eXtended Reality (XR) and holographic telepresence place stringent Quality of Service (QoS) demands on network infrastructure, requiring ultra-low latency, high throughput, and reliable connectivity. Meeting such QoS demands is critical in dynamic, distributed cloud environments, but does not always guarantee a satisfactory user experience. Quality of Experience (QoE) captures the user’s perception of service performance, which may be influenced by factors not fully reflected in systemlevel metrics. Thus, novel orchestration strategies must consider both QoS and QoE. This paper proposes a Reinforcement Learning (RL)-driven approach to edge-cloud orchestration capable of adapting to dynamic network conditions, leveraging a multiobjective reward function, including both QoS and QoE aspects, to guide service placement decisions. Evaluation shows that our RL approach reaches a 21.3% QoE gain over heuristics and 14.7% over balanced strategies, with 100% request acceptance. The results highlight the robustness and scalability of RL-driven orchestration, particularly for latency-sensitive 6G applications. Our findings also reveal the limitations of traditional heuristics under complex objectives and highlight the potential of RL as a transformative tool for intelligent network and service management in next-generation communication systems.

The emergence of the 6 G era demands seamless orchestration across an increasingly heterogeneous and distributed compute continuum-spanning edge, fog, and cloud resources. Moreover, the next generation of the mobile network is poised to redefine the digital landscape by enabling pervasive intelligence, ultra-low latency communication, and extreme heterogeneity across the entire network infrastructure. This transformation introduces unprecedented orchestration challenges due to the dynamic, multi-domain, and resource-constrained nature of emerging workloads such as Generative Artificial Intelligence (GenAI) inference, immersive eXtended Reality (XR), and autonomous systems. To tackle this complexity, we advocate for a Hybrid Hierarchical Digital Twin (DT) architecture that serves as a foundation for intelligent, adaptive, and real-time orchestration in 6 G environments. We present a comprehensive vision for integrating DTs as enablers of intelligent, context-aware, and adaptive orchestration mechanisms that span across multiple domains. The proposed architecture introduces a multi-layered DT hierarchy combining local and global views, enabling scalable coordination and real-time decision-making. We highlight key architectural enhancements required to realize this vision, including inter-twin interoperability and behavioral modeling for QoE estimation. This work aims to guide researchers and practitioners in shaping the foundations of resilient and efficient orchestration frameworks for 6 G systems.

Future 6G networks will be called upon to support increasingly sophisticated applications based on the massive use of Artificial Intelligence (AI). This poses serious challenges to mobile devices that do not have adequate computational and energy capacity to support the effective execution of AI tasks. According to the emerging Split-AI paradigm, 6G networks can help user devices to perform complex AI tasks by distributing and executing a Neural Network (NN) collaboratively among different processing nodes, including network nodes in the 6G User Plane (UP). However, the extent to which these offloading mechanisms can improve the end-to-end latency of AI tasks is limited by the computational resources available in an operator’s network. The presence of Graphical Processing Units (GPUs) in some elements of the UP can certainly help, but it is necessary to quantify the cost of the task acceleration in terms of increased energy consumption entailed by GPU usage. In this work, we study on the example of Split-AI, the impact of GPU acceleration on inference time and energy consumption. Thereby, we demonstrate that GPU inclusion can reduce the latency induced by the DNN computation by up to 90%, with moderate energy consumption increase of 22% at most.

The future of human interaction is virtual. Thus it will require effective collaboration on tasks among users in remote settings. eXtended Reality (XR) is playing a leading role in this transition, offering a realm where virtual collaboration becomes not just possible but essential in situations where physical presence is limited by risk, cost, or complexity. However, while networks are continuously evolving, they can still introduce unexpected impairments that potentially degrade the user perception, i.e., the Quality-of-Experience (QoE), of such Collaborative Virtual Reality (CVR) scenarios. In response to this challenge, this paper presents a demonstrator designed to explicitly showcase the effects of network conditions on CVR. Our platform, centered around a pizza-making game, allows for exploration of the real-time impact of different network parameters, such as packet delay, loss, and throttling on the user engagement and perception in CVR. The framework employs a combination of subjective, objective, and physiological assessments, including the capture of heart rate and skin conductivity, to gain comprehensive insights into user experiences. Our platform not only allows users to directly experience the impact of network impairments on CVR interactions but also provides initial evidence of how such distortions affect both subjective perceptions and objective performance metrics.

In-Network Computing (INC) is a currently emerging paradigm. Realizing INC in 6G networks could mean that user plane entities (UPEs) carry out computations on packets while transmitting them. These computations may have specific requirements in terms of their completion time. In case of high compute pressure at one UPE, migrating computations to another UPE may be beneficial, in order to avoid exceeding the completion time requirement. Centralized migration approaches suffer from increased signaling and are prone to react too slow. Therefore, this paper investigates the applicability of distributed intelligence to tackle the problem of compute task migration in the 6G User Plane. Each UPE is equipped with an intelligent agent, enabling autonomous decisions on whether computations should be migrated to another UPE. To enable the intelligent agents to learn and apply an optimal task migration policy, we investigate and compare two state-of-the-art Deep Reinforcement Learning (DRL) approaches: Advantage Actor-Critic (A2C) and Double Deep Q-Network (DDQN). We show, via simulations, that the performance of both solutions, in terms of the percentage of tasks exceeding their completion time requirement, is near-optimal and training A2C is at least 60% faster than DDQN.

The deployment of network elements in 6G is expected to be significantly more distributed than the existing 5G deployments. Distributed management paradigms are compatible with such distributed network deployments. Further, owing to their ability to solve complex problems by evaluating the impact of actions on the environment, intelligent solutions based on Reinforcement Learning (RL) for distributed management are promising. However, there are still several unsolved challenges before distributed intelligence could be seamlessly integrated in 6G. This work defines relevant research questions, reports on the progress made in the PhD project and presents the next steps and future directions for the advancement of this topic.

While the concept of AI4Net has been widely discussed in the past decade and adopted in 5G, its counterpart, Net4AI, has not gained that much attention so far. This is mostly due to the absence of solutions for the network to support AI applications beyond providing the communication infrastructure. In-Network Computing (INC) is a promising paradigm, potentially being integrated into 6G, which opens new solutions for realizing Net4AI. This paper focuses on the specific case of INC-assisted Split-AI. A Neural Network (NN) is split vertically and the executions of some layers of the split NN are offloaded to the entities of an Intelligent 6G User Plane. With the example of INC-assisted Split-AI, we elaborate on the challenges of Net4AI and discuss key requirements for the 6G architecture in terms of novel capabilities and information exchange.

In future 6G networks, the deployment of network elements is expected to be highly distributed, going beyond the level of distribution of existing 5G deployments. To fully exploit the benefits of such a distributed architecture, there needs to be a paradigm shift from centralized to distributed management. To enable distributed management, Reinforcement Learning (RL) is a promising choice, due to its ability to learn dynamic changes in environments and to deal with complex problems. However, the deployment of highly distributed RL – termed massive distribution of intelligence – still faces a few unsolved challenges. Existing RL solutions, based on Q-Learning (QL) and Deep Q-Network (DQN) do not scale with the number of agents. Therefore, current limitations, i.e., convergence, system performance and training stability, need to be addressed, to facilitate a practical deployment of massive distribution. To this end, we propose improved Double Deep Q-Network (IDDQN), addressing the long-term stability of the agents’ training behavior. We evaluate the effectiveness of IDDQN for a beyond 5G/6G use case: auto-scaling virtual resources in a network slice. Simulation results show that IDDQN improves the training stability over DQN and converges at least 2 times sooner than QL. In terms of the number of users served by a slice, IDDQN shows good performance and only deviates on average 8% from the optimal solution. Further, IDDQN is robust and resource-efficient after convergence. We argue that IDDQN is a better alternative than QL and DQN, and holds immense potential for efficiently managing 6G networks.

5G mechanisms typically rely on resource over-provisioning or static reservations to satisfy the stringent demands of critical connections. Consequently, 5G networks are inefficient in fulfilling application's low delay and high reliability requirements. Emerging 6G use cases will be even more demanding in terms of delay and reliability. Therefore, the usage of existing mechanisms would result in immense operational costs for the mobile network operators due to their in-efficiency. This highlights the need for investigating sophisticated mechanisms with the evolution towards 6G networks, providing reliability in a cost-efficient manner. This paper examines how to prioritize critical connections over non-critical ones, while constructively exploiting the available resources. To achieve this goal, we propose an intelligent admission control (AC) scheme for a radio access node (AN). More specifically, using AI/ML techniques, the AN forecasts the number of incoming critical connections and supports their reliability requirements by dynamically reserving resources for them, consequently affecting the admission of non-critical connections. Our simulation-based evaluations show that the proposed approach improves a baseline approach – not using intelligence –as it is capable of providing reliability to a larger number of connections, while efficiently utilizing system's resources. Thus, our work provides evidence of the potential of using intelligence for next-generation admission control design.

Driven by the development of programmable networking hardware, In-network Computing (INC) has gained a considerable amount of attention in recent years. However, INC has so far barely been studied in the context of mobile networks, despite the vast advantages shown for fixed networks, such as latency or traffic reduction. Motivated by an Augmented Reality (AR) use-case, our work envisions an INC-enabled Intelligent User Plane (IUP) for 6G networks, which allows offloading computational tasks to UP entities having enhanced computational capabilities. The 6G IUP thus helps to keep mobile end-devices lighter and supports meeting the stringent delay requirements of novel applications, such as AR. Besides elaborating on the involved prospects and challenges, we identify key enablers for realizing the INC-enabled IUP. We show that embedding INC into the 6G system entails major changes in the architecture, as compared to the current 5G design.

Recent studies on designing 6G networks, especially when it comes to satisfying the stringent latency and reliability requirements, identify native integration of Artificial Intelligence (AI) as a key enabler. Although 5G systems support features to provide high-reliability communication, they mostly rely on resource over-provisioning and are hence inefficient. We identify the need for addressing the problem of resource-efficiency from the perspective of 6G systems. The dynamic behaviors of radio access networks, due to varying radio channel conditions, pose an additional challenge in achieving high resource-efficiency. In this respect, we present and evaluate a Machine Learning (ML)-based mechanism for efficient support of safety-critical communication with stringent latency and reliability requirements. This mechanism is embedded in the admission control (AC) of an access node (AN) and uses Least Absolute Shrinkage and Selection Operator (LASSO)-based resource budget prediction while admitting new connection requests to the network. The goal is to maximize the number of clients admitted into the system while maintaining the reliability of the previously admitted critical connections. Our simulation-based evaluations highlight that the proposed approach outperforms the baseline approach - not featuring ML - by about 17% in terms of the average per-session reliability of safety-critical connections at different network load conditions, and hence, provide further evidence on the potentials of ML for next-generation networks design.

Monitoring and providing customers with a satisfying Quality of Experience (QoE) is a crucial business incentive for mobile network operators (MNOs). While the MNO is capable of monitoring a vast amount of network-related key performance indicators (KPIs), it typically does not have access to application-specific performance metrics. Among others, this is due to practical obstacles, such as missing standardized interfaces between the network and the application. Existing QoE models allow to map collected KPIs to the user-perceived quality. However, they are not dynamic, cumbersome to obtain, and often rely on application-level information, such as the stalling duration in the case of video streaming. The 5G networking architecture provides new features which can potentially overcome current limitations of in-network QoE monitoring. More specifically, the Application Function (AF) provides a standardized interface for communicating between 5G systems and third parties, such as application providers. The Network Data Analytics Function (NWDAF) is capable of collecting a vast number of network statistics from other 5G network functions and is dedicated to training and deploying Machine Learning (ML) models. This opens new possibilities, unimaginable for earlier mobile network generations, to dynamically learn the relationship between network KPIs and QoE by utilizing ML. Besides elaborating on how the new capabilities introduced with 5G can support an ML-based QoE estimation, we perform a simulation-based feasibility study which evaluates the estimation accuracy of different state-of-the-art regression techniques. In addition, we discuss them with respect to various qualitative aspects from an MNO’s point of view.

Network operators generally aim at providing a good level of satisfaction to their customers. Diverse application demands require the usage of beyond best-effort resource allocation mechanisms, particularly in resource-constrained environments. Such mechanisms introduce additional complexity in the control plane and need to be configured appropriately. Within 5G mobile networks, two new mechanisms for QoS-aware resource allocation are introduced. While QoS Flows enable specifying various QoS profiles on a per flow granularity, slices are dedicated virtual networks, strongly isolated against each other, with aggregated QoS guarantees. It is, however, unclear how QoS Flows and network slicing can optimally be exploited to ensure a high customer QoE while efficiently utilizing the available network resources. We address this research question and evaluate the outlined interplay using the OMNeT++ simulation environment in a multi-application scenario. We show that resource isolation induced by slicing may negatively affect application quality or system utilization, and that this impact can be overcome by finetuning the system parameters.

Since their first release, 5G systems have been enhanced with Network Data Analytics Functionalities (NWDAF) as well as with the ability to interact with 3rd parties' Application Functions (AFs). Such capabilities enable a variety of potentials, unimaginable for earlier generation networks, notable examples being 5G built-in Machine Learning (ML) mechanisms for QoE estimation, subject of this paper. In this work, an ML-based mechanism for video streaming QoE estimation in 5G networks is presented and evaluated. The mechanism relies on an ML algorithm embedded in NWDAF, the collection of 5G network KPIs, and the collection of QoE information from video streaming service provider, i.e., the 3rd party AF. The mechanism has been evaluated in terms of QoE estimation accuracy against the cost in terms of required input sources and data for the estimation, and its performance has been compared to alternative methodologies not making use of ML. The evaluation, via simulation activity, clearly highlights the benefits of the proposed mechanism. Based on the derived results, the required input sources are ranked with respect to their importance.

HTTP Adaptive Streaming (HAS) is the de-facto standard for video delivery over the Internet. It enables dynamic adaptation of video quality by splitting a video into small segments and providing multiple quality levels per segment. So far, HAS services typically utilize a fixed segment duration. This reduces the encoding and streaming variability and thus allows a faster encoding of the video content and a reduced prediction complexity for adaptive bit rate algorithms. Due to the content-agnostic placement of I-frames at the beginning of each segment, additional encoding overhead is introduced. In order to mitigate this overhead, variable segment durations, which take encoder placed I-frames into account, have been proposed recently. Hence, a lower number of I-frames is needed, thus achieving a lower video bitrate without quality degradation. While several proposals exploiting variable segment durations exist, no comparative study highlighting the impact of this technique on coding efficiency and adaptive streaming performance has been conducted yet. This paper conducts such a holistic comparison within the adaptive video streaming eco-system. Firstly, it provides a broad investigation of video encoding efficiency for variable segment durations. Secondly, a measurement study evaluates the impact of segment duration variability on the performance of HAS using three adaptation heuristics and the dash.js reference implementation. Our results show that variable segment durations increased the Quality of Experience for 54% of the evaluated streaming sessions, while reducing the overall bitrate by 7% on average.

HTTP Adaptive Streaming (HAS) is the de-facto standard for video delivery over the Internet. Splitting the video clip into small segments and providing multiple quality levels per segment allows the client to dynamically adapt the quality to current network conditions. The performance of HAS, and as a consequence the user Quality of Experience (QoE), is influenced by a multitude of parameters. This includes adjustable settings like quality switching thresholds, the initial buffer level, or the maximum buffer, as well as video characteristics like segment duration or the variation of segment sizes along the video. Recently, a couple of analytical models for video streaming have been proposed, allowing to compare these input parameters and derive their impact on QoE-relevant metrics for HAS-based video delivery. The outcome of these models are typically asymptotic probabilities, distribution functions, or centralized and standardized moments. For instance, these models do not yield any temporal information in terms of stalling events or requested video quality. This contradicts to QoE prediction models like P.1203, which compute the QoE based on the chronological sequence of a specific video playback. So far, it is unclear how and to which extent the generalized results of analytical models can be utilized to derive sequence-based QoE values or the QoE distribution for a set of sequences for similar input parameters with stochastic variations. To address this problem, we compare testbed measurements with the output of a GI/GI/1 model with pq-policy and buffer-based quality switching capability to conclude to which extent the results still allow to approximate the video QoE.

HTTP Adaptive Streaming (HAS) is the de-facto standard for video delivery over the Internet. It enables the dynamic adaptation of video quality by splitting the video clip into small segments and providing multiple quality levels per segment. Current HAS streaming services typically utilize segments of equal durations. However, this leads to video encoding overhead as segments have to start with I -frames, independently of the encoded video content. In this paper we evaluate the prospects of variable segment durations, where video segments are aligned to the video characteristics. We evaluate the reduction of the encoding overhead and investigate its impact on the stalling probability using a theoretical model. It turns out that the variable approach outperforms the fixed approach in 86% of the evaluated cases with respect to video stalls.