AI/ML in 5G RAN protocol stack
- Venkateshu
- 11 hours ago
- 18 min read
Introduction
Artificial Intelligence (AI) and Machine Learning (ML) are becoming integral to the evolution of 5G Radio Access Networks (RAN), transitioning them from static, rule-based systems to dynamic, self-optimizing infrastructures. The immense complexity, stringent performance demands, and diverse service requirements of 5G and beyond necessitate intelligent automation. AI/ML models are being embedded directly into the RAN protocol stack to enhance efficiency, reduce latency, and improve user experience in real-time.
The 3rd Generation Partnership Project (3GPP) has been actively studying and standardizing the integration of AI/ML into the network architecture, particularly from Release 18 onwards, signalling a shift towards an AI-native framework for future wireless systems like 6G. A framework for RAN intelligence has been described in 3GPP TR 37.817, as shown in below.
Core Components
Data Collection: This is the starting point. This component is responsible for gathering all raw data from the environment. This data can come from various sources, such as user interactions, network sensors, or the outcomes of previous actions.
Model Training: This is where the "learning" happens. A subset of the collected data, known as Training Data, is used to build or refine a machine learning model. The goal of this stage is to create a model that can recognize patterns and make accurate predictions.
Model Inference: This is the "live" or "production" stage where the trained model is put to use. It takes in new, real-time data (Inference Data) and uses its learned patterns to produce an Output, which is typically a prediction, classification, or a decision.
Actor: This is any system, application, or process that takes the model's Output and performs an action based on it.
The core of Release 18's contribution is the standardization of data collection and signaling support for three specific, high-impact use cases :
Network Energy Saving: AI models can analyze traffic load, user distribution, and other Key Performance Metrics (KPMs) to predict periods of low demand. This allows the network to intelligently switch off cell carriers or other components to reduce power consumption without impacting user experience. Release 18 standardizes how a gNodeB (gNB) collects and reports the necessary data for these predictive models to function effectively.
Load Balancing: To prevent congestion and optimize resource utilization, AI/ML algorithms can predict traffic hotspots and proactively shift user load between different cells or frequency bands. Release 18 defines the procedures for gNBs to exchange the required data, such as cell load and resource block utilization, to enable intelligent load-balancing decisions.
Mobility Optimization: For use cases like high-speed mobility in vehicles, AI can enhance handover reliability by predicting user trajectory and selecting the optimal target cell based on more than just signal strength. Release 18 provides the signaling support to collect and share the rich data needed for these advanced mobility algorithms, including information from neighboring network nodes.
Key Architectural and Interface Enhancements
To enable these use cases, Release 18 introduces crucial enhancements to the NG-RAN architecture and its interfaces, which directly impacts how AI/ML can be deployed:
Standardized Signaling over Xn Interface: A major step is the specification of procedures for the Xn interface (the interface between gNBs). This allows neighboring gNBs to formally request, configure, and exchange the specific datasets needed for AI/ML model inference. The framework includes the "Data Collection Reporting Initiation" and "Data Collection Reporting" procedures, creating an interoperable way for network nodes to cooperate on AI-driven tasks.
Flexible Functional Placement: The specifications are designed to support different deployment architectures. AI/ML models for training and inference can be located in various places: training could occur in the Operations, Administration, and Maintenance (OAM) system, with inference running in the gNB, or both could be co-located within the gNB itself (or the gNB-CU in a split architecture). This provides flexibility for network implementation.
Foundation for AI/ML Model Lifecycle Management: By standardizing data collection and management functions, Release 18 supports the broader AI/ML operational workflow. This includes data acquisition for model training, deploying trained models to the target network function for inference, and monitoring performance, laying the groundwork for a complete MLOps framework in the RAN
Types of Machine Learning Algorithms in 5G RAN
Different ML paradigms are suited for various RAN optimization tasks:
Supervised Learning: Used when labeled data is available. For example, training a CNN on a dataset of radio measurements and corresponding optimal beams for beam selection.
Unsupervised Learning: Useful for finding hidden patterns in data, such as using clustering algorithms for pseudo-localization of users based on signal characteristics.
Reinforcement Learning (RL): Ideal for real-time decision-making in dynamic environments. An RL agent learns by interacting with the network (the "environment") and receiving rewards or penalties for its actions. This is highly effective for MAC scheduling and real-time resource allocation.
Deep Learning: A subfield of ML using neural networks with many layers (e.g., CNNs, LSTMs). It excels at learning complex patterns from vast amounts of data, making it suitable for CSI prediction, channel estimation, and beam management.
Generative AI: Models like Generative Adversarial Networks (GANs) are being explored for creating synthetic but realistic channel models for training and testing other AI systems.
AI/ML across the 5G RAN Protocol Stack
AI/ML functionalities are being applied across various layers of the gNodeB (gNB) and User Equipment (UE) protocol stack, from the physical layer to the upper layers of the user plane.
In5G new radio (NR),the 5G layer1 (L1) is physical layer.The 5G layer2 (L2) is data link layer consisting of Medium Access Control (MAC), Radio Link Control (RLC) and Packet Data Convergence Protocol (PDCP).
Physical (PHY) Layer
The PHY layer is responsible for the transmission and reception of data over the air interface. In radio access networks, the base band processing in physical layer runs on an user equipment and a base station. Although the computational power of the user equipment increases in every generation, the workload of the base band processing is already heavy at the user equipment and it would not be efficient to run AI algorithms on user equipment. Thus, the base band processing at base stations would be suitable for adopting AI algorithms to physical layer. In order to run AI algorithms, the intelligent application with high computations will be deployed at the edge of radio networks.
AI/ML introduces significant enhancements to its core functions:
Beam Management: Selecting the optimal transmission and reception beams is a complex challenge in 5G's high-frequency bands. Deep learning algorithms can analyse radio conditions and user location data to predict the best beam pairs, significantly improving signal quality and reliability, particularly for users at the cell edge.
The primary use cases for ML in beam management can be categorized into prediction, optimization, and contextual awareness.
Predictive Beam Management
The core of ML's contribution is its ability to predict the optimal beam configuration without performing an exhaustive, brute-force search. This predictive capability addresses the significant overhead associated with legacy beam-sweeping procedures.
Time-Domain Beam Prediction (TBP): This is one of the most critical use cases, especially for mobile users. ML models, particularly Recurrent Neural Networks (RNNs) and LSTMs, are trained on a sequence of historical beam measurements. They learn the temporal patterns of how the best beam changes as a user moves, allowing the model to forecast the optimal beam for future time instances. This is crucial for maintaining a stable link for users in vehicles or other high-mobility scenarios, as it allows the network to switch to the correct beam proactively.
Spatial-Domain Beam Prediction (SBP): Instead of relying on time-series data, SBP uses spatial information to predict the best beam. An ML model can be trained to correlate a User Equipment's (UE) geographical location, angle of arrival (AoA), or other spatial features with the optimal transmit and receive beams. This allows the gNB to infer the best beam for a UE based on its position, significantly reducing the need for initial beam sweeping. This is particularly useful for initial access or when a UE becomes active after a period of inactivity.
The process is divided into two distinct phases:
1. Training Phase (Offline)
This is the preparatory phase where the AI model learns how to make predictions.
Provide Historical Data: The gNodeB (gNB) gathers a large amount of historical data. This data includes past user movements (UE positions) and which beams provided the best signal quality at those positions (beam reports).
Train Predictive Model: This historical data is fed into an AI/ML Engine. The engine uses this data to train a predictive model. The diagram notes an example of using an LSTM (Long Short-Term Memory) model, which is excellent for sequence prediction tasks like forecasting a user's trajectory. The goal is to create a model that, given a user's recent movement history, can accurately predict their future location and the corresponding optimal beam.
2. Inference Phase (Real-time)
This is the live operational phase where the trained model is used to manage a connection in real time.
UE Sends Live Data: The User Equipment (UE) sends uplink signals and periodic reports to the gNB. This provides real-time information about its current location and signal quality.
Model Predicts Future Beam: The gNB takes this live data (such as the current location and the sequence of beams used recently) and inputs it into the trained AI/ML model.
The model makes a prediction, forecasting the future optimal beam(s) for the user. It essentially anticipates where the user is moving next.
Apply Predicted Beam: The AI engine returns the Predicted Beam Index to the gNB. The gNB then proactively selects and applies this beam for its next transmission.
Transmit Data: The gNB transmits data to the UE using this predicted beam, ensuring the signal is already aimed at the UE's next location.
Optimization of Beam Configuration
Beyond prediction, ML can optimize various aspects of the beamforming process itself.
Super-Resolution Beam Prediction: This technique aims to reduce beam training overhead. Instead of sweeping through all narrow, high-gain beams, the system can first perform a sweep with wider, lower-resolution beams. An ML model, trained on this sparse data, can then infer or "super-resolve" which narrow beams within the selected wide beam are likely to be optimal. This allows the system to perform a much more targeted, fine-grained search, saving significant time and resources.
Beam Selection Optimization: In complex scenarios with many potential beam pairs, ML can solve the beam selection problem more efficiently than traditional exhaustive searches. A model can be trained to take various inputs (e.g., UE location, surrounding obstacles, signal-to-noise ratio) and directly output a ranked list of the most promising beam pairs. This is often framed as a classification or regression problem, where the model learns the mapping from environmental context to the best beam choice.
Context-Aware Beam Management
ML enables the beam management system to become aware of and adapt to the broader network context.
Blockage Prediction: Millimeter-wave (mmWave) signals are highly susceptible to blockage from obstacles like buildings, vehicles, or even people. By using inputs from sensors (like LiDAR or cameras on a vehicle) or analyzing historical radio data, an ML model can predict an imminent blockage event. This allows the gNB to proactively initiate a handover to a different cell or switch to a non-line-of-sight (NLOS) beam path before the link is dropped, ensuring service continuity.
Dynamic Beamforming Adjustment: In a multi-user MIMO (MU-MIMO) environment, the optimal beams for one user depend on the beams being used for other users to minimize interference. An ML model, such as a deep reinforcement learning agent, can analyze the real-time spatial distribution of all users and dynamically adjust the beamforming parameters for the entire cell. This holistic approach optimizes overall cell throughput and reduces inter-user interference, a task that is computationally prohibitive for traditional algorithms.
2. Channel State Information (CSI) Feedback: In Massive MIMO systems, obtaining accurate CSI is critical but generates significant overhead. Deep learning models, such as Convolutional Neural Networks (CNNs) and Long Short-Term Memory (LSTM) networks, can predict CSI more accurately with less pilot overhead. This leads to improved link robustness, especially in highly mobile and dense urban environment.
3. Channel Estimation: AI models can implicitly capture the complex characteristics of wireless channels, including in novel scenarios like those involving Reconfigurable Intelligent Surfaces (RIS). This allows for more accurate channel estimation than traditional methods, leading to gains in diversity and multiplexing.
In a 5G system, particularly in a Time Division Duplex (TDD) configuration, there is an unavoidable delay between the moment the gNB measures the channel (via sounding reference signals) and the moment it uses that CSI to schedule data and perform beamforming. This delay is typically a few milliseconds.
In slow-fading channels (e.g., a stationary user), the channel state doesn't change much during this delay, so the measured CSI remains a valid representation of the channel.
In fast-fading channels (e.g., a user in a high-speed vehicle), the radio channel's characteristics can change significantly within milliseconds. By the time the gNB applies the CSI for transmission, the channel has already evolved, rendering the once-accurate CSI outdated or "aged."
This channel aging leads to a mismatch between the beamforming applied by the gNB and the actual state of the channel, causing degraded Multi-User MIMO performance, increased interference, and reduced overall system throughput. The traditional approach is to simply use the last known CSI until a new measurement arrives (a method known as Zero-Order Hold), which does not solve the aging problem.
The ML Solution: Predictive CSI
Instead of relying on outdated measurements, ML models can forecast the future state of the channel. The core idea is to treat the dynamic wireless channel as a complex time series. By analyzing a sequence of past CSI measurements, an ML model can learn the underlying temporal patterns of how the channel evolves and predict what the CSI will be in the near future, effectively counteracting the system's inherent delay.
This turns CSI acquisition into a regression problem, where the model learns a mapping from historical channel data to future channel data.
The diagram is divided into two main phases:
1. Training Phase (Offline)
This is the preparatory step where the AI model learns the behavior of the radio channel.
Provide Historical Data: The gNB provides the AI/ML Engine with large datasets of Historical CSI Data Sequences. These are time-stamped records of how the channel quality has changed over time in various conditions.
Train Predictive Model: The AI/ML Engine uses this data to train a predictive model (the diagram suggests an LSTM, which is ideal for time-series forecasting). The model learns the complex temporal patterns of channel fluctuations.
2. Inference Phase (Real-Time Operation)
This is the live process where the trained model is used to make real-time predictions.
UE Sends Reference Signal: At a specific moment, time 't', the User Equipment (UE) sends a Sounding Reference Signal (SRS). This is a known signal used for channel measurement.
Estimate Current CSI: The gNB receives this SRS and estimates the current state of the channel, denoted as CSI(t).
Input for Prediction: The gNB sends a sequence of its most recent CSI estimates (e.g., CSI(t), CSI(t-1), etc.) to the AI/ML Engine.
Predict Future CSI: The AI model performs inference. It analyzes the sequence of recent channel states and predicts what the CSI will be at a future time 't + Δt'. Here, Δt represents the system's internal processing latency—the exact amount of time the channel is expected to "age."
Use Predicted CSI: The model returns this Predicted CSI to the gNB. The gNB now has a highly accurate estimate of what the channel will look like at the exact moment of the upcoming data transmission. It uses this future-proof CSI value for its resource scheduling and beamforming calculations.
Transmit Data: Finally, at time 't + Δt', the gNB transmits data to the UE using the optimized settings. Because the settings were based on the predicted CSI, they are perfectly matched to the actual channel conditions at the moment of transmission.
Medium Access Control (MAC) Layer
The MAC layer schedules radio resources among users. AI/ML brings predictive and adaptive capabilities to this critical function:
Intelligent Scheduling: Traditional schedulers follow predefined rules. In contrast, schedulers based on Reinforcement Learning (RL) can learn complex traffic patterns and application requirements on the fly. An RL agent can continuously monitor network conditions and user demands to dynamically adjust scheduling decisions, optimizing for latency, throughput, or fairness based on real-time needs.
Step-by-Step Flow Explanation
The process is a continuous learning loop, which is the core concept of Reinforcement Learning.
Users Report Their Needs: Multiple UEs send Buffer Status Reports (BSRs) to the gNB. These reports tell the scheduler how much data each user has waiting to send and imply their service requirements.
State Observation: The MAC Scheduler gathers all relevant information about the current network situation, which is known as the "State." This includes:
The buffer status from all UEs.
The radio channel quality for each user (CQI).
Overall network load and historical data traffic patterns.
It then feeds this complete State information to the RL Agent.
Action Selection (The Decision): The RL Agent observes the current State and, based on its learned "policy," selects an "Action." In this context, the Action is the actual scheduling decision: how to allocate the available time and frequency radio resources among the competing users. The goal is to choose the action that it predicts will yield the best outcome based on past experience.
Applying the Decision: The RL Agent passes this decision back to the MAC Scheduler in the form of Scheduling Grants. The scheduler executes this decision, allocating the radio resources to UE1 and UE2 according to the agent's intelligent plan.
Data Transmission and Outcome: The UEs use their allocated resources to transmit data.
Reward Calculation (The Feedback): This is the most crucial step for learning. The MAC Scheduler evaluates the outcome of its last decision. Did the action lead to a good result? This evaluation is quantified as a "Reward" signal, which is sent back to the RL Agent. For example:
Positive Reward: If the latency-sensitive user's data was sent on time, or if the throughput-hungry user got a high data rate.
Negative Reward (Penalty): If a user's latency target was missed or if resources were allocated inefficiently, resulting in unfairness.
Policy Update (The Learning): The RL Agent receives this Reward. The reward tells the agent whether the action it took in that specific state was good or bad. It uses this feedback to update its internal policy. Over time, through trial and error across thousands of these cycles, the agent learns to associate states with actions that consistently maximize the cumulative reward. It gets progressively better at making scheduling decisions that optimize for the desired KPIs like latency, throughput, and fairness.
Adaptive Modulation and Coding (AMC): AI models can predict link quality with high accuracy, enabling the gNB to select the optimal MCS for each user. This maximizes spectral efficiency while ensuring link reliability, adapting instantly to changing channel conditions.
Step-by-Step Workflow Explanation
1. UE Reports Channel Quality: The UE continuously measures the quality of its radio connection and reports this back to the gNB in the form of Channel State Information (CSI). This includes metrics like the Channel Quality Indicator (CQI).
2. AI Model Gathers Inputs: Instead of just using the instantaneous CQI report, the gNB feeds a much richer set of data to the AI/ML Link Quality Predictor. This includes:
The real-time CSI from the UE.
Historical link quality data for that user and location.
User mobility information (is the user stationary or moving quickly?).
Statistics from previous transmissions (e.g., success/failure rates).
3. AI Prediction Phase: The AI model analyzes all these inputs to make a highly accurate prediction of the future link quality. This is the key advantage over traditional methods; it doesn't just react to the current state but anticipates what the channel will look like in the immediate future.
4. MCS Recommendation: Based on its prediction, the AI model recommends the optimal MCS Index. This index corresponds to the perfect balance between maximizing data throughput (speed) and ensuring the link remains stable and error-free.
5. Selection and Transmission: The gNB receives this recommendation and uses it to select the MCS for the next data transmission to the UE.
6. Feedback Loop for Learning: After the data is sent, the UE reports back whether it received the data correctly (ACK) or if there was an error (NACK). This ACK/NACK feedback is fed back into the AI model. This critical feedback loop allows the model to learn from the outcome of its decisions, continuously refining its predictions to become even more accurate over time.
Radio Link Control (RLC) Layer
The RLC layer is responsible for data segmentation, reassembly, and ensuring reliable data transfer. A key use case for AI/ML here is congestion control.
Active Queue Management (AQM): AI-driven AQM algorithms can prevent bufferbloat and reduce latency. For instance, the Dynamic RLC Queue Limit (DRQL) algorithm uses communication between the SDAP and RLC layers. An ML model can predict impending congestion by analyzing RLC queue lengths and incoming data rates from the SDAP layer, and then dynamically adjust the RLC queue limits to manage data flow proactively. This prevents packet drops and maintains low latency for sensitive applications.
Below is a sequence diagram illustrating an AI-driven AQM process involving a Near-Real-Time RAN Intelligent Controller (Near-RT RIC).
Step-by-Step Explanation
Data Transmission Begins: The User Equipment (UE) sends uplink data to the base station, which is first received by the gNB-DU. The gNB-DU then forwards these data packets to the gNB-CU-UP for higher-layer processing.
AI/ML Congestion Control Loop: This is the core of the process where the intelligence lies.
Reporting Queue Status: The gNB-DU constantly monitors the status of its RLC queues (where data waits before being sent over the air). It reports this status, such as queue length and packet arrival rate, as Key Performance Metrics (KPMs) to the Near-RT RIC.
AI Analysis and Prediction: The Near-RT RIC feeds this real-time data into its ML model. The model analyzes the trends and predicts potential congestion before it actually happens. For example, if it detects that a queue is filling up too quickly, it can anticipate a future bottleneck.
Sending Control Commands: Based on this prediction, the RIC sends a Control Message back to the gNB-CU-UP. A key command is to "Adjust Flow Rate."
Proactive Adjustments:
The gNB-CU-UP receives the command and adjusts the packet forwarding rate down to the RLC layer in the gNB-DU. It acts like a smart valve, slowing the flow of data to prevent the radio queue from overflowing.
Simultaneously, the gNB-DU can use this intelligence to dynamically adjust its RLC queue limits, further optimizing resource use and preventing packet drops.
Radio Resource Control (RRC) Layer
The RRC layer manages the control plane signaling, including mobility.
Handover Optimization: In dense, heterogeneous networks (HetNets) with overlapping macro cells and small cells, deciding the optimal handover target is complex. AI models can analyze a wide range of inputs—such as user mobility patterns, cell load, and application type—to predict the best handover decision. This ensures seamless connectivity and balances load across the network, improving Quality of Experience (QoE) for users in vehicles or at cell boundaries.
Measurement Reporting: The process begins when the UE, which is mobile (e.g., in a car or walking near a cell edge), sends Measurement Reports to the Source base station it is connected to. These reports contain vital information about the signal strength and quality from its current cell and neighboring cells.
Handover Decision Request: The Source base station receives these reports and determines that a handover might be necessary. Instead of relying on a simple "strongest signal" algorithm, it queries the AI Handover Decision Model. It forwards not only the signal data but also a richer set of inputs, including:
User mobility patterns: How fast and in what direction the user is moving.
Cell load: How congested the neighboring cells are.
Application type: Whether the user is streaming video, browsing the web, or on a voice call, as each has different network requirements.
AI Analysis and Prediction: The AI model processes this multi-faceted data. It evaluates potential target cells against several criteria to predict the best outcome, focusing on:
Seamless Connectivity: Preventing dropped calls or buffering video.
Load Balancing: Avoiding sending the user to an already congested cell.
Quality of Experience (QoE): Ensuring the user's application continues to perform well after the switch.
The AI model then sends its decision—the Optimal Target eNB/gNB—back to the Source station.
Handover Initiation: The Source station receives the AI's recommendation and initiates the formal handover process by sending a Handover Required message to the Core Network (MME), specifying the target cell chosen by the AI.
Target Cell Preparation: The MME communicates with the Target base station, sending a Handover Request to prepare it to receive the UE's connection. The Target station acknowledges this, confirming it has allocated the necessary resources.
UE Connection Command: Once the target is ready, the Core Network authorizes the handover. The Source station then sends a RRC Connection Reconfiguration message to the UE. This command contains the necessary information for the UE to detach from the Source and connect to the Target.
Handover Completion: The UE successfully switches to the Target base station and sends a RRC Connection Reconfiguration Complete message to confirm the handover. From the user's perspective, this transition is seamless.
Path and Resource Cleanup: The Target station notifies the MME that the handover is complete. The MME then updates the network path so that all data for the UE now flows through the new Target station. Finally, the MME instructs the Source station to release the resources it was holding for the UE, completing the cycle.
Service Data Adaptation Protocol (SDAP) Layer
The SDAP layer, present in the 5G user plane, is responsible for mapping Quality of Service (QoS) flows to Data Radio Bearers (DRBs).
QoS Flow Management: AI/ML can optimize the mapping of QoS flows to DRBs. By analyzing the characteristics of different data flows (e.g., latency requirements of an online game vs. throughput for a file download), an ML model can ensure that each flow is assigned to a DRB with the appropriate configuration. As seen in the AQM use case, the SDAP layer can also regulate the flow of packets to the RLC layer based on intelligence from a RIC, effectively managing QoS from a higher layer.
This flow diagram illustrates how a 5G base station (gNB) uses an AI/ML model, managed by a RAN Intelligent Controller (RIC), to intelligently handle different types of data traffic (QoS Flows).
Multiple Data Flows: The process begins when the User Equipment (UE) generates data from different applications, each with unique performance needs (e.g., a low-latency online game and a high-throughput file download).
Flow Analysis: This data arrives at the SDAP (Service Data Adaptation Protocol) layer in the gNB. The gNB reports the characteristics of these flows to the RIC.
AI-Powered Decision: The RIC forwards this information to the AI/ML Model. The model analyzes the requirements of each flow and determines the most efficient way to handle them. It decides:
Which Data Radio Bearer (DRB) each QoS flow should be mapped to. A DRB is a logical channel configured with specific quality-of-service parameters.
The optimal configuration for each DRB (e.g., settings for low latency vs. high throughput).
An Active Queue Management (AQM) policy to prevent network congestion.
Configuration and Mapping: The AI model sends its decision back to the RIC, which translates it into a concrete policy for the gNB. The SDAP layer receives this policy and performs two key actions:
Maps each QoS flow to the correct DRB as instructed (e.g., game traffic to a low-latency bearer, file download to a high-throughput bearer).
Regulates the flow of packets being sent down to the lower radio layers (PDCP/RLC/MAC). This is a form of intelligent traffic shaping that ensures packets are sent at a rate the radio link can handle, preventing buffer overload and maintaining QoS.
Data Transmission: The data, now properly sorted and managed, is processed by the lower layers and transmitted over the air to the UE, with each application's unique performance needs being met effectively.
Technical Challenges in Implementation
Integrating AI/ML into the 5G RAN is not without its difficulties:
Data Acquisition and Quality: AI models require large, high-quality datasets for training. Collecting and labeling this data from a live, dynamic network is a significant challenge.
Real-Time Operation: Many RAN functions, especially at the PHY and MAC layers, operate on a sub-millisecond timescale. AI models must perform inference with extremely low latency to be effective, which demands highly optimized software and hardware.
Complexity and Interpretability: Deep learning models can be "black boxes," making it difficult to understand their decision-making process. This lack of interpretability can be problematic in a carrier-grade network where reliability and predictability are paramount.
Standardization: Ensuring interoperability between AI models and network functions from different vendors requires standardization. Efforts by groups like the O-RAN Alliance are defining open interfaces (e.g., the E2 interface for the RIC) to facilitate this, but the process is ongoing.
Lifecycle Management: AI models are not static. They need to be continuously monitored, retrained, and updated to adapt to changing network conditions, which introduces operational complexity.
Increased Attack Surface: Integrating AI could introduce new security vulnerabilities. Adversarial attacks, where malicious actors feed manipulated data to fool an AI model, are a potential threat that must be addressed.
References:
3. Principles and Methodologies for AI/ML Testing in Next Generation Networks, https://mediastorage.o-ran.org/ngrg-rr/nGRG-RR-2024-05-Principals%20Methodologies%20AIML%20testing%20next%20Generation%20Networks-v1.9.pdf
4. https://www.techedgewireless.com/post/ai-for-ran-enabling-smart-and-adaptive-radio-access-networks
5. “Artificial Intelligence for 6G” by Haesik Kim