Certain things are at the core of the Internet, and without them the Internet simply could not exist as we know it today. Among these is the Transmission Control Protocol, or TCP, which has among its responsibilities regulating the rate at which data is sent between computers, to ease congestion. For years human engineers have been tweaking TCP to improve performance, and now researchers at MIT have developed an artificial intelligence to do the same.
There are many things the human brain is able to do well, but there are some things computers can do better, and among them is managing complex systems. A human engineer can certainly work with a network consisting of a few connections, but when a handful or more computers are added, the network becomes very hard to work with. The machine learning program, Remy, however thrives in these situations, as it can quickly test a multitude of protocols, to determine which is best for the situation. Unlike the human-made protocols, Remy's can consist of over a hundred rules, and exactly why these rules work as well as they do is still being examined.
The researchers, who will be showing Remy off at the conference for the Association for Computing Machinery's Special Interest Group on Data Communication, have tested it on a standard simulation system called ns-2 and on a simulation of Verizon's cellular data network. Remy's solution for ns-2 ended up doubling throughput and cutting delay by more than half, while the solution for Verizon improved throughput by 20-30% and delay by 25-40%.