Feb 25 2019
New research leveraging machine learning could boost the efficiency of optical telecommunications networks. As the world becomes progressively interconnected, fiber optic cables provide the ability to convey more data across longer distances than traditional copper wires. Optical Transport Networks (OTNs) have come up as a solution for packaging data in fiber optic cables, and enhancements aim to make them more economical.
An artificial intelligence technique used for self-driving cars can also be used to make Optical Transport Networks run more efficiently. (Image credit: Getty Images/iStockphoto)
A team of scientists from Universitat Politècnica de Catalunya in Barcelona and the telecom company Huawei have retooled an artificial intelligence method used for self-driving cars and chess to help OTNs work more efficiently. They will showcase their research at the forthcoming Optical Fiber Conference and Exposition, to be held between 3rd and 7th March in San Diego, California, USA.
OTNs need procedures for how to divvy up the large amounts of traffic they manage, and writing the procedures for making those instant decisions becomes very difficult. If the network gives more space than required for a voice call, for instance, the unused space could have been put to better use ensuring that an end user streaming a video does not get “still buffering” messages.
What OTNs need is a better traffic guard.
The new approach formulated by the scientists to tackle this problem integrates two machine learning methods: The first, referred to as reinforcement learning, develops a virtual “agent” that learns via trial and error the specifics of a system to enhance how resources are managed. The second, known as deep learning, integrates an additional layer of superiority to the reinforcement-based method by using so-called neural networks, which are computer learning systems stimulated by the human brain, to obtain more abstract deductions from each round of trial and error.
Deep reinforcement learning has been successfully applied to many fields. However, its application to computer networks is very recent. We hope that our paper helps kickstart deep-reinforcement learning in networking and that other researchers propose different and even better approaches.
Albert Cabellos-Aparicio, Study Researcher, Universitat Politècnica de Catalunya.
Thus far, the most progressive deep reinforcement learning algorithms have been able to enhance a few resource allocations in OTNs, but they become trapped when they encounter novel situations. The scientists aimed to overcome this by changing the manner in which data are revealed to the agent.
Subsequent to learning the OTNs through 5,000 rounds of simulations, the deep reinforcement learning agent directed traffic with 30% better efficiency than the existing advanced algorithm.
One thing that amazed Cabellos-Aparicio and his team was how the new method was able to learn without difficulty about the networks after beginning with a blank slate.
This means that without prior knowledge, a deep reinforcement learning agent can learn how to optimize a network autonomously. This results in optimization strategies that outperform expert algorithms.
Albert Cabellos-Aparicio, Study Researcher, Universitat Politècnica de Catalunya.
With the mammoth scale certain OTNs already have, Cabellos-Aparicio said, even small progress in efficiency can obtain large returns in decreased latency and operational costs.
Going forward, the team plans to apply their deep reinforcement strategies along with graph networks, an evolving field within artificial intelligence with the prospect to change scientific and industrial fields, such as chemistry, computer networks, and logistics.