The Loon division of Alphabet, the parent company of Google, has just celebrated an unprecedented feat: responsible for sending internet to remote locations on the planet using stratospheric balloons, the company was able to deploy a navigation system in these vehicles whose software is an artificial intelligence ” developed “capable of learning alone.
Currently, Loon’s internet balloons are conducted around the Earth by artificial intelligence (AI), which is based on a set of algorithms written and executed by a flight control system. The novelty is that this system is no longer controlled by people, but self-regulated by a technique called “reinforcement learning” (RL).
The system is already in operation, and managing the helium balloon fleet in Kenya, where Loon launched its first commercial internet service last July.
Source: Loon / DisclosureSource: Loon
An AI-piloted flight system
Unlike tests developed in AI that teach computers how to play video games or learn how to manipulate a robotic hand, RL allows software to learn by trial and error. However, since operating balloons involves very high risks and costs, Loon taught its flight control to fly balloons using computer simulations.
Thus, the system was improved over time, before being deployed in the balloon fleet in the real world. Loon’s chief technology officer, Sal Candido, says that RL is feasible and that “nowadays, the most complex task of Loon’s navigation system is solved by an algorithm learned by a computer, which learns with a virtual system balloon navigation ”.
In the first real-world test, held in Peru in July 2019, the machine competed with Candido himself and his team, in building an algorithm that would be its controller. The researchers had the joy of losing to AI, which, in a few weeks, developed a better system than the human developed in many years.