Logo: to the web site of Uppsala University

Federated learning (FL) is a promising paradigm that can enable collaborative model training between vehicles while protecting data privacy, thereby significantly improving the performance of intelligent transportation systems (ITSs). In vehicular networks, due to mobility, resource constraints, and the concurrent execution of multiple training tasks, how to allocate limited resources effectively to achieve optimal model training of multiple tasks is an extremely challenging issue. In this paper, we propose a mobility-aware multi-task decentralized federated learning (MMFL) framework for vehicular networks. By this framework, we address task scheduling, subcarrier allocation, and leader selection, as a joint optimization problem, termed TSLP. For the case with a single FL task, we derive the convergence bound of model training. For general cases, we first model TSLP as a resource allocation game, and prove the existence of a Nash equilibrium (NE). Then, based on this proof, we reformulate the game as a decentralized partially observable Markov decision process (DEC-POMDP), and develop an algorithm based on heterogeneous-agent proximal policy optimization (HAPPO) to solve DEC-POMDP. Finally, numerical results are used to demonstrate the effectiveness of the proposed algorithm.

On-orbit intelligence, such as federated learning (FL) on low Earth orbit satellites, poses challenges, as the environment is shaped by network dynamics and resource heterogeneity. This paper brings knowledge distillation (KD) into space-based FL. The use and training of both teacher and student models with KD is not meant just to handle — but to leverage — the system’s variability over time. Within this framework, we derive upper bounds on the two models’ global empirical loss; the bounds serve as a metric for effectively guide joint optimization of client selection and inter-satellite routing over dynamic network topology. We further provide dedicated convergence analysis tailored to the KD-enabled FL design. Beyond learning, we investigate novel multicast routing for global model distribution, as well as the optimization of client selection jointly with routing for collecting local model updates. Simulations on real-world satellite constellations showcase a significant and consistent performance edge of dual-model FL with KD in pushing the boundaries of on-orbit intelligence.

With the increasing uncertainty in production environments and changes in market demand, flexible and efficient scheduling solutions have become particularly critical. However, existing research mainly focuses on static scheduling or relatively simple dynamic scheduling problems, which are inadequate to address the complexities of actual production processes. This paper considers the dynamic flexible flow shop scheduling problem (DFFSP) characterized by diverse processes, complexity, and high flexibility, and proposes a multi-agent reinforcement learning algorithm based on cross-attention networks (MARL_CA). First, this paper proposes a novel state feature representation method, which represents the job processing data and the production Gantt chart as a state matrix, fully reflecting the environment state in the scheduling process. In addition, a cross-attention network is proposed to extract state features, enabling efficient discovery of complex relationships between jobs and machines, thereby enhancing the model's ability to understand intricate features. The model is trained using an independent proximal policy optimization (IPPO) based on the actor-critic method to help agents learn accurate and efficient scheduling strategies. Experimental results on a large number of static and dynamic scheduling instances demonstrate that the proposed algorithm outperforms traditional heuristic rules and other advanced algorithms, exhibiting strong learning efficiency and generalization capability.

Federated learning (FL) emerges as a promising approach to empower vehicular networks, composed by intelligent connected vehicles equipped with advanced sensing, computing, and communication capabilities. While previous studies have explored the application of FL in vehicular networks, they have largely overlooked the intricate challenges arising from the mobility of vehicles and resource constraints. In this paper, we propose a framework of mobility-aware decentralized federated learning (MDFL) for vehicular networks. In this framework, nearby vehicles train an FL model collaboratively, yet in a decentralized manner. We formulate a local iteration and leader selection joint optimization problem (LSOP) to improve the training efficiency of MDFL. For problem solving, we first reformulate LSOP as a decentralized partially observable Markov decision process (Dec-POMDP), and then develop an effective optimization algorithm based on multi-agent proximal policy optimization (MAPPO) to solve Dec-POMDP. Finally, we verify the performance of the proposed algorithm by comparing it with other algorithms.

Passively cooled base stations (PCBSs) have emerged to deliver better cost and energy efficiency. However, passive cooling necessitates intelligent thermal control via traffic management, i.e., the instantaneous data traffic or throughput of a PCBS directly impacts its thermal performance. This is particularly challenging for outdoor deployment of PCBSs because the heat dissipation efficiency is uncertain and fluctuates over time. What is more, the PCBSs are interference-coupled in multi-cell scenarios. Thus, a higher-throughput PCBS leads to higher interference to the other PCBSs, which, in turn, would require more resource consumption to meet their respective throughput targets. In this paper, we address online decision-making for maximizing the total downlink throughput for a multi-PCBS system subject to constraints related on operating temperature. We demonstrate that a reinforcement learning (RL) approach, specifically soft actor-critic (SAC), can successfully perform throughput maximization while keeping the PCBSs cool, by adapting the throughput to time-varying heat dissipation conditions. Furthermore, we design a denial and reward mechanism that effectively mitigates the risk of overheating during the exploration phase of RL. Simulation results show that our approach achieves up to 88.6% of the global optimum. This is very promising, as our approach operates without prior knowledge of future heat dissipation efficiency, which is required by the global optimum.

Federated Learning (FL) on low earth orbit (LEO) satellites represents a promising frontier for on-orbit edge intelligence. However, the inherent network dynamics and heterogeneity of datasets and resource across satellites pose challenges to efficient on-orbit FL. In this work, we model client selection and inter-satellite routing as a joint optimization problem. We derive and minimize an upper bound of the global empirical loss as the objective function, to enable fast convergence. We model the constraints of inter-satellite routing via time-varying graphs and network flow theory. We propose both exact and approximate solutions for the joint optimization problem. In addition, we formalize and prove the convergence property of our approach. Last, by simulation we demonstrate the efficiency and superiority of the proposed scheme for realistic satellite networking scenarios.

The rapid advancement of AI-generated content (AIGC) is poised to reshape content delivery, by enabling AIGC capabilities at the network edge or directly on end-user devices. This will allow content requests to be satisfied with AIGC based on prompts, rather than transmitting the original content. Each option presents unique trade-offs. On-device AIGC minimizes network traffic by transmitting only prompts, but it produces lower content quality than on-edge AIGC, which supports larger AI models. AIGC on the edge, in turn, results in lower quality than the original content. Delivering the original content requires more resource on radio access (and backhaul if not cached), while on-edge AIGC consumes computing power.  To optimally exploit these trade-offs, we formulate a utility maximization problem where for each content the system can opt for the original content, on-edge AIGC, or on-device AIGC, accounting for backhaul capacity, computational resources, and radio access constraints.  For this discrete optimization problem, we prove that, by applying Lagrangian multipliers to the three resource constraints, the problem relaxation can be efficiently solved to optimality. We then propose a solution approach that combines the problem relaxation with an algorithm for reaching feasible solutions of the overall problem. Simulation results demonstrate that our approach outperforms the baseline strategies and the solutions are close to the global optimum.

Caching popular contents at cell edge has been recognized as a promising way to facilitate rapid content delivery and alleviate backhaul burden. The content popularity is greatly influenced by recommendations by content providers. In this paper, we leverage this fact to jointly optimize caching and recommendation towards higher caching efficiency. We focus on both personalized and incumbent-aware recommendation. The incumbent content refers to the content that a user is currently browsing, resulted by the user's short-term interest. We model and formulate the resulting cache efficiency maximization problem subject to user satisfaction requirements. We prove the NP-hardness of the problem, and reformulate it using integer linear programming, enabling to solve optimally small-scale instances. Based on problem analysis with a graph representation, we derive three polynomial-time algorithms, where the recommendation sub-problem is solved to global optimum. Among these algorithms, the first two are based on sub-modularity, with  1−e⁻¹ approximation guarantee under mild conditions, while the last one is an alternation-based algorithm with fast convergence. Numerical results show the close-to-optimal performance of the proposed algorithms.

Passively cooled base stations (PCBSs) are highly relevant for achieving better efficiency in cost and energy. However, dealing with the thermal issue via load management, particularly for outdoor deployment of PCBS, becomes crucial. This is a challenge because the heat dissipation efficiency is subject to (uncertain) fluctuation over time. Moreover, load management is an online decision-making problem by its nature. In this paper, we demonstrate that a reinforcement learning (RL) approach, specifically Soft Actor-Critic (SAC), enables to make a PCBS stay cool. The proposed approach has the capability of adapting the PCBS load to the time-varying heat dissipation. In addition, we propose a denial and reward mechanism to mitigate the risk of overheating from the exploration such that the proposed RL approach can be implemented directly in a practical environment, i.e., online RL. Numerical results demonstrate that the learning approach can achieve as much as 88.6% of the global optimum. This is impressive, as our approach is used in an online fashion to perform decision-making without the knowledge of future heat dissipation efficiency, whereas the global optimum is computed assuming the presence of oracle that fully eliminates uncertainty. This paper pioneers the approach to the online PCBSs load management problem.

The optimization of sawing processes in the wood industry is critical for maximizing efficiency and profitability. The introduction of computerized tomography scanners provides sawmill operators with three-dimensional internal models of logs, which can be used to assess value and yield more accurately. We present a methodology for solving the sawing optimization problem employing a flexible sawing scheme that allows greater flexibility in cutting logs into products while considering product quality classes influenced by wane defects. The methodology has two phases: preprocessing and optimization. In the preprocessing phase, two alternative algorithms are given that generate and evaluate the potential sawing positions of products by considering the 3D surface of the log, product size requirements, and product quality classes. In the optimization phase, a maximum set-packing problem is solved for the preprocessed data using mixed-integer programming (MIP), aiming to obtain a feasible cut pattern that maximizes value yield. This is implemented in a system named FlexSaw, which takes advantage of parallel computation during the preprocessing phase and utilizes a MIP solver during the optimization phase. The proposed sawing methods are evaluated on the Swedish Pine Stem Bank. Additionally, FlexSaw is compared with an existing tool that utilizes cant sawing. Results demonstrate the superiority of flexible sawing. While the practical feasibility of implementing a flexible way of sawing logs is constrained by the limitations of current sawmill machinery, the potential increase in yield promotes the exploration of alternative machinery in the wood industry.