In November the Forum will publish an extensive Trend Analysis Report on AI and machine learning, analyzing the results of surveys of CSPs and suppliers (choose the right one for you). Take the survey and make your opinion count. As a thank you, you’ll be entered into a draw for an Apple Watch.
I remember going on a school trip to see my first computer. We were in awe of how large it was and had fun playing noughts and crosses (tic-tac-toe) on a teleprinter connected to the massive computer room. At university, I began programming computers in Fortran with punch cards and paper tape. Later I built my first home computer, amazed to have access to this technology at home.
Things have really accelerated since then, but one thing has remained the same: Humans write the programs and give computers the data they need to generate outputs. This is beginning to change, however, with advances in machine leaning now making it possible for computers to program themselves.
The effects of this artificial intelligence (AI) revolution will be great. Many believe it will be much more profound than the industrial revolution. Some even believe it is an existential threat.
Without a doubt, AI and machine learning will have a huge impact on the communications networks of the future. Both will absolutely be required not only to control and optimize networks but also the IT support systems underpinning them.
What is machine learning?
AI refers to the capability of a machine to imitate human behavior. Machine learning, which evolved from the study of pattern recognition and computational learning theory in AI, explores construction of algorithms that can learn from and make predictions about data.
Machine learning needs data. In fact, you could say data is the ‘currency’ of machine learning.
In traditional programming, a human writes a computer program and provides the data, which the computer processes to create the output. In machine learning, humans provide the data along with the desired output, rules and constraints, and the computer writes the program to deliver this.
Types of machine learning
There are many types of machine learning. Here’s a brief look at three of the most common:
- Reinforcement learning uses the concept of a reward and feedback of state. The role of the machine is to maximize the reward by adjusting the state of the environment it is controlling. The reward is generated by an interpreter which looks at optimal target outcomes. The key thing here is that the machine has no real knowledge of the actual outcome it is trying to maximize; it only has knowledge of the reward and the state. There are analogies here with Pavlov’s dogs.
- In supervised leaning, humans provide feedback to the machine on which outputs are correct and which are not. Imagine feeding multi-colored cards into a machine and asking the machine to recognize only the red cards. We ‘tell’ the machine each time it gets the recognition right or wrong, and the machine uses this feedback to constantly adjust until the recognition becomes accurate.
- Unsupervised leaning is similar to supervised except that humans are out of the picture. The machine uses clustering techniques to look for patterns in the data. I recently looked at a start-up company in the UK that is using unsupervised machine learning to detect anomalies in the behavior of enterprise networks. For example, if there are unusual large data flows or if a PC in the HR department is constantly accessing a development server, it flags this up. The company’s aim is to move from simple detection to remediation over time.
Machine leaning and communications networks
One of the biggest challenges when applying machine learning for network operation and control is that networks are inherently distributed systems, where each node (for example, switch, router, etc.) has only a partial view and partial control over the complete system. Learning from nodes that can only view and act over a small portion of the system is difficult and complex, particularly if the end goal is to exercise control beyond the local domain.
The emerging trend towards logical centralization of control (via software-defined networking – SDN) will ease the complexity of learning in an inherently distributed environment.
Why is this? SDN and network functions virtualization (NFV) are becoming well established as technology platforms for future communication networks. Using these technologies, networks will be highly dynamic. Potentially thousands or more virtual machines (VMs) will carry out network functions, and new VMs will be deployed and removed in seconds.
This will make it impossible for humans to optimize and control such networks at the super human speeds required; machine learning will be necessary, not only for network control and optimization but in other areas of communications service providers’ (CSPs’) businesses as well.
The knowledge-defined network
David Clark, a Senior Research Scientist at the MIT Computer Science and Artificial Intelligence Laboratory, and others are proposing a Knowledge Plane for the Internet, a new construct that relies on machine learning and cognitive techniques to operate the network.
This knowledge-defined network (KDN) operates by means of a control loop to provide automation, recommendation, optimization, validation and estimation. I believe the KDN will be a vital part of future combinations of network infrastructures.
Where are CSPs using AI and machine learning?
CSPs are beginning to use AI and machine learning in three areas:
- Customer experience management;
- Service management and optimization; and
- Network management and optimization.
For example, operators are using virtual agents and chatbots to improve customer experience. When it comes to network and service management in the Internet of Everything (IoE), both must be zero-touch because it isn’t feasible to support the volume and velocity of changes that must happen in a software-defined network made up of millions of nodes running thousands of applications.
Management must happen at multiple levels, from the physical network to logical resources to services to the customer or IoE device, and it must extend beyond a network operator’s borders to include partners’ networks and services. To accomplish this, operators are beginning to use intent-based management that relies on orchestration, data analytics, policy and machine learning to autonomically provision, configure and assure their networks and the services they deliver to customers.
Analytics are critical for their role in automating closed control loops, which help operators guarantee and optimize performance and reduce operating costs. The massive amount of data that operators collect about performance, traffic patterns and customers’ locations, preferences and history also can be used to sell additional services and create new ones, which leads to increased revenue.
How can TM Forum help?
TM Forum has been at the heart of operational and business system support systems (OSS/BSS) standardization and innovation for many years. The Frameworx suite of best practices and standards, which provide the blueprint for effective business operations, have become the common language for operators around the world. Key components of Frameworx include the Business Process Framework (eTOM), the Information Framework (SID) and the Application Framework (TAM).
With its vibrant collaboration community building on the Frameworx heritage, the Forum is ideally positioned to become a leader in the development of the KDN and future OSS/BSS to support it. As a result of recent workshops, several of TM Forum’s senior CSP members are proposing a new architectural vision for service providers to replace traditional OSS and BSS. We refer to this new architecture as the Open Digital-business Enablement System (ODES).
The ODES essentially is a new operating system for service providers. This single evolutionary architecture will deliver all the capabilities previously siloed in independent BSS and OSS. It removes the barriers between the two, which increases flexibility and lowers costs, and helps CSPs deliver a truly digital customer experience end-to-end. This approach supports the requirements of future agile, real-time, service provider infrastructures based on SDN and NFV.
The architecture is data-centric, meaning a consistent and holistic data architecture runs end to end though the various layers of the stack. The KDN will be connected into this data architecture, as well as into other layers, so that AI-driven intents can be fed into the architecture in line with the closed-loop model.
I will be presenting on AI/machine learning and the ODES during a day-long workshop at TM Forum Action Week Sept. 25-29. Join us in Vancouver to discuss the future of OSS/BSS and provide your input into the evolving architecture.