An abstract approach to reasoning about games with mistaken and changing beliefs

Benedikt Löwe, Eric PacuitAustralas. J. Log 2008 · DOI

Abstract

Traditional game theory often relies on the assumption of "common knowledge of rationality," where all players perfectly understand the rules and each other's preferences. However, in real-world social interactions and narratives, agents frequently operate under mistaken beliefs and revise their strategies as new information emerges. This paper presents a formal, descriptive framework for modeling such agents. By indexing preferences to specific players and game states, the authors introduce a subjective backward induction algorithm that mimics human-like reasoning in scenarios involving misunderstandings and surprises.

Introduction

Classical game theory, particularly the backward induction algorithm, is normative—it describes how an ideal, perfectly rational agent should behave. In contrast, "Epistemic Game Theory" attempts to handle belief complexities but often results in mathematically dense models focused on ideal behavior. Löwe and Pacuit identify a significant gap: the need for a simpler, descriptive way to model how agents (such as those in video games or virtual worlds) behave when they are operating under false assumptions. Their work aims to provide a framework that captures the "drama" of social interactions where players can be wrong about others' desires or change their minds based on unexpected moves.

Methodology: Subjective States and Backward Induction

The core of the proposed "abstract approach" is the encoding of beliefs directly into the game's state, avoiding the need for complex, external belief-revision functions.

  1. The State Function ($S$): The model employs a function that assigns a "description" to every node in a game tree. This description is indexed:
  2. $S(t, \emptyset)$ represents the "true" state of affairs at node $t$.
  3. $S(t, \mathbb{A})$ represents Player A's belief about the state.
  4. $S(t, \mathbb{AB})$ represents Player A's belief about Player B's belief.
  5. Subjective Backward Induction: The authors modify the standard Zermelo algorithm. Instead of a single "true" path, the model calculates subjective labels for every node. Each player performs backward induction based on their current, potentially mistaken, belief about the preferences of all other players.
  6. Belief Revision as State Change: When a player makes a move that contradicts another player's expectations, the "State" at the subsequent node is updated to reflect the revised belief.

Case Study: Modeling Narrative Dynamics

To demonstrate the framework, the authors utilize a qualitative benchmark: a fictitious "TV drama" script involving characters Sue, Jeff, Walter, and Mary. The narrative involves an affair, an ultimatum, and a series of miscalculations.

Figure 1: A formal game tree representing a social drama with players Mary, Walter, and Sue, illustrating subjective states and belief revisions.

The story is translated into a formal game tree (Figure 1) where outcomes are determined by the players' subjective interpretations of the situation (represented by the sun and rain cloud icons). For example, at node $t_0$, Walter must decide whether to say "YES" or "NO" based on his belief about Mary's future actions. The model successfully captures the "aha!" moment where Walter realizes he has misjudged Mary and is forced to revise his strategy, ultimately leading to a "true run" that matches the narrative's conclusion.

Key Insights and Results

The framework prioritizes simplicity over normativity. It does not require agents to be globally rational; it only assumes they follow backward induction based on their current subjective beliefs. By treating beliefs as strings of player symbols (e.g., $\mathbb{WMS}$ for Walter's belief about Mary's belief about Sue), the model handles deeply nested misunderstandings without complex modal logic.

A key result is the decoupling of labels from moves. In standard game theory, a node's label dictates the optimal move. Here, a player's subjective label might point to one outcome, but the actual move is determined by another player's potentially different subjective label. This creates the "surprise" element essential for realistic social modeling. Furthermore, the authors prove that if all beliefs are correct and unchanging, their algorithm converges to classical backward induction (Proposition 1).

Limitations and Future Work

While powerful, the model currently faces the "modeller's burden," requiring a human to manually translate natural language stories into formal state-description trees. Additionally, the framework allows for belief revision but does not dictate the specific logic of the update, leaving that to the application designer.

Future applications include Game Design, where this logic could create more lifelike NPCs capable of being "tricked" or "surprised," and Narrative Theory, providing a formal tool to analyze why certain plot twists are effective.

Summary updated

BibTeX
@article{lowe2008abstract,
  author = {Benedikt Löwe and Eric Pacuit},
  title = {{An abstract approach to reasoning about games with mistaken and changing beliefs}},
  journal = {Australas. J. Log.},
  year = {2008},
  doi = {10.26686/ajl.v6i0.1800},
}
Figure