Artificial intelligence isn’t always intelligent
We all got excited when we heard about Google planning to launch balloons into the stratosphere to provide Internet in regions where it is otherwise unavailable. However, we have not heard much about any progress.
This summer Google X Lab, or just X as it is known under the Alphabet’s umbrella, launched a balloon into the stratosphere over Peru where is stayed for 98 days. This is an impressive achievement given that much like regular balloons, the balloons of Lab X have one big problem – they tend to float away.
The navigation system that scientists at X use allows them to only move the balloons up and down, while pumping air into a balloon inside the balloon or pumping the air out — just like hot air balloons. One reason for this is the fact that a more complex navigation system would probably prove to be too heavy for the balloons. Taken this big limitation into account, the scientists had to search for other solutions.
Given the recent success another Google spin-off had lately with AlphaGo, an AI system that beat one of the world’s best player at the Chinese game of Go, the project Loon turned to artificial intelligence. To ensure the victory for the AlphaGo system, the scientists used reinforcement learning. When the system learned how to play Go by analysing human moves, it then played game after game, analysed its success and used this data to improve its behaviour. Following the vision of the creators of AlphaGo, reinforcement learning started to being used for various other tasks, including Project Loon.
To demonstrate the progress of the algorithms, it is sufficient to note that the balloons were navigated by handcrafted algorithms that were not able to react to the uncertainty that the balloons were facing in stratosphere. They were able to respond only to a predetermined set of variables, like altitude, location, wind speed, etc. Thus, there would be a chance of the balloon floating away over the ocean if it was exposed to an extreme situation, which the algorithm did not account for. However now, with the help of reinforcement learning the balloons are using all those different scenarios they have encountered in order to improve their behaviour and navigation in the future. After collecting data on over 17 million kilometres of balloon flights, the navigation system can more effectively predicting the course balloon should take, whether it should move the balloon up or down. While these predictions are not always perfect, they represent a starting point of a new large shift across the tech world as a whole.
What do you think, are we going to have an internet balloon during our next hike in the mountains?
