Towards Action Prediction Using a Mental-Level Model

Ronen I. Brafman, Moshe TennenholtzIJCAI 1995 · OpenReview

Abstract

We propose a formal approach to the problem of prediction based on the following steps: First, a mental-level model is constructed based on the agent's previous actions; next, the model is updated to account for any new observations by the agent, and finally, we predict the optimal action w.r.t. the agent's mental state as its next action. This paper formalizes this prediction process. In order to carry out this process, we need to understand how a mental state can be ascribed to an agent and how this mental state should be updated. Here we investigate a particular update operator and show that its ascription requires making only weak modeling assumptions.

Introduction

Predicting the behavior of complex agents—whether robots, software, or humans—is a fundamental challenge in artificial intelligence. Traditional approaches often rely on analyzing low-level specifications, such as source code or mechanical blueprints. However, this is frequently computationally prohibitive or impossible when the underlying structure is inaccessible.

This paper advocates for an abstract alternative: the mental-level model. By treating an agent as if it possesses beliefs, goals, and a rational decision-making process (the "intentional stance"), we can predict its actions without knowing its internal "wiring." The central challenge addressed here is formalizing how to update these ascribed beliefs when the agent encounters new information.

Methodology

The authors propose a three-step iterative loop for action prediction:

  1. Ascription: Based on past behavior and known goals, identify a set of beliefs that renders the agent's actions "rational."
  2. Update: Adjust these ascribed beliefs based on new observations the agent makes (e.g., detecting an obstacle).
  3. Prediction: Determine the "best" action for an agent in that updated mental state.
The three-step prediction process: from past behavior to predicted behavior via a mental-level model.

This process relies on the Agency Hypothesis, which assumes the agent follows a protocol maximizing its perceived outcome according to a decision criterion like maximin. The technical core is the concept of Admissibility: a belief update is admissible if the agent simply narrows its focus to previously plausible scenarios that remain consistent with new data.

Theoretical Insights and Case Study

The framework is demonstrated through a toy example: a robot tasked with finding a target at one of three locations (A, B, or C). By observing the robot's path, a modeler can "back out" its internal beliefs. For instance, if the robot moves toward A but then changes course, we can infer its updated beliefs about the target's location.

An observed path allows a modeler to ascribe specific beliefs to the robot.

Key theoretical findings include:

  • Stability of Rationality: Theorem 2 shows that for agents with admissible beliefs, a plan that was initially optimal remains optimal as the agent progresses, provided observations don't contradict initial theories.
  • Generality of Admissibility: Theorem 4 proves that if an agent's behavior can be modeled using beliefs at all, it can be modeled using admissible beliefs, making it a robust modeling assumption.
  • Computational Simplification: Theorem 5 demonstrates that admissibility allows behavior prediction using local utility functions rather than complex future path analyses.

Conclusion and Limitations

While the mental-level model provides a powerful abstraction, it faces hurdles with unexpected observations. If an agent encounters data that completely contradicts its current beliefs, the model lacks a clear predictive path. Future work aims to incorporate "inductive leaps" and formal languages for state description to handle such revisions. Additionally, the current model assumes perfect recall, which may not apply to memory-constrained systems.

Summary updated

Abstract

We propose a formal approach to the problem of prediction based on the following steps: First, a mental-level model is constructed based on the agent's previous actions; next, the model is updated to account for any new observations by the agent, and finally, we predict the optimal action w.r.t. the agent's mental state as its next action. This paper formalizes this prediction process. In order to carry out this process, we need to understand how a mental state can be ascribed to an agent and how this mental state should be updated. In [Brafman and Tennenholtz, 1994b], we examined the first stage. Here we investigate a particular update operator and show that its ascription requires making only weak modeling assumptions.
BibTeX
@inproceedings{brafman1995towards,
  author = {Ronen I. Brafman and Moshe Tennenholtz},
  title = {{Towards Action Prediction Using a Mental-Level Model}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {1995},
  url = {https://openreview.net/forum?id=SkV6WVM_bS},
}
Figure