NTT Corporation (NTT) and the National Institute of Informatics (NII), in collaboration with CNRS Institut de Recherche en Informatique et Systèmes Aléatoires (IRISA) and Wavely, have been chosen through a competitive selection process led by the Japan Science and Technology Agency (JST) and the French National Research Agency (ANR). This selection falls under the Japan-France Strategic International Collaborative Research Program (SICORP), focusing on “Edge AI”. Our joint venture, known as LIGHT-SWIFT, aims to develop a lightweight edge AI technology tailored for smart factories, optimizing wireless communications and sensing while conserving power.
Our research endeavors to create an edge AI solution capable of operating efficiently on small-scale devices with limited computational and battery resources. Additionally, we aim to pioneer deep learning and reinforcement learning-based wireless access technology, derived from our collective research efforts. The focal point lies in crafting intelligent wireless access technologies capable of learning and adapting to dynamic environmental changes, such as those encountered within factory settings where radio channel conditions may rapidly degrade due to moving objects and equipment.
- Background: In the envisioned 6G era of the 2030s, the proliferation of Industrial Internet of Things (IIoT) terminals for factory automation and autonomous driving underscores the importance of wireless communications. However, ensuring reliable communication amidst fluctuating environmental conditions poses a significant challenge. Traditional centralized control mechanisms are insufficient for managing the multitude of IIoT terminals efficiently, consuming excessive power and bandwidth.
- Key Research Points: Our project aims to develop a high-performance, low-power edge AI technology—termed “lightweight AI”—suitable for edge devices like IoT terminals. This technology seeks to optimize wireless communications and sensing in dynamic environments, such as factories. Unlike conventional centralized approaches, lightweight edge AI empowers distributed terminals to autonomously optimize their wireless access, ensuring rapid adaptation to environmental changes.
NTT and NII will spearhead the development of a novel radio access technology utilizing lightweight edge AI, building upon previous joint research endeavors. The objective is to halve power consumption at edge devices and decrease reinforcement learning convergence time by half compared to current benchmarks.
- Multi-Radio Access Technology Leveraging Reinforcement Learning: Through collaborative efforts, NTT and NII have devised a multi-radio access technology based on Risk-Averse Q-learning (RA-QL), an advanced reinforcement learning method. This technology enables autonomous selection of the best radio interface and packet transmission rate to achieve reliable communications in dynamic environments. Additionally, we propose a wireless access method leveraging deep reinforcement learning, designed for edge AI, to further reduce processing and energy consumption.
- Institutional Roles: NTT will contribute to technology creation, architecture examination, and experimental evaluation. NII will lead the design and evaluation of core algorithms, representing the Japanese organization within SICORP.
- Future Prospects: Our research aims to advance smart wireless access technologies employing lightweight edge AI, ensuring high-quality wireless communications amidst complex mobile environments while achieving significant energy savings. Through our collaborative project, we strive to establish practical edge AI-based technologies for future smart factor.
