[ad_1]
Alphabet’s Loon, the team responsible for transmitting the Internet to Earth from stratospheric helium balloons, has taken a new step: its navigation system is no longer run by man-made software.
Instead, the company’s internet balloons are driven around the world by artificial intelligence – in particular, a set of algorithms both written and executed by a reinforcement learning-based flight control system. deep which is more efficient and skillful than the old, man-made one. a. The system now manages Loon’s balloon fleet over Kenya, where Loon launched its first commercial internet service in July after testing its fleet as part of a series of disaster relief and disaster relief initiatives. ‘other test environments for much of the past decade.
Similar to how researchers achieved groundbreaking advances in AI by teaching computers to play sophisticated video games and helping software learn to manipulate robotic hands realistically, reinforcement learning is a technique that allows software to teach itself skills through trial and error. Obviously, such repetition is not possible in the real world when it comes to high altitude balloons which are expensive to operate and even more expensive to repair in the event of a crash.
So Loon, like many other AI labs who have turned to reinforcement learning to develop sophisticated AI programs, taught his flight control system how to fly balloons using the computer simulation, with the help of the Google AI team in Montreal. That way, the system could improve over time before being deployed to a real-world balloon fleet.
“While the promise of RL (reinforcement learning) for Loon was always great, when we first started exploring this technology it was not always clear that deep RL was practical or viable for high altitude platforms. drifting through the stratosphere autonomously for long periods of time, ”explains Sal Candido, technical director of Loon and co-author of an article on the new flight control system published this week in the scientific journal Nature, in a blog post. “It turns out that RL is handy for a fleet of stratospheric balloons. Loon’s navigation system’s most complex task today is solved by a computer-learned algorithm experimenting with balloon navigation in simulation.
Loon says his system is considered the world’s first deployment of this variety of AI in a commercial aerospace system. And not only that, but it actually surpasses the system designed by humans. “To be frank, we wanted to confirm that using RL a machine could build a navigation system equal to what we had built ourselves,” writes Candido. “The learned deep neural network that specifies the flight controls is wrapped in an appropriate safety assurance layer to ensure that the agent is always driving safely. Across our simulation benchmark, we were able to not only replicate, but also significantly improve our navigation system using RL. “
In its first real-world test in Peru in July 2019, the AI-controlled flight system clashed with a traditional flight system, controlled by a human-built algorithm called StationSeeker, which was designed by Loon engineers themselves. “In a sense it was the machine – that spent a few weeks building its controller – against me – that, along with many others, had spent many years carefully tweaking our conventional controller based on a decade of experience working with Loon balloons. We were nervous… and we were hoping to lose, ”says Candido.
The AI-controlled system has vastly outperformed the human system in staying ever closer to a device the team uses to measure LTE signals in the field, and this test paved the way for further experiments to prove the efficiency of the system before it formally replaced the one the team had spent years building by hand. Loon now believes his system can “serve as proof that RL can be useful in controlling complex and real systems for fundamentally continuous and dynamic activity.”
In his closing remarks, Candido raises the question of whether this type of AI is worthy of the name, due to its specialization and its resemblance to a traditional automated system but not self-learning like those that operate machines. heavy. or public transport control elements.
“While there is no chance that a pressure balloon effectively drifting through the stratosphere will become sensitive, we have moved from designing its navigation system to building it by computers in a data-driven manner.” , he said. “Even though this isn’t the start of an Asimov novel, it’s a good story and maybe something worth calling AI.”
[ad_2]
Source link