AI/ML: The next big step in the telco industry!

In this era of Artificial Intelligence and Machine learning it’s hard to find any industry that is not actively benefiting from it, whether it’s the retail world or the ever so changing telco world.

The telecommunication industry has seen so many advancements in the recent years and is no longer limited to providing basic phone or internet services. Especially as the 5G evolution has brought in the use cases of massive IoT applications, Low Power Wide Area Networks, and Massive MIMO. These use cases require a high level of intelligence in the network, to look for patterns within the data (identifying dynamic change and forecasting the user distribution by analysing historical data), enabling telecom operators to both detect and predict network anomalies, and allowing them to proactively fix problems (like dynamically optimize the weights of antenna elements using the historical data or to improve the coverage in a multi-cell scenario considering the inter-site interference) before customers are negatively impacted, hence creating a self-optimizing and self-updating network.

Along with AI/ML, the concept of disaggregating the hardware and software and creating open interfaces between them is also becoming popular within the telecom industry. When these concepts are applied with the Radio Access Network (RAN), which is the part of the mobile network responsible for wirelessly connecting the devices to the core network, then this disaggregation and use of open interface is called an Open RAN. There are a variety of organizations focused on standardising these concepts, such as the O-RAN Alliance and the Open Networking Foundation.

Open RAN also promotes the easy adaption of AI/ML functionalities into the network with the help of components called the “RAN Intelligent Controllers”, or RIC. A RIC is further split as the Near-Real Time RIC and the Non-Real Time RIC, which have a <1s latency and a >1s latency respectively.

Any ML application deployed within the O-RAN architecture will be found either in the Near-RT RIC (xApps) or in the Non-RT RIC (rApps). The xApps and rApps can either work independently (such that both the training host and the inference host are in the same RIC) or together (such that the training host is in the non-RT RIC and the inference host is in the near-RT RIC).

Here, the ML Training host is responsible for training the model and different approaches can be taken for training a model based on the use case. Some of the training/learning approaches are mentioned below:

• Supervised learning is best for regression and classification problems. Regression is where the ML needs to predict a quantity, whereas classification is where the ML needs to predict a label. Supervised learning also requires the training data to be given with the label already set, so that the algorithm can learn these and check the accuracy of the results produced by the model.
• Unsupervised learning is useful at finding a hidden structure or pattern within the data. For example, this can be useful for separating the data into x amount of groups, where the algorithm will decide what the distinguishing factor between the groups is.
• Reinforcement learning is used for more complex problems, where a sequence of decisions will need to be made. The way this learning is performed is comparable to a game, where the algorithm is trying to maximise the total reward. Each decision will come with a reward, either positive or negative, and once the model is trained, the algorithm should know the best sequence of decisions which gives the highest points.

A model is ready to be used by the inference host once it has achieved a good accuracy such that any further training will not improve the results.

The ML Inference host will use the model created by the ML Training host to decide how to act next, based on the live data it’s receiving from the RAN. In order words, this live data from the RAN will be given to the model, which will output the best course of action. Whether the inference host is found in the Near-RT RIC or in the Non-RT RIC, it will send the desired action back down to the RAN components to follow through with this.

Based on these principles there are already various operators and vendors actively developing the xApps like the:

• KPI Monitoring (KPIMON) xApp: Gathers the radio and system Key Performance Indicators (KPI) metrics from E2 Nodes and stores them in the Shared Data Layer (SDL).
• Anomaly Detection (AD) xApp: Fetches UE data regularly from SDL, monitors UE metrics and sends the anomalous UEs to Traffic Steering xApp.
• Traffic Steering (TS) xApp: Consumes A1 Policy Intent, listens for badly performing UEs, sends prediction requests to QP Driver, and listens for messages that show UE throughput predictions in different cells to make a decision about UE Handover.
• Quality of Experience (QoE) Prediction Driver (QP Driver) xApp: Generates a feature set of metrics to input to QoE Prediction, based on SDL lookups in UE-Metric and Cell-Metric namespaces.
• QoE Prediction (QP) xApp: Receives a feature set of metrics for a given UE, and output Throughput predictions on the Serving and any Neighbor cells to Traffic Steering xApps.

Together these xApps implement a use case where anomaly detection is combined with QoE prediction and traffic steering action to move the affected UEs to a different cell. Similar use cases for SLA assurance are also being worked on by the operators, where the rApps are capable of managing and orchestrating each slice autonomously, in order to prevent or reduce network congestion and its impact across the network, ultimately delivering more varied services.

In conclusion, this blog has briefly touched upon the role AI/ML plays in the telecom world and how it can accelerate the time to market for the offerings promised by 5G (as shown in the figure above). Currently Reply is also actively engaged in creating Proof of Concepts (PoC) and developing 5G solutions catered to its clients’ infrastructure, while also contributing to the open-source communities like O-RAN Alliance and SD-RAN.

If you have any questions or would like to understand how Net Reply can help you with this or similar solutions, get in touch with Jasmine Dixit and Thayná Dorneles.