Outsourcing Understanding of Telco to AI

By: Vinayak Sharma and David Cherney, Members of Scientific Staff, DISH Wireless

Problem statement:

Telecom in the cloud is enormously complicated; data collection and movement, understanding of that data, and acting on that data are three problems to be tackled. Let’s talk about understanding data. A human could spend a lifetime studying data gathered about 5G network functions running on Kubernetes in the public cloud, the meaning of the myriad features recorded, and not come to a complete understanding of how best to act to tune the network.

To address the problems of moving data, DISH Wireless is implementing a radically transparent data mesh called DISH Data Platform, with game changing storage and accessibility. A DISH project called FONPR (first open network pattern reactor) is addressing how best to take action.

But one can not act on even the most well organized and accessible data unless that data is understood. As Said Berrahil, Vice President of Technology Integration at DISH Wireless, said: “I would like to stress…[the DISH 5G] network cannot be managed by humans.” That is, strategies for access to big data and for data based action need to be complemented with a strategy to automate the understanding of data.

What we did:

The DISH Wireless Pattern Detection team is outsourcing the understanding of telecom networks to machines.

On the data side, DISH’s cloud-native 5G core is running on Elastic Kubernetes Service (EKS) in the AWS Cloud. A combination of custom and AWS built in services record time series data about that core.

On the machine learning side, the team used a neural network structure called a time series autoencoder to learn patterns. In general, such autoencoders are given the task of taking in multivariate time series, creating a smaller number of new variables from them, and then reconstructing the original multivariate time series. This learning is supervised; an autoencoder learns by comparing its reconstruction to the original. Think of it as learning how to compress data and then faithfully decompress. With each comparison it learns to be better at its job. But, and this is the most important part, the model learns the most from the patterns it sees most often; it faithfully decompresses the input data if that data is typical of what it has seen before, and is less faithful in reconstructing unfamiliar patterns. Thus, when an autoencoder is done learning, it is a tool for classifying data patterns as typical or anomalous.

DISH Wireless autoencoder architecture:

As a conceptual example, imagine an autoencoder is trained on video screen captures of black circles being drawn on a white background at constant speed, like in the animated GIF below. The original data might have a thousand frames, each with a thousand pixels, each with red, green and blue pixel values. That is 3 million numbers. An autoencoder might learn to take in this multivariate time series and to create just a few new variables: the radius of the circle, the center of the circle, and the speed with which the circle is drawn. The decoding stage would then take these three variables and re-generate the 3 million numbers. However, the videos that the autoencoder was trained on did not include any videos of animals; if fed a video of a giraffe eating a tree, the autoencoder would try to capture the essence of the video with three variables (radius, position of center, and speed), and then try to use those variable to reconstruct the video of the giraffe. It would do a poor job; it would reconstruct the giraffe as a circle. This poor reconstruction provides information; this video is not like the ones the autoencoder was trained on. It is anomalous.

In particular, the DISH Wireless team trained an autoencoder on features like CPU, memory, and network transmitted bytes used by each network function to find anomalous behavior in the telecom infrastructure.

Results:

The team has a pre-trained model that understands what is typical for one 5G core. That pre-trained model can be deployed to observe any other 5G core. It will be available on an upcoming Data Product Marketplace. Further, that model can be tuned to a new 5G network, transferring its learning of the core it was trained on to the new core.

Also, the team built an open source infrastructure to capture, store, and route data about a 5G core. That data product will also be available on the marketplace. And last, there will be a data product for routing that data through model training environments, and to cataloging versions of the model. Through this combination of products, the team will be able to make the model available to customers who wish to apply it to their private 5G core. At this time, all of the code for the upcoming products is open source and free for use on GitHub.

What’s next:

The three problems, data collection and movement, understanding of that data, and acting on that data, are all three being solved by the team. The next step is to enable 5G core networks to automatically react to learned patterns it observes in itself. DISH Wireless is now building a mechanism to autonomously react to observed patterns in the telco cloud: First Open Network Pattern Reactor (FONPR).

First Open Network Pattern Reactor (FONPR) paradigm:

Using the understanding and infrastructure built from the autoencoder, we aim to have a reinforcement learning agent autonomously understand and take actions upon the EKS clusters running DISH Wireless’ telco network.

Connect with us:

If you are interested in or would like to learn more about how DISH Wireless is engaging engineers and developers, please reach out to DevEx@DISH.com. The DISH Pattern Detection team’s GitHub repository is publicly available. We would love to hear from you and collaborate on our research.

About the Authors:

Vinayak Sharma is a Data Scientist for the DISH Wireless Scientific Staff. He hopes to enable DISH with creative machine learning solutions and is passionate about working on cutting edge research.

David Cherney is Lead Data Analyst for DISH Wireless Scientific Staff. His background is in Mathematical Physics, and his current research interest is AI for massive machine type communication in precision agriculture.