A new AI can pilot balloons in the stratosphere
A type of AI can be used to navigate balloons in the unpredictable conditions of the stratosphere, researchers report this week.
Why it matters: It's a demonstration of an AI making decisions in a messy real-world environment.
What's happening: Researchers at Google Brain and Loon used a form of AI called reinforcement learning to control a balloon over the Pacific Ocean for a period of 39 days.
- The AI used historical weather observations and current reports to determine when to gain or lose altitude in order to find the best winds to keep the balloon's position. (It also monitored the balloon's power.)
- The researchers augmented wind direction and speed data with randomly generated "noise" — giving the AI a range of possible wind conditions it could encounter in order to assess its current situation and future path.
- The algorithm improved decision-making time and uses less power than the existing navigator. The balloon also spent a greater proportion of time close to its station, they report this week in the journal Nature.
The work "represents a big advance in the use of reinforcement learning for real-world applications," Scott Osprey of the U.K.'s National Centre for Atmospheric Science wrote in an accompanying article.
- Balloons could be used for monitoring air quality and other aspects of the environment.
The big picture: Unlike board games or card games that algorithms have mastered where all the information is available or known, the real world presents partial information and plenty of unknowns.
- "For me, the most exciting part and what will be a differentiator is the ability to reason about what we don’t know," says Marc Bellemare, a co-author of the study and a researcher at Google Brain.
What's next: Loon has already deployed the controller on its balloons that provide internet service to parts of Kenya, says the company's chief technology officer Sal Candido.
- The next step, he says, is to try to use the new AI pilot to navigate balloons from launch sites, for example in Puerto Rico, to their final destination.