This particular AI was developed using deep reinforcement learning, in which a system teaches itself based on trial and error. However, allowing the system to train itself that way in the real world with actual high-altitude balloons would have been very costly when they crashed. So Loon teamed up with Google’s Montreal-based team to train the AI with a computer simulation, allowing it to crash simulated balloons as many times as it needed in order to learn how to fly them.
In a real-world test last year, the AI flight-system proved to be a better pilot than the human-built algorithm, so Loon has now officially handed over the reins. “It turns out that RL (reinforcement learning) is practical for a fleet of stratospheric balloons,” said Loon Chief Technology Officer Sal Candido. “These days, Loon’s navigation system’s most complex task is solved by an algorithm that is learned by a computer experimenting with balloon navigation in simulation.”
