Modern networking infrastructure is extremely complex due to increased network size, volume of traffic, and diversity of devices and applications. Manually configuring these networks has become too time-consuming, error-prone, and difficult. To be effective, network operators need at least scripting-type automation, or better yet, intelligent automation driven by AI that is aware of all aspects a human operator would consider in managing a network and applies the sum of its knowledge to the larger scale and increased complexity.
In this blog post, we’ll take a high-level look at the essential building blocks of an AI-operated self-optimizing network and how they work together.
The objective of a self-optimizing network is to provide its users with stable connectivity and to satisfy the users’ traffic demand such that they can experience the highest service quality. In practice, this human-perceived user experience is very difficult to measure automatically. To build, train, and evaluate the AI, we use substitute metrics that are directly measurable, such as throughput, latency, or resource efficiency. For each use case, we choose a metric – a numeric representation of how well the network is doing and how it compares to other networks.
We identify the factors that influence the chosen metric, and separate them into two types: those that the AI is allowed to control (controllable factors), and those that are assumed as given (environmental factors).
Take the use case of optimizing radio transmission parameters for the access points in a Wi-Fi network. The controllable factors include RF channel bandwidth and RF transmission power levels. The Wi-Fi access point hardware is physically able to operate at various settings of these factors, and we choose to let the AI decide which setting is optimal.
Environmental factors include, for example, the spacing between adjacent access points in the deployment, the propagation characteristics of the RF signals depending on the building materials in the coverage area, and the RF characteristics of the client devices that connect. The AI is not free to modify these factors (and they may vary naturally over time).
Equipped with the definitions of a measurable objective function and its controllable and environmental factors, we instrument the network to collect this data continuously, and report it to the cloud. For each network and each instant in time, the cloud AI is thus aware of how well the network is running, the current setting of the controllable factors, and the current value of the environmental factors.
To evaluate performance fairly, we rank each network against its peer group, which is the set of other networks with the same environmental factors. For example, a network may rank at the 10th percentile of its peers, that is, it performs worse than 90% of its peers. Since the principal difference between the peer networks is the assignment of controllable factors, the AI algorithm can move this network up in the ranking by modifying the controllable factors.
The optimal setting of the controllable factors is determined in one of two modes:
With supervised learning, the past data of all networks is distilled into an AI model that can predict the objective function value of a network with its current environmental factors with any possible setting of the controllable factors. Using these predictions, it is easy to select the best setting among all possible settings. In this mode, the AI exploits the knowledge embedded in the available data in order to optimize each network.
Alternatively, with active learning, the AI is allowed to acquire new knowledge that is not available in past data. The algorithm purposefully and cautiously sets the controllable factors to combinations of values that have never been tried before. In this mode, the AI explores the space of opportunity to identify previously unknown optimization potential for each network.
The key building blocks of an AI-operated self-driving network are a meaningful and measurable objective function, its controllable and environmental influencing factors, a large and diverse collection of data, a definition of peer groups to compare against, and AI models and algorithms that automatically select the optimal assignment of the controllable factors in each operating condition.
In future follow-up blog posts, we’ll shine a spotlight on these building blocks, to understand the challenges in detail and showcase the solutions that are available in Aruba’s products.