Learning to Efficiently Pursue Communication Goals on the Web with the GOSMR Architecture

Kevin GoldAAAI 2013 · DOI

Abstract

Simulating realistic human behavior at scale is a critical requirement for cybersecurity research, particularly when evaluating network resilience under disruption. This paper presents the Goal Oriented Scenario Modeling Robots (GOSMR) architecture, a lightweight cognitive framework that enables agents to pursue communication goals on the web. GOSMR combines partial-order planning, temporal difference learning, and an "affordance" abstraction layer to allow agents to adapt to varying site performance and network conditions. Our evaluation shows that agents utilizing utility-proportional selection significantly outperform those using standard Boltzmann distributions, achieving nearly double the task completion rate with minimal computational overhead.

Introduction

In cybersecurity emulation, researchers often need to generate "background" traffic that mimics human users. When a network channel is disrupted, realistic agents should adapt—perhaps by switching to a different communication platform or adjusting their retry frequency. Traditional cognitive architectures like ACT-R or SOAR, while sophisticated, are often too computationally intensive to run in the hundreds on standard virtualized hardware. Furthermore, their rule-based nature struggles with the probabilistic delays and failures inherent in web environments. GOSMR was developed to provide a "smart" yet frugal alternative for large-scale experimentation, capable of making rational decisions under the uncertainty of the modern internet.

The GOSMR Architecture

GOSMR integrates several key AI techniques to achieve adaptive, goal-directed behavior:

  • Partial-Order Planning: Agents search backward from a goal (e.g., "Agent A has communicated Fact B") to generate sequences of actions. By sampling multiple plans rather than just picking the most optimal one, GOSMR ensures behavioral diversity.
  • Affordances: This layer separates high-level intent (e.g., Login) from low-level execution (e.g., specific Selenium clicks). This allows the same agent logic to operate across different website UIs without hard-coding every interaction.
  • Temporal Difference (TD) Learning: Agents learn the expected duration of actions through experience. This allows them to dynamically prioritize faster websites and avoid those that are consistently slow or down.
  • Utility-Proportional Selection: The authors argue that the widely used Boltzmann distribution for goal selection is problematic because it is not scale-free. Instead, GOSMR uses selection probabilities directly proportional to a plan's utility, which proves more robust for producing adaptive behavior.
  • Frugal Sensing: To maintain a low footprint, GOSMR only checks for web elements that confirm or deny its current beliefs, significantly reducing the overhead of running headless browsers.

Experimental Setup

The architecture was tested in a live, controlled web environment featuring several custom sites designed to test different performance profiles:

  • FakeFacebook: A social network built on the Elgg framework requiring multi-step navigation.
  • QuickMail & SlowMail: Custom mail sites, with SlowMail featuring a hard-coded 10-second artificial delay.
  • DownMail: A site that remained offline for the duration of the experiment to test persistent failure handling.
The experimental web environment: (a) FakeFacebook and (b) SlowMail.

Results and Discussion

A demonstration with 20 agents over one hour revealed significant differences in performance based on the selection strategy. Agents using utility-proportional selection successfully gravitated toward the most efficient site (FakeFacebook) as they learned its performance characteristics. Once logged in, FakeFacebook proved faster than the single-threaded QuickMail site. These agents completed 1,604 successful communication actions, compared to only 856 for agents using Boltzmann selection.

Comparison of agent activity over time using (a) Boltzmann and (b) Utility-proportional selection strategies.

Crucially, GOSMR is highly efficient. The 20 agents and their associated infrastructure consumed only 5% of a standard laptop's CPU. This footprint is comparable to a single instance of a modern web browser, proving GOSMR is lightweight enough for massive scaling in virtualized environments.

Conclusion

GOSMR provides a scalable solution for simulating adaptive human behavior on the web. By abstracting UI interactions through affordances and employing robust selection mechanisms, it bridges the gap between simple scripts and heavy cognitive models. Future work will focus on automating affordance discovery via machine learning and implementing more sophisticated trust models for multi-agent communication and human-agent cooperation.

Summary updated

BibTeX
@inproceedings{gold2013learning,
  author = {Kevin Gold},
  title = {{Learning to Efficiently Pursue Communication Goals on the Web with the GOSMR Architecture}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2013},
  doi = {10.1609/aaai.v27i1.8469},
}
Figure