[1H3-OS-40] OS-40

2:25 PM - 2:40 PM JST(5:25 AM - 5:40 AM UTC)

[1H3-OS-40-04] AutoML Arena: Predictive Modeling Competitions Driven by AutoML Agents

〇Yuki Ochiai¹, Kyohei Atarashi¹, Koh Takeuchi¹, Yuko Sakurai², Satoshi Oyama³, Hisashi Kashima¹ (1. Kyoto University, 2. Nagoya Institute of Technology, 3. Nagoya City University)

Keywords:

LLM Agent,Multi-Agent System,Machine Learning Competition,Leaderboard

AutoML is a framework that automates machine learning processes, and in recent years, numerous AutoML agents incorporating large language models (LLMs) have been proposed. However, much of the existing research has focused on improving the performance of individual agents, with insufficient consideration given to the cooperation and competition between multiple agents. Meanwhile, prediction modeling competitions, exemplified by Kaggle, are known to yield models achieving performance levels difficult for individuals to attain through competition among participants. Therefore, this study verified whether competition among agents leads to performance improvement by deploying multiple AutoML agents in an environment simulating a predictive modeling competition. Experiments conducted on a custom-built platform with a leaderboard demonstrated that agents effectively utilized information shared via the leaderboard, resulting in improvements across several evaluation metrics.