Adaptive Cooperative Learning for Spatio-Temporal Environmental Variation in Vehicular Communications

2023 Fiscal Year Final Research Report

• Project/Area Number: 21K17734
• Research Category: Grant-in-Aid for Early-Career Scientists
• Allocation Type: Multi-year Fund
• Review Section: Basic Section 60060:Information network-related
• Research Institution: The University of Tokyo (2022-2023); Aoyama Gakuin University (2021)
• Principal Investigator: Akihito Taya 東京大学, 生産技術研究所, 助教 (10867948)
• Project Period (FY): 2021-04-01 – 2024-03-31
• Keywords: 車車間通信 / 分散協調学習 / Federated Learning / 環境適応

Outline of Final Research Achievements

Developing connected cars equipped with communication functions is essential for autonomous driving. In automotive communication, high reliability and low latency are required for safety, and to achieve this, communication control tailored to the environment, such as radio wave propagation and congestion, is necessary. In this study, we developed an algorithm that reduces the amount of communication needed during decentralized machine learning by vehicles when learning the environment. We also extended a method to predict the communication characteristics in an unknown area from the model of the learned area for a communication environment with spatial dependence.

Free Research Field

情報通信

Academic Significance and Societal Importance of the Research Achievements

開発した分散機械学習手法は従来手法とは異なり、ニューラルネットワークのパラメータを車両間で共有しながら共通のパラメータに収束させるのではなく、異なるパラメータであっても同一入力に対して同一出力すれば同じ予測モデルであるという考えのもと、関数そのものを最適化対象とすることが特徴である。この手法の考案にあたり、数値を収束させるのではなく、関数である予測モデルを収束させるという発想の転換があり、今後の分散機械学習の発展への貢献が期待される。