The research groups of the Information and Communication Research Organization (NICT), Saitama University, and Keio University have succeeded for the first time in the world in reinforcement learning applying ultra-high-speed photonics using optical chaos generated from semiconductor lasers.
Technologies called "deep learning" and "reinforcement learning" are important for machine learning that supports modern artificial intelligence (AI).Reinforcement learning is a method of learning through trial and error in an unknown environment.The central issue is the problem of "how to maximize profits in a casino with many slot machines" (= "multi-armed bandit problem").
To maximize profits, it is necessary to try out the hitting table, but if you try too much, you will lose money, and if you stop early, you will miss the hitting table.The solution to this "search" or "decision" problem is to randomly select a platform, which used to use "pseudo-random numbers" that are generated based on calculation procedures on a computer, but it takes time and the random numbers " There was also a limit to "quality".Therefore, the research group proposed "reinforcement learning using physical phenomena in the natural world directly" and succeeded in realizing a method using a single photon.
In this research, we have realized an ultra-fast solution to the problem of selecting a machine with a high probability of hitting from two slot machines by using the "chaos phenomenon" (Note) of light that occurs when a semiconductor laser is used.Laser chaos light generated from a semiconductor laser is sampled at high speed, and decision making is made only by determining the magnitude of the "threshold value" based on the "past battle history".As a result, the time from input to output of information is one billionth of a second (2 nanosecond), which is a high-speed decision-making, and the performance is superior to the conventional method using pseudo-random numbers.In the future, it is expected to make a great contribution as a basic technology of AI and IoT, such as instantaneous frequency allocation.
(Note) A phenomenon in which when the light emitted from a semiconductor laser is reflected by a mirror and returned to the laser, the operation of the laser becomes unstable and chaos (a phenomenon in which a small difference in the initial state increases explosively over time) occurs.
Paper information:[Scientific Reports] Ultrafast photonic reinforcement learning based on laser chaos
