While there still might be no general theory of networking for machine learning to learn from, it’s possible to chip away at the problem. Startup Aria Networks has been doing that, using genetic algorithms to build up artificial intelligence for networking.
Some of those results will be on display at Mobile World Congress in Barcelona, Spain. Specifically, Aria will show off its AI algorithms working inside a software-defined networking (SDN) controller from Chinese firm BOCO Inter-Telecom.
Facebook has given Aria a “like” as well. Aria will get name-checked next week when Facebook explains its backbone network at the Asia Pacific Regional Internet Conference on Operational Technologies (Apricot).
Genetic algorithms tackle a problem by evolving their way toward a solution. They adapt and tweak, seeking out changes that improve the results. In Aria’s case the algorithms are working on developing AI algorithms for networking. It’s AI writing AI, Perrett says.
Click here for SDxCentral’s full coverage of Mobile World Congress 2017.
His previous startup, Darwinian Neural Networks Industries (DNNI), applied this approach to genomics for the pharmaceutical industry. Perrett has since moved into networking, where the time to produce new products is faster. “There was more money in telecom than drug discovery, believe it or not,” he says.
So, Aria applies AI to the problem of building a path computation engine for networking.
Intelligence at the Edge
BOCO Inter-Telecom, one division among many inside China’s massive BOCO holding company, is China’s equivalent to Amdocs. It’s a specialist in OSS/BSS that sells software and professional services to the major telcos and mobile carriers.
BOCO developed a cross-domain SDN controller, meaning it could work with equipment from both Huawei and ZTE, the two vendors most frequently used by China’s operators. Aria’s intelligence is being used for path computation, but it’s also being applied to capacity planning. It can tell operators with the placement of virtual components, or flag when a network element such as a base station is about to be overloaded.
What’s interesting is that BOCO chose to put this intelligence in the controller rather than in the orchestrator. That turns Aria’s decisions into local ones, an architecture that BOCO and Aria think will scale further than a centralized model would. It’s similar to the thinking behind mobile edge computing (MEC) that’s starting to permeate the Internet of Things (IoT).
“My own personal view is that this will eventually be in the network elements itself,” Perrett says.