Tuncay Cil and Haleema Mehmood

Posted by Tuncay Cil and Haleema Mehmood


Tuncay Cil and Haleema Mehmood has published 2 blog posts.

ASSIA’s Approach to Machine Learning

Part 2: Machine Learning improves Wi-Fi Quality of Experience

In the previous Machine Learning Primer article (published as a blog on ASSIA’s web site on May 30, 2018), we tried to give a short overview of basic AI/Machine learning definitions and the associated challenges.  In this article, we aim to provide some real-world examples of how ASSIA applies these AI/machine learning concepts on its CloudCheck® product line.

CloudCheck uses AI and machine learning techniques to solve a wide range of Wi-Fi related operational problems including AP (Access Point) level diagnostics, optimization, and network-wide trends and predictions.

AI and Machine Learning Examples:

Here we outline a step-by step process that tracks two example applications – Client Steering and Temperature Detection –  at the AP level:

  • Measure Health, Environment and Performance: CloudCheck measures the AP’s health, its environment and the performance of connected clients periodically. The period of observation varies by the type of data being collected, ensuring sufficient coverage of the observed phenomenon. For example, the RSSI of connected clients, which can show high variation with time, is measured at a small time-scale, while the temperature of the CPU, which remains largely constant for long time periods, is collected at a larger time-scale.
  • Diagnose: CloudCheck analyzes the data to determine trends and patterns at various levels. For example, for each client, a mapping of RSSI and throughput is established in both 2.4GHz and 5GHz frequency bands. In contrast, the temperature measurement is aggregated by the device model of the AP to determine the temperature distributions of different chipsets.
  • Set Thresholds and Policies: CloudCheck uses the detected patterns to establish thresholds or policies for diagnostics and optimization. For example, for each client, CloudCheck individually determines the best RSSI to transition from the 5GHz to 2.4GHz band, and the best RSSI to transition from the 2.4GHz to 5GHz band, optimizing for both throughput and coverage. For CPU temperature distributions, CloudCheck determines thresholds for finding outlying or problem temperatures.
  • Monitor Deviations: CloudCheck uses these policies to monitor deviations from the defined behavior. For Wi-Fi clients, any RSSI transition that crosses the established threshold for the particular band results in a real-time alert. For temperature distribution, daily temperature data collected for each AP is compared with the appropriate model-specific threshold.
  • Recommend or Optimize: The last step of the process is either an optimization action, or a diagnostic output combined with a trouble-shooting recommendation. Just like the period of observation, the output/response time is in accordance with the need of the application. In the client RSSI example, clients requiring band-steering are steered in near real-time to the appropriate band. For temperature, alerts are raised on a daily basis for each AP if required.

As more data is collected, CloudCheck is able to learn from the incoming data, improving the performance of the diagnostics and optimization actions over time.

QoE Assessment

Another application of machine learning techniques within CloudCheck is the prediction of user Quality of Experience (QoE). A supervised learning approach is employed for this purpose, using network-wide statistics:

  • Training data: Supervised learning for QoE requires a labeled data set for training. The features in this data come from CloudCheck’s daily AP diagnostics. The diagnostics are assessments of AP performance along individual dimensions/features including but not limited to client coverage and Wi-Fi throughput, interference in various bands, broadband disconnections and throughput, etc. The labels for the training data come from an assessment of user experience either based on customer surveys, or customer call/churn data provided by the service provider.
  • QoE Model Training: Supervised machine learning techniques are used on this training data-set to build a model that maps CloudCheck diagnostics to user QoE. This diagnostic is applied for daily assessment of customer QoE.
  • QoE Determined: Based on the needs of the Service Provider, this model can be used either for predictive purposes or for trouble-shooting. CloudCheck further employs feature selection techniques on the training data-set to identify the key drivers of customer dissatisfaction in the service provider’s network.

Over its 15 year history, ASSIA has pioneered development of AI/machine learning applications for last-mile and home network management challenges.  From their inception, ASSIA’s DSL, GPON and Wi-Fi solutions have utilized AI and machine learning to enable many of the world’s most progressive carriers and service providers to manage well over a hundred million internet connections.  Stay tuned for more innovation in AI at ASSIA in the future!

ASSIA’s Approach to Machine Learning/Artificial Intelligence

Part 1: A Primer


Arthur Samuel in 1959 defined machine learning as the “field of study that gives computers the ability to learn without being explicitly programmed”. Typically, machine learning and artificial intelligence (AI) are used interchangeably in the industry although AI is a superset concept that is not just limited to machine learning.  Only recently, the advances in cost, processing power and storage on devices and network equipment, have delivered the possibility of AI techniques previously only available for super computer clusters.

In a typical AI/machine learning system for network management, the following steps form the foundations of any prescriptions or predictions that get generated by it:

  • Start with observing a user or a system
  • Describe, categorize and make sense of the observation
  • Attempt to define the behavior of the user or system by establishing a set of baselines
  • Monitor and detect deviations from baselines of defined behaviors
  • Predict future behavior, prescribe trouble-shooting solutions, and self-optimize

Applying these AI techniques to better monitor, detect, troubleshoot and optimize Wi-Fi networks, is part of ASSIA’s core engineering capabilities. While the feasibility of real-world applications of AI techniques is there, the data scientist generalists approach falls short of delivering tangible value to service providers. This is where ASSIA brings real-world performance and value accepted by service providers to the field of AI driven network management.

Basic vocabulary for machine-learning/AI

Data is the input to the overall system and in the context of Wi-Fi management can be anything from physical layer performance metrics, device characteristics and environment, alerts end-user inputs, and Wi-Fi client performance indicators.

A feature is an individual data element relevant to the specific machine learning model and use case. It can be directly extracted from the data or derived after passing the raw data through a pre-processing algorithm. For example, Wi-Fi throughput between an access point and a client can be derived from traffic counters.

A model maps the input features to the desired output or label.  In Wi-Fi management, an output may be the best channel to pick in the user’s environment or the user’s perceived quality of experience score.

Labels augment individual unlabeled data with meaningful tags that are informative. Machine learning models can be applied to the data so that new unlabeled data can be presented to the model and a likely label can be guessed or predicted for that piece of unlabeled data.

Supervised machine learning is a “teaching” technique to develop a relationship between a known set of outputs (labels) and their inputs (features). Once the model is developed, it can be used to predict the output for a new set of inputs.

Training prepares a supervised machine learning model using features extracted from training data along with the appropriate labels. Training can be performed periodically with the update of training data or features.

Unsupervised machine learning is “self-learning,” and no prior training, or preparation is required before it’s deployed. The algorithm automatically discovers relevant structure and relationships among the inputs.

Baselining is a method to determine “expected” behavior for an entity like a user or a system. Once these baselines are established, the models look for deviations from the baselines.

How ASSIA makes machine learning/AI possible

ASSIA’s CloudCheck solution makes AI possible using its agent-cloud architecture. The agent resides on the home Wi-Fi router, collecting the data necessary for meaningful AI. The data is processed and stored in the cloud, allowing for powerful machine learning/AI applications.

For many years, three main issues hindered proper data collection necessary for meaningful cloud-based AI applications in last-mile and Wi-Fi network management:

  • Non-exhaustive platform coverage:There are a variety of CPEs with various chipsets in use in service providers’ networks. It is important to collect data from all types of nodes for network level machine learning. ASSIA solves this problem with its hardware-agnostic CloudCheck agent that provides similar data collection across more than 100 CPE models.
  • Insufficient sampling across time, and network nodes:Overnight data collection that is prevalent in the industry is insufficient for identifying problems in network nodes. Furthermore, it is gravely inadequate for machine learning applications. CloudCheck solves this problem by providing a much higher frequency of data collection. This data is stored and transported efficiently so as to minimize network load.
  • Data collection methods that don’t scale:While trying to collect more data, more frequently, in a more intelligent way, we can’t (yet) assume unlimited computational resources. CloudCheck uses a smart way of mixing event and time triggered data collections that takes into account the dynamic nature of the observed trouble patterns. This resolves the scalability problem of the growing data collection needs without missing critical information.

ASSIA has developed patented data collection techniques that enable application of AI-based algorithms in the real-world environment.

ASSIA’s CloudCheck solution makes use of this data in unsupervised ways to learn baseline behaviors along various dimensions. The learnt policies or thresholds are either fed back to the CloudCheck agent, so it can report deviations from the thresholds to trigger an optimization action, or the policies are used by server algorithms to generate diagnostic outputs. The diagnostic outputs are further used in a supervised learning model that assesses customer satisfaction.

In the next article we will explore specific applications of how ASSIA applies AI and machine learning in Wi-Fi Management.