A Framework for Studying Communication Pathways in Machine Learning-Based Agent-to-Agent Communication

Sathish Purushothaman, Michael Granitzer, Florian Lemmerich, Jelena Mitrovic

2024

Abstract

: The rise of Large Language Models (LLMs) has increased the relevance of agent-to-agent communication, particularly in systems where agents learn from their interactions. However, current LLMs offer limited insights into the communication dynamics among multiple agents, especially in large-scale settings when multiple agents are involved. Particularly training LLMs - in contrast to in-context learning - becomes nearly infeasible without large-scale computing infrastructure. In our work we present a machine-learning based agent framework to investigate the role of different communication pathways for studying language emergence between machine learning-based agents. We designed a transformer-based image auto-encoder as the agent architecture. A Gumbel SoftMax layer encodes images in form of symbols forming the language between agents. We study two pathways: In the first pathway, the sender reads an image and sends a message to the receiver. The receiver uses the message to reconstruct the sender’s image. In the second pathway, the sender and receiver read an image and minimize the distance between the generated symbols. In the first pathway, language emerges with the Levenshtein distance of ≤ 2 for 96% of messages. In the second pathway, no language emerges with the Levenshtein distance of ≤ 2 for 3% of messages.

Paper Citation

in Harvard Style

Purushothaman S., Granitzer M., Lemmerich F. and Mitrovic J. (2024). A Framework for Studying Communication Pathways in Machine Learning-Based Agent-to-Agent Communication. In Proceedings of the 16th International Conference on Agents and Artificial Intelligence - Volume 1: ICAART; ISBN 978-989-758-680-4, SciTePress, pages 117-126. DOI: 10.5220/0012423500003636

in Bibtex Style

@conference{icaart24,
author={Sathish Purushothaman and Michael Granitzer and Florian Lemmerich and Jelena Mitrovic},
title={A Framework for Studying Communication Pathways in Machine Learning-Based Agent-to-Agent Communication},
booktitle={Proceedings of the 16th International Conference on Agents and Artificial Intelligence - Volume 1: ICAART},
year={2024},
pages={117-126},
publisher={SciTePress},
organization={INSTICC},
doi={10.5220/0012423500003636},
isbn={978-989-758-680-4},
}

in EndNote Style

TY - CONF

JO - Proceedings of the 16th International Conference on Agents and Artificial Intelligence - Volume 1: ICAART
TI - A Framework for Studying Communication Pathways in Machine Learning-Based Agent-to-Agent Communication
SN - 978-989-758-680-4
AU - Purushothaman S.
AU - Granitzer M.
AU - Lemmerich F.
AU - Mitrovic J.
PY - 2024
SP - 117
EP - 126
DO - 10.5220/0012423500003636
PB - SciTePress