Two weeks ago, we attended AI Net Conference 2018 in Paris. The conference lasted three days and consisted of tutorial sessions and talks about artificial intelligence and how it can be used in networking. While AI is a huge topic of discussion in most every field, it is currently being underutilized in networking. It has the potential to be used in many areas of networking, including fault isolation, log analysis, design optimization, and many others.
Conference Attendees and Talks
The conference brought together a set of renowned experts, including researchers from Inria, Tokyo City University, and Cataluna University; operators from Orange, Google, and Expedia; industrials at Juniper Networks, Nokia, Huawei, Cisco, and Aria Networks; and consultants at companies including Rethink Research, OVUM, STL Partners, and Disruptive Analysis, to name a few.
In general, AI has gained a lot of traction in thought leadership in recent years due to the increase in computation power, storage, and cloud technology. This conference brought together people from various backgrounds and institutions to share ideas and find common ground for the future. The motto of the conference, “Towards Self Driven Networks”, exemplified this.
Our CEO, Varun Singh, gave a talk on Optimizing AI for Communications, where he discussed indispensible AI techniques we use at callstats.io to deliver optimal media quality and our latest product release, Optimize. Optimize uses artificial intelligence algorithms to provide media and network settings that deliver optimal audio and video quality for each device, connection, media, and network setup.
Evaluating AI In Networks
Currently, networks are managed manually by humans without much automation. The network industry is making strides to change this with the help of artificial intelligence. As most telecommunications operators have a vast amount of data, this goal is within reach.
Changes to software-defined networking (SDN) and network functions virtualization (NFV) are two big steps for automating networks with AI. With SDN, the controller for network components is being centralized. In NFV, partitioning network resources is becoming automated. From a networking perspective, the changes to SDN and NFV provide a common place for a network configuration. This is an important step for the AI integration, as this gives a common place for all configurations. The AI has a central place to detect, predict, and take control.
During these talks, the need for a skilled data scientist with a background in networking became apparent. Network data scientists are able to analyze the network data in-depth with the right techniques, which provides more understanding of the data. This enables any organization that has an interest in AI for networks to further develop their AI system.
AI Approaches for Networking Problems
AI can benefit many use cases in networks, either by automating processes or decreasing response times. Many in the networking community feel that artificial intelligence is an appropriate response to an increase in the complexity of networks, which may become unmanageable manually by humans. During the conference, AI techniques were applied to several topics, including routing updates, load distribution and link utilization, equipment faults, addressing the underlying cause for a series of events, network attacks, and exchange to exchange interactions.
There were many interesting and intriguing talks at the conference, but of the ones we attended, these stood out the most.
Intelligent Network Infrastructure
Huawei demoed a new AI they have developed that moves flows to lower latency paths as the need arises. This is not an easy task, as the AI needs to understand the network topology, and small changes in the topology can have big impacts. If the network topology is changed even a small amount, the AI needs to be made aware of this and will often need to be retrained on the new topology entirely. Otherwise, the AI will not work properly, and may make incorrect decisions or overreact to changes, including its own, which can result in route flapping.
AI for Network Automation and Service Design
Juniper Networks uses AI for detection and protection against attacks, especially for insecure IoT devices. Since the vast majority of IoT devices are insecure, this technology has the potential for a huge impact. In their methodology, devices are grouped together and changes in their behavior are detected to predict attacks.
ADVA focuses in on employing artificial intelligence and machine learning to open up efficient ways to monitor and operate networks. They utilize artificial intelligence to predict network equipment failure with decent success. In order to do so, they use the amount of bit errors produced to identify potential failure.
Proactive Service Restoration and the Role of Machine Learning
Packet Design is concerned with enabling self-driving networks. They use artificial intelligence to optimize the network for link utilization and the rerouting of traffic in the case of link failures.
For all of these instances, artificial intelligence can be applied at different stages of the process:
Detect: The AI searches for notable events and presents them for consideration to human end-operators.
Analyze: After detection, the AI identifies what to do and presents a potential solution.
Apply: The AI is fully autonomous and configures the network appropriately on its own.
The best way to use the AI depends heavily on the use case and how much it can be trusted to perform the task at hand. The overarching opinion at the conference was that artificial intelligence algorithms should start at properly implementing detection, then move on to analyzing, and then applying. Each stage should only be addressed when the previous stage has been established thoroughly, and is reliable both in general and additionally with edge cases.
In addition to the procedure of detection, an AI can also be used for prediction of these events where the stages are similar.
AI Techniques in Networks
With different approaches in artificial intelligence come different techniques. During individual talks, the speakers sometimes mentioned the technique they used. While this does not give a full picture of their work, it at least gives a strong indication of which techniques have been applied successfully to build automated networks.
For detecting or classifying common situations, including heavy hitters, low latency paths, and shift traffic, neural networks are preferred. These include artificial neural networks (ANN), recurrent neural networks (RNN) and convolutional neural networks (CNN).
When the goal is to apply changes to the network state, there are techniques that allow the AI to alter the network settings in a simulated environment before actual deployment. This allows the AI to learn how to change parameters in different situations and achieve better performance. These techniques, also known as supervised or reinforced learning, can be very helpful for complex states. For complex states, deep nets or deep learning is preferred.
Support Vector Machines
When detecting anomalies and outliers, neural networks are not ideal, as the training on low amounts of data which contains the anomaly is not as effective. Instead, support vector machines (SVM) or isolation forests are used.
AI: The Network Solution?
In general, it is important to note that artificial intelligence is not a perfect solution for every issue in networking.
A panel discussion at the conference referenced this and compared rule-based systems to AI approaches. The consensus of the panel was that a rule-based system is the best place to start, and if the network is not exceedingly complex, it may be sufficient. In some instances, it may even be faster and less work-intensive than an AI.
Still, the AI approach is incredibly powerful and can solve complex tasks when handled correctly. When used in conjunction with a rule-based system, the rule-based system can give guidance and boundaries to the AI to prevent excessive errors.
Check out Optimize, our new AI powered product!