Predictive Modelling of Human Reasoning Using AGM Belief Revision

Clayton K. BakerIJCAI 2023 · DOI

TL;DR

This paper proposes an empirical investigation into how humans revise their beliefs when faced with conflicting information, aiming to bridge the gap between formal belief revision theories (like AGM) and actual human reasoning. By comparing human responses to predictions from formal models, the research seeks to develop a more psychologically realistic model of belief change.

What Problem Does This Paper Address?

Understanding how humans update their beliefs in the face of new, potentially contradictory information is a fundamental challenge in cognitive science and artificial intelligence. Traditional logic struggles with the nuances of human reasoning, which often relies on context and intuition rather than strict rules. Non-monotonic reasoning and belief change theories offer more flexible frameworks, but their psychological validity remains an open question. Specifically, this work addresses the gap between formal belief revision theories, such as the Alchourrón, Gärdenfors, and Makinson (AGM) framework, and how humans actually perform belief revision. The authors want to know if humans follow logical postulates when revising beliefs and how factors like relevance and confidence influence this process.

Methods

The core of this project involves designing and conducting experiments with human subjects. The proposed methodology includes two key experiments:

  1. Information Relevance: This experiment aims to understand how subjects identify relevant information within a given context.
  2. Information Confidence: This experiment will assess how subjects value or esteem beliefs after learning new information.

The data collected from these experiments will be used to build representations of subjects' beliefs. These representations will then be analyzed using standard statistical tests. A key aspect of the analysis will be to determine if a cognitive model called MMT (Mental Model Theory), implemented in the CCOBRA software platform, can predict individual human responses. The predictive power of MMT will be compared against the predictions derived from the 8 AGM postulates and desirable postulates for information relevance.

Datasets

Not discussed in the paper. The paper outlines proposed experiments and data collection from human subjects, but does not refer to pre-existing datasets for training or evaluation.

Key Insights and Ideas

  • Bridging Formal Logic and Human Cognition: The paper highlights the ongoing challenge of aligning formal, logical models of belief revision with the often less formal and more context-dependent nature of human reasoning.
  • The Role of Relevance and Confidence: The research posits that human belief revision is not solely driven by logical consistency but is significantly influenced by the perceived relevance of new information and the confidence subjects have in their existing beliefs.
  • Mental Models as a Predictive Tool: The project explores the potential of Mental Model Theory (MMT) as a psychologically grounded framework for predicting human belief revision, contrasting its approach with purely logical systems.
  • Empirical Validation of Belief Revision: A central idea is the need for empirical validation of belief revision postulates, moving beyond theoretical constructs to understand how real people actually update their knowledge.

Key Results

Not discussed in the paper. The paper outlines a proposed research project and its methodology, including planned experiments and analyses. It does not present any experimental results or findings at this stage.

Limitations and Future Work

Limitations: * Previous Work Limitations: The authors acknowledge limitations in their prior work [Baker and Meyer, 2021], including using only one translation per AGM postulate and a lack of detailed comparison across postulates. * AGM Completeness: While the AGM framework is a sound basis, it is not considered complete. The paper mentions recent developments in relevance and independence that address this. * Psychological Realism: The paper questions whether current belief revision theories are perceived as natural and useful by humans, suggesting a potential disconnect.

Future Work: * Varied Material for Testing: The authors plan to test each postulate repeatedly using varied material to better understand human responses, as humans construct visual representations of their environment. * Broader Social Science Impact: In the future, the research aims to consider the impact of broader work in social science literature. * Enhancing Cognitive Models: A long-term goal is to add significant features of belief revision to the existing MMT model within the CCOBRA platform.

Summary updated

BibTeX
@inproceedings{baker2023predictive,
  author = {Clayton K. Baker},
  title = {{Predictive Modelling of Human Reasoning Using AGM Belief Revision}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2023},
  doi = {10.24963/ijcai.2023/811},
}
Figure