Modeling Situation Awareness in Human-Like Agents Using Mental Models

Mark Hoogendoorn, Rianne van Lambalgen, Jan TreurIJCAI 2011 · DOI

Abstract

This paper presents a cognitive agent model designed to replicate human Situation Awareness (SA) through structured mental models. By simulating the three levels of SA—perception, comprehension, and projection—the model accounts for human-like cognitive constraints such as limited time and working memory. Evaluated in an F-16 combat flight scenario, the model demonstrates how these constraints lead to graceful degradation of awareness, mirroring human performance in high-stakes environments.

Introduction

In high-stakes, dynamic environments like air traffic control or military combat, effective performance relies on Situation Awareness (SA). SA is the ability to maintain a coherent internal representation of the external world. Traditional AI agents often rely on rigid "if-then" rules or purely statistical methods, failing to capture the nuances of human cognition. Specifically, they often overlook the hierarchical nature of SA (perception, comprehension, and projection), the impact of cognitive constraints (time and memory), and the role of domain expertise in organizing mental models.

The Proposed Cognitive Architecture

The authors propose a generic cognitive architecture for agents that incorporates four functional components:

  1. Observation Component: Collects raw environmental data with varying degrees of certainty.
  2. Current Belief Formation: Aggregates observations into simple beliefs, which are then integrated into complex beliefs (e.g., combining "radar blip" and "closing speed" to conclude "under attack").
  3. Future Belief Formation: Projects current complex beliefs into potential future states (e.g., predicting being "shot down").
  4. Observation Selection: A feedback mechanism where goals and projections prioritize relevant observations based on a "cost" function, simulating limited attention.
The proposed cognitive agent architecture showing the flow from observations to belief formation and the role of the mental model.

Mental Models and Belief Dynamics

Central to this architecture is the Mental Model, a network of domain-specific concepts and weights stored in long-term memory. Beliefs within the model are characterized by activation values that decay over time unless reinforced, mimicking human working memory.

The hierarchy of beliefs in the mental model, ranging from raw observations to future projections.

The model supports anytime updating, allowing the agent to reach conclusions even when reasoning is interrupted by time pressure. This leads to "rushed" thinking where complex integration (Level 2 SA) or projection (Level 3 SA) might fail while basic perception remains intact.

Evaluation in F-16 Combat Scenarios

The model was tested using a custom F-16 fighter pilot training scenario. Expert knowledge was used to define the mental model weights for cues like radar warning receiver signals.

  • Sufficient Resources: With ample time and memory, the agent achieved high activation levels (~0.95) for complex beliefs like "hostile" or "engaged."
  • Time Pressure: Under restricted reasoning time, activation levels for critical beliefs dropped significantly (~0.7), and complex beliefs were transient.
  • Memory Constraints: When observation costs exceeded capacity, the agent prioritized immediate threats but lost "big picture" context, such as the direction of the front line.

Key Insights and Future Work

The model successfully demonstrates graceful degradation of awareness under stress, aligning with psychological observations. Key insights include the role of attention as a resource and the power of projection in driving perception.

Future work aims to automate the learning of mental model weights from human data rather than relying on hand-tuned expert weights. Additionally, the authors plan to bridge the gap between awareness and decision-making by linking this SA model to naturalistic decision-making frameworks to observe how degraded awareness directly impacts tactical choices.

Summary updated

BibTeX
@inproceedings{hoogendoorn2011modeling,
  author = {Mark Hoogendoorn and Rianne van Lambalgen and Jan Treur},
  title = {{Modeling Situation Awareness in Human-Like Agents Using Mental Models}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2011},
  doi = {10.5591/978-1-57735-516-8/ijcai11-285},
}

Loading related content...

Figure