The Google X lab has been launching balloons into the stratosphere in an attempt to provide internet access to underserved parts of the world. This effort is called Project Loon, and was first announced in the summer of 2013. The first balloons stayed aloft for about 5 days, and increasing their flight time has been the primary challenge of the project ever since.
This past summer, the lab launched a balloon into the stratosphere over Peru that stayed aloft for 98 days.
Fighting the balloon’s natural tendency to simply float away has been the biggest obstacle to keeping them in the air long enough to provide viable internet access. As with a hot air balloon, the navigation system of these balloons are only able to move them up and down, relying on (or avoiding) weather patterns, since more complex navigation systems (like jet propulsion) would be too heavy and expensive.
When the project got started, the team used handcrafted algorithms, prepared to respond to a fixed set of variables such as altitude, location, and wind speed. To get results such as with the Peru balloon, the project turned to artificial intelligence. These new algorithms use machine learning to analyze huge quantities of data, and in the process actually learn over time. With more control over the navigation of these balloons, the project is able to use fewer balloons to provide internet. Instead of launching huge quantities, and essentially hoping for the best, they are able to launch a few balloons with improved navigation to stay where they are needed.
Google, along with other leading companies such as Facebook and Twitter, have increasingly turned towards the concept of deep neural networks – algorithms which loosely model the network of neurons in the human brain. Instead of engineers needing to hand-code each algorithm, they are able to learn on their own, and expand their capabilities. While the Project Loon navigation does not use deep neural networks exactly, it employs a similar but simpler approach to machine learning, called Gaussian Processes. It uses data from over 17 million kilometers of balloon flights since the start of Project Loon, to predict whether the balloon should be navigated up or down at any given time. To deal with the unpredictable nature of the weather the team has encountered, they’ve added reinforcement learning to the process. This allows the AI to collect additional data on what actions work and which don’t, to continue to modify future behavior.
The project’s navigation systems rely on access to massive Google data centers. Access to such vast oceans of data is improving the capabilities of all sorts of technologies, and will continue to reshape people’s lives, and relationships to technology. Providing internet access to formerly ignored regions of the globe is just the beginning.