Large language models (LLMs) have shown remarkable abilities by significantly scaling up model size, but this has also greatly increased their computing overhead. Traditional solutions to reduce the overhead are offloading inference tasks to cloud servers. With the evolution of computing power networks, computing resources are increasingly distributed at the network edge, allowing inference tasks to be handled locally. This edge inference can reduce traffic stress on backbone networks from cloud offloading and improve the utilization of heterogeneous edge computing power. However, the challenge is how to balance the gigantic computing workloads of LLMs with the limited computing power at the edge. To overcome this, a mixture-of-depths (MoD) Transformer based tri-parallel distributed approach was introduced. By dynamically allocating computing power to specific positions in the Transformer and parallel computing, this approach maximizes the capabilities of heterogeneous edge nodes to achieve resource-efficient inference. Simulation results showed that the proposal performs best in various edge environments, reducing inference delay by up to 24.3 % and energy consumption by up to 37.5%, respectively.

The evolution toward next-generation networks emphasizes the integration of distributed intelligence and the convergence of data, operation, information, and communication technologies (DOICT). While the 5th generation mobile network (5G) pioneered the cloudification of core networks to advance ICT integration, the 6th generation mobile network (6G) aims to further enhance this convergence by incorporating intelligent endogenous capabilities. This paper investigates the architecture of distributed intelligent endogenous, proposing a layered DOICT fusion framework designed to leverage distributed intelligence. Additionally, this paper defines the capability requirements between interfaces of the 6G intelligent endogenous network and explores prototype implementation, paving the way for the realization of next-generation networks.

To meet the multidimensional extreme performance requirements of intelligent services in sixth-generation mobile (6G) networks, it is crucial to implement the joint management of sensing, communication, and computation resources. However, the competition between services and the inherent conflicts among multidimensional resources result in a prominent contradiction between the efficiency of joint resource management and its high complexity. To address the challenges, a multi-service coexistence model is proposed, incorporating sensing, communication, and computing requirements. The optimization problem is decomposed to enable a low-complexity solution. Initially, a service resource management and mode selection algorithm is proposed, leveraging attention-assisted multi-agent reinforcement learning to effectively coordinate service resource competition. Subsequently, a one-to-one matching game is developed for radio resource blocks and users, ensuring stable maximization of joint sensing and communication performance while optimizing radio resource reuse. Finally, a computing resource management algorithm is designed using the Lagrange multiplier method and Karush-Kuhn-Tucker conditions to enhance computing performance. Theoretical analysis and numerical simulations validate the proposed schemes in terms of low complexity and high effectiveness, achieving approximately 20% overall performance improvement over baseline schemes.

The growing demand for edge computing is driving the proliferation of wireless computing power infrastructures and poses significant challenges to network energy efficiency (EE). Traditional offloading schemes rely solely on multi-access edge computing (MEC) servers. High-quality communication links and adequate distributed resources are expected to improve EE. Inspired by reconfigurable intelligent surface (RIS) and device-to-device communication technologies, this paper first proposed an edge-end collaborative computing system in RIS-empowered heterogeneous wireless computing power networks. Through resource virtualization, heterogeneous computing powers on MEC servers and nearby devices are unified into resource pools for efficient utilization. In this wireless system, task offloading is closely coupled with channel allocation, power coordination, RIS phase shift design, and base station receive beamforming. To tackle it, this paper suggested a block coordinate descent (BCD)-based framework that decouples the problem into three sub-problems. For each sub-problem, specialized solutions are applied: the Rayleigh quotient maximization and concave-convex procedure for the beamforming and power allocation co-design sub-problem, dimensionality reduction for 3-dimensional task offloading and channel allocation co-pairing sub-problem, and convex optimization for the RIS phase shift control sub-problem. Numerical results showed that the proposed method can outperform benchmark approaches in terms of EE and delay by up to 28.5% and 22.9%, respectively.

Generally, Reinforcement Learning (RL) agent updates its policy by repetitively interacting with the environment, contingent on the received rewards to observed states and undertaken actions. However, the environmental disturbance, commonly leading to noisy observations (e.g., rewards and states), could significantly shape the performance of agent. Furthermore, the learning performance of Multi-Agent Reinforcement Learning (MARL) is more susceptible to noise due to the interference among intelligent agents. Therefore, it becomes imperative to revolutionize the design of MARL, so as to capably ameliorate the annoying impact of noisy rewards. In this paper, we propose a novel decomposition-based multi-agent distributional RL method by approximating the globally shared noisy reward by a Gaussian Mixture Model (GMM) and decomposing it into the combination of individual distributional local rewards, with which each agent can be updated locally through distributional RL. Moreover, a Diffusion Model (DM) is leveraged for reward generation in order to mitigate the issue of costly interaction expenditure for learning distributions. Furthermore, the monotonicity of the reward distribution decomposition is theoretically validated under nonnegative weights and increasing distortion risk function, while the design of the loss function is carefully calibrated to avoid decomposition ambiguity. We also verify the effectiveness of the proposed method through extensive simulation experiments with noisy rewards. Besides, different risk-sensitive policies are evaluated in order to demonstrate the superiority of distributional RL in different MARL tasks.

The forthcoming sixth generation (6G) mobile communication system aims to advance technologies that span and integrate computation and communications. Computing power networks (CPNs) and virtualized radio access networks (vRANs) are regarded as two fundamental techniques to achieve this integration. Network functions of virtualized next-generation Node Bs (vgNBs) are implemented on general-purpose servers to process protocol stacks. The energy consumption of vgNBs accounts for a significant portion of energy consumption. However, the proliferation of computing power nodes results in increased energy consumption in CPNs. Power usage effectiveness (PUE) reflects the efficiency of computing nodes while efficiency of computing power (ECP) is adopted to indicate data rates per computing power unit. In this work, a joint load adjustment and sleep management scheme was designed to maximize ECP while minimizing PUE. The optimization problem was formulated as a mixed integer non-linear programming (MINLP) problem, which is NP-hard. A quantum genetic algorithm (QGA) with non-equal size quantum register was suggested to solve this problem. Simulation results demonstrated that the proposed algorithm could outperform benchmark approaches in terms of convergence speed, ECP, PUE, and computing power consumption. When compared to other methods, the proposed approach could improve ECP and computation energy consumption by up to 19.5% and 21.7%, respectively.

With the rapid advancement of global satellite communications, the integration of terrestrial and satellite communication has emerged as a focal point in the telecommunications field. The development of integrated space-ground networks represents a significant trend for the future of networking, presenting substantial challenges in mobility management for integrated ground-satellite networks. We propose several scenarios involving mobility management in path switching, approached from three distinct angles: Path Switching Modes, Terminal Communication Methods, and Connection Methods. These scenarios are designed to support seamless connectivity for users moving between different networks, especially in high-speed mobility contexts. The implementation of these schemes not only enhances connectivity in remote areas but also improves the overall reliability and resilience of the network through smooth transitions between terrestrial and satellite networks.

In recent years, the Intent-driven Network (IBN) has been proposed to further enhance the intelligence of communication systems. In IBN, users can express their resource expectations through intents while the IBN performs resource scheduling to fulfill these intents. Many challenges of complex networks can be tackled by IBN such as the mobile edge computing (MEC) system. In a MEC system, limited computing resources are competed by the users which may cause the intent conflict. To solve this, in this paper, we introduce the IBN concept to the MEC system, where an intent conflict detection module is proposed. The proposed module is based on the Open Network Automation Platform (ONAP) architecture. Moreover, by formulating the computing resource conflict problem as a Markov decision process (MDP) model, we employ an improved deep Q-network (DQN) algorithm to improve the efficiency of resource utilization. Simulation results demonstrate the completion time of intents is remarkably reduced in the proposed intent conflict resolution scheme.

Ultradense networks (UDNs) are considered to be a key technology that can meet the growing rate requirements caused by the explosion of user equipments (UEs) in the Internet of Things (IoT) applications. The dense deployment of small-cell base stations (SBSs) facilitates the reuse of spectrum resources but also leads to significant interference among adjacent SBSs. Joint transmission (JT) technology can alleviate intercell interference and improve throughput. However, signal processing and backhaul during BS cooperation require additional power consumption, which reduces the energy efficiency (EE) of UDNs. Additionally, the arrival time of received signals from different cooperative BSs at UEs results in asynchronous interference, which poses a significant challenge for JT. To improve EE, we need to determine the clustering strategy and address the interference issue of asynchronous JT. Specifically, the EE-centric (EEC) clustering scheme was proposed based on the maximal independent set of graph theory to determine the SBS clusters. In each cluster, asynchronous gap generation and gap compensation operations were employed to eliminate tail interference of asynchronous JT and compensate for the gap between adjacent received blocks, respectively. This approach effectively mitigated the asynchronous interference at UEs. Simulation results demonstrated that the proposed asynchronous interference cancelations in EEC clusters can significantly improve sum rate and EE compared to other schemes.

The Internet of Things (IoT) system provides sensing and computing services via terrestrial networks. However, the restricted coverage of terrestrial networks, such as base stations, limits the ubiquitous IoT services. Low-Earth orbit (LEO) satellites are able to provide network coverage for terrestrial IoT devices in unconnected scenarios, e.g., maritime. IoT devices in such scenarios usually have restricted onboard computation and power resources. In this article, we present an LEO-assisted IoT network (L-IoT) architecture where a device splits its task and offloads a portion of its task to the LEO to process within the coverage time. We formulate a task split problem with communication and computation resource allocation (SCC) to minimize the L-IoT energy consumption. We proposed an alternating optimization for split ratio and resource allocation (AOSR) algorithm. In particular, we use the outputs of Karush-Kuhn–Tucker (KKT) for resource allocation as part of the reward that feeds twin-delayed deep deterministic policy gradient. Lastly, the results of numerical simulations show that the proposed AOSR approach reduces 12.7% energy consumption compared to soft actor-critic (SAC) and 15% to deep deterministic policy gradient (DDPG).

Currently, Internet of Things (IoT) services in radio access networks require access to multi-dimensional network resources such as communication, computation, and caching to provide customized services. When resources are limited, there is always competition for resources and multi-dimensional resource requirement conflicts (MRRCs), which will lead to performance degradation of the IoT services. Moreover, the diverse resource requirements of IoT services and the fact that multi-dimensional resources are involved in scheduling make it extremely difficult to solve the MRRCs problem. To depict the above issues, we formulate a hierarchical MRRCs model, which applies the Stackelberg model and the multi-objective optimization model to describe the conflicts among services and users, respectively. Then, to address the aforementioned problem, we propose a deep reinforcement learning scheme with a hierarchically structured action space. Additionally, a case study is designed to simulate the resource conflicts of three different types of services on the spectrum, computation capacity, and caching resources. The numerical simulation results show that the proposed scheme has the best convergence ability and overall performance in terms of the MRRCs' coordination compared with the baseline schemes.

The popularity of 5G and cloud technology have created conditions for the development of novel video services, and at the same time, the services have put forward higher requirements for the network. By analyzing the problems of complex configuration, long setup cycle and poor scalability of traditional Metropolitan Area Network (MAN) architecture in the deployment of video services, and introduce the metropolitan area cloud network (MACN) architecture. Introduce how to deploy Ethernet Virtual Private Network (EVPN) over SRv6 based on MACN to satisfy the differentiated guarantee of different video services with simple configuration and flexible efficiency taking the cross-provincial unicast video service as an example. Taking Internet Protocol Television (IPTV) service as an example, introduce the solution of deploying Bit Index Explicit Replication IPv6 (BIERv6) multicast technology based on MACN to simplify multicast service configuration, reduce redundant video traffic in the network, shorten service turn-up cycle and improve user experience.

This paper presents a comprehensive framework for Intent-Based Networking (IBN). The framework is an open-source project, and its implementation is standards-based. Relevant IBN concepts from the standards organizations, the framework’s architecture, and its implementation on key IBN aspects and features including Intent life-cycle, Intent translation, Intent orchestration, and Intent assurance using closed-loops are discussed. The paper also demonstrates a real intent-based use case realized by the framework in order to show and validate the proof-of-concept. The Future work of this project is also discussed.

Ultrareliable and low-latency communications (uRLLC) is one of the most significant requirements for future wireless networks. The conventional terrestrial base stations cannot always provide the required uRLLC for emerging applications and scenarios, e.g., Tactile Internet services or when a large number of users get connected during an event. Therefore, multilayer airborne networks with low/medium/high altitude platforms can be deployed as an effective solution to offer capacity and coverage along with required latency and reliability for wireless networks. In this article, we propose three layers of the airborne network to support the uRLLC requirement in wireless networks. Optimized link selection schemes have been provided based on polychromatic sets (PSets) theory to focus on the uRLLC. With the optimized link selection algorithm, multiple properties of the airborne platforms are exploited, and the links are selected based on the multiconstrained requirements to support the desired performance of the airborne network. Moreover, two links distribution schemes have been proposed as distributed greedy scheme and centralized greedy scheme to demonstrate the deployment of the proposed airborne network. Numerical results show that both PSets-based links distribution schemes outperform the general distribution schemes on average latency and overall reliability also known as the unassociated ratio, which strongly supports the uRLLC in considered airborne networks.

Molecular Nanoscale Sensor Networks (MNSNs) introduce a new molecular routing paradigm where the molecular communication is enabled by activating the nanosensors on the routing path to release the molecules. This new molecular activation mechanism is a new many-to-many scheme where the transmitting nanosensors transmits molecules to multiple receiving nanosensors and the receiving nanosensors receives the molecules from multiple transmitting nanosensors. Molecular activation mechanism poses two new capacity constraints where the received molecules must be above a threshold to activate the receiving node and the molecules released from the transmitter should not exceed its molecular capacity. These two new criteria (many-to-many communication and capacity constraint) make the molecular activation and communication scheme a very challenging issue in an MNSN, which is totally different from the communication and routing scheme in existing wireless IP networks. In this paper, for the first time, we propose a sound mathematical model to capture the many-to-many communication scheme, activation capacity constraint and molecular capacity constraint in the MNSN. We then propose a novel QoS (cost and capacity) aware algorithm, CACAMA, to identify the cost efficient molecular activation and communication path in the MNSN. From the computational experiments, it shows that the CACAMA algorithm is superior to other two heuristics, SP and MCST, in all network settings.