AI/ML enhanced mobility in 5G-Advanced paving the way for an AI-native 6G
Future 6G use cases will bring a new set of challenging, more stringent requirements, as the mobile networks move beyond voice, video and mobile broadband communications.
Artificial intelligence and machine learning (AI/ML) are key enablers to address these new challenges. Therefore, 6G is expected to be the first mobile network generation, where AI/ML will be a native part of the system design from the very beginning.
The AI/ML framework for the air-interface studied in 3GPP as part of 5G-Advanced Release-18 is the foundation for the common enablers across various ML-enabled use-cases. A continuation of this study focusing on ML-optimized inter-cell mobility has been proposed by Nokia, and will be studied in the upcoming Release-19. This initiative aims to study how AI/ML can be applied to enhance network mobility performance and to further refine, extend and validate the AI/ML framework with the AI/ML-optimized mobility use cases, paving the way for a future-proof 6G AI-native air interface.
Addressing future needs with AI/ML-optimized mobility
Figure 1: AI/ML-enhanced mobility challenges, capabilities and advantages
Mobility procedures have been extensively studied and optimized in every 3GPP release. For instance, to improve handover robustness, conditional handover (CHO) has been introduced, interruption time has been reduced with the Dual Active Protocol Stack (DAPS) handover and Lower-layer Triggered Mobility (LTM) methods. Thus, handovers have reached a high level of maturity and can generally be considered reliable.
However, the increasing concentration of users, utilization of higher-frequency bands and denser network deployments leads to frequent handovers becoming a more prominent challenge, while the need for high energy efficiency and reduced network complexity require efficient and robust mobility management. New network applications are posing new requirements for the mobility methods. For instance, Time Sensitive Communication (TSC) and XR use-cases are typically subject to strict packet delay budgets (PDBs).
Machine learning solutions will enable further mobility optimization in unique and challenging scenarios that cannot be addressed with the non-ML methods.
Figure 2: Mobility context features that ML can (implicit) learn and adapt to
AI/ML can learn how different features shown in Figure 2, such as different mobility profiles (pedestrian, ambulance) and speeds (slow, fast), trajectories (road, sky) and radio environment features (antennas, buildings, trees), affect the mobility in each scenario. The learned insights can be used to improve handover reliability, reduce interruption time, optimize resource usage, and reduce radio measurements that the user equipment needs to take and report.
Figure 3: ML-based inter-frequency measurement reduction
Let us look at one example, in which AI/ML is used to reduce the number of costly inter-frequency measurements that a UE needs to take. Inter-frequency measurements are costly as they require energy at the UE and while the UE is measuring, it cannot communicate. Figure 3 shows a scenario, where a macro cell layer A provides coverage and micro cell layer B provides additional capacity, but not full coverage. Connecting to layer B is preferred whenever possible, for improved network performance and for better load balancing. A drone, for example, may connect to layer B when approaching its destination for higher throughput and lower latency to stream live video for better situational awareness to the first responders. In the baseline scenario, when connected to the macro layer A, in addition to making measurements in the layer A, the drone must also measure the micro layer B to determine if it would be in the coverage of a micro cell. However, with measurement data from both layers, we can train an ML model that learns the correlation between the radio measurements in the two layers and can predict if a drone connected to the macro layer A would be able to connect to a micro layer B cell based on layer A measurements. Only when the AI/ML model predicts that a connection is possible, inter-frequency measurements of layer B are taken to determine, which small cell in layer B can be connected to, and to make sure that the ML prediction was correct.
Figure 4: Simulation results for inter-frequency measurements
Unavoidably ML will sometimes make wrong predictions. In this case, a false positive prediction will lead to an unnecessary measurement of layer B, reducing the intended measurement savings. A false negative prediction, on the other hand, can delay connecting to the preferred layer B. The trade-off between these two impacts can be configured and the better the ML model performance, the less trade-off we need to make. Our simulation experiment results are shown in Figure 4. We can save from 20% up to over 50% of the inter-frequency measurements over a non-ML baseline, depending on if we want the maximum utilization of layer B, as established without any measurement reduction, or if we accept slight reduction in the time connected to it.
Other examples of how AI/ML can be utilized to optimize mobility include, for example, optimizing resource allocations in handovers based on predicted serving beams, short-stay handovers or other mobility related events.
Enabling machine learning optimized mobility
Let’s look at the three key considerations enabling ML-optimized mobility: network- and UE-sided deployment, reliability and testability.
Network- and UE-sided models: In mobility use cases, one key capability of ML is the ability to learn the local mobility context and to adapt to it. This means that ML model instances may be trained for a specific set of cells and their adjacencies. These so-called cell-, site-, or area-specific ML models are suitable for network-side mobility optimization solutions, where sufficient data for that context and the right model instance are always available. UE-sided models are better suited for higher ML model generalization, but on the other hand may have better availability of radio measurements and other information not accessible by the network.
Reliability: Mobility methods need to be very reliable, even in case of unexpected circumstances. This includes any ML solutions used to optimize mobility, which must perform in a wide variety of scenarios and be robust, resilient, and failsafe. For example, radio measurements are by nature noisy and including unpredictable elements. Handover decisions are influenced by factors, where predictability is limited in varying degrees, like transient blocking objects, the pose and movement of the UE and more. In many use cases, it may be only possible to give accurate predictions statistically rather than for individual handovers. Furthermore, ML is introducing memory into the mobility methods. What this means is that the models predict based on the network configuration and environment captured in the training data. If any of the features shown in Figure 2 change, for example radio environment (construction of a new building) or the mobility profiles (advent of electric bikes), this may reduce the accuracy of the ML models and require re-training. For this reason, coordinated ML Life-Cycle Management (LCM) and monitoring is required, and ML-enabled mobility solutions must be robust and failsafe, even in unforeseen situations. Figure 5 shows the effect of a radio environment change to the inter-frequency measurement reduction model in our simulation study. With ML monitoring the ML performance change can be detected and ML LCM can do the required re-training and/or ML model switch. Furthermore, to be prepared for any unforeseen events, the algorithm is designed to be inherently failsafe, i.e., any false predictions from the machine learning model do not lead to failures.
Figure 5: Impact of radio environment change to the inter-frequency measurement reduction ML model
Testability: Introducing machine learning to mobility methods requires new methods to evaluate and verify mobility procedures. Especially when the model is trained with localized context and adapted over time. This may include, for example, studying any changes required in the reference scenarios used to evaluate the solutions as well as new testing procedures and methods in the UE conformance testing.
Work towards an AI/ML-native network continues
In Release-19, we are taking the next steps towards 6G, where the AI/ML capabilities are a native part of the network design. The advantages of AI/ML and the required enablers are further studied with the mobility use cases. At the same time, this study will contribute to creating a future-proof AI/ML framework for the air interface, which will meet the requirements of the future 6G applications.