through rigorous proof and philosophical thought experiments rather than empirical datasets.

  • The Lectureship Scenario: By analyzing a case involving three job applicants, the authors demonstrate why desiring two separate outcomes (e.g., Alice getting the job or Bob getting the job) should not necessarily force a desire for their disjunction in all contexts.
  • The Vegetarian Dinner Example: This scenario illustrates how agents process "good news" versus "bad news" based on their existing beliefs and utility functions.

Key Findings

  • Side-Effect Freedom: The BPA logic is more robustly side-effect free than previous frameworks, such as those proposed by Cohen & Levesque.
  • Bratman’s Requirements: The logic satisfies the "Asymmetry Thesis," which posits that a rational agent should not desire what they believe to be impossible.
  • Introspectivity: The model assumes agents possess self-awareness regarding their pro attitudes; if an agent desires a state, they believe that they desire it.

6. Conclusion and Future Work

The BPA logic successfully bridges the gap between abstract modal logic and the practicalities of probabilistic programming. However, several avenues for future research remain. The current model is limited to a single-agent perspective and is static, meaning it does not yet account for how desires evolve over time or change in response to environmental actions. Future iterations could refine the $H$ modality to distinguish between different types of pro attitudes—such as goals versus mere wishes—and explore the model checking problem to verify these logical formulas within active computer programs.

Summary updated

BibTeX
@inproceedings{su2007modal,
  author = {Kaile Su and Abdul Sattar and Han Lin and Mark Reynolds},
  title = {{A Modal Logic for Beliefs and Pro Attitudes}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2007},
}
Figure