August 22, 2017
Ultra-high-speed Reinforcement Learning, a Key Challenge in AI, Achieved Using Laser Chaos
Enabling "Instantaneous Adaptation" in Frequency Allocation and Other Applications
Highlights
■ Ultra-high-speed implementation of reinforcement learning, one of the fundamental problems in AI, using optical chaos generated from a laser.
■ Achieved instantaneous "decision-making" by leveraging the high speed of light and natural physical phenomena, confirming superior performance.
■ Expected to contribute as a foundational technology for AI and IoT, such as in the instantaneous allocation of frequencies.
Principal Investigator Makoto Naruse of the National Institute of Information and Communications Technology (NICT; President: Hideyuki Tokuda), Professor Atsushi Uchida of the Graduate School of Science and Technology at Saitama University (President: Hiroki Yamaguchi), and Project Associate Professor Song-Ju Kim of the Graduate School of Media and Governance at Keio University (Dean: Yasushi Kiyoki) have succeeded for the first time in the world in applying ultra-high-speed photonics to reinforcement learning, achieving an adaptation speed of 1 GHz (gigahertz: 1 billion times per second) using optical chaos *1 (hereafter referred to as laser chaos) generated from a semiconductor laser.
Focusing on the high speed of light, NICT and its collaborators have successfully solved the "two-armed bandit problem"—the challenge of selecting the slot machine with the higher winning probability from two machines with unknown probabilities—physically and at high speed. This was achieved by combining the random signals generated by the chaotic phenomenon in a semiconductor laser with a uniquely developed reinforcement learning method, leveraging the ultimate performance of light. Thanks to the ultra-high speed of the laser chaos phenomenon, a rapid decision-making process with a latency *2 (the time from information input to output) of 1 ns (nanosecond: one-billionth of a second) was confirmed. Furthermore, it was verified that this method demonstrates superior performance compared to fast, virtually generated pseudorandom numbers *3 (colored noise).
This technology is expected to make significant contributions as a foundational technology for AI and IoT, for applications such as arbitration *4 for instantaneous mediation of computing resources for faster computing, and the instantaneous allocation of frequencies in wireless communications. This research was published in Scientific Reports at 6:00 p.m. (JST) on Friday, August 18.
The press release can be viewed at the following link: Ultra-high-speed Reinforcement Learning, a Key Challenge in AI, Achieved Using Laser Chaos: Enabling "Instantaneous Adaptation" in Frequency Allocation and Other Applications.pdf
For inquiries regarding this matter:
Keio University
Project Associate Professor, Graduate School of Media and Governance
Song-Ju Kim
E-mail: songju@sfc.keio.ac.jp
Distributed by:
Office of Research Development and Sponsored Projects at Shonan Fujisawa Campus