Consider an uplink rate splitting multiple access (RSMA) transmission in low earth orbit (LEO) terrestrial-satellite networks, by jointly optimizing the user access, transmit power splitting and receive beamforming. Despite advantages of deep reinforcement learning in online optimization, some issues still remain, i.e., the hybrid nature of discrete and continuous actions, low sample quality, and intractable representation of irregular environments. To circumvent these issues, we first formulate a parameterized Markov decision process (PAMDP) to represent the environment evolution, by parameterizing each discrete action as a continuous real vector, upon which the hybrid proximal policy optimization (PPO) is applied. Then, generative diffusion models (GDMs) are embedded into hybrid PPO to output the optimal action distribution, by adding the pure Gaussian noise with states into denoising networks. Experiments results corroborate that GDMs-assisted hybrid PPO achieves larger rewards with stability and faster convergence. Further, the proposed method dominates other benchmarks, under different terminal numbers, transmit power budgets and noise levels.

Traditional low earth orbit (LEO) satellite networks are typically independent of terrestrial networks, which develop relatively slowly due to the on-board capacity limitation. By integrating emerging mobile edge computing (MEC) with LEO satellite networks to form the business-oriented “end-edge-cloud” multi-level computing architecture, some computing-sensitive tasks can be offloaded by ground terminals to satellites, thereby satisfying more tasks in the network. How to make computation offloading and resource allocation decisions in LEO satellite edge networks, nevertheless, indeed poses challenges in tracking network dynamics and handling sophisticated actions. For the discrete-continuous hybrid action space and time-varying networks, this work aims to use the parameterized deep Q-network (P-DQN) for the joint computation offloading and resource allocation. First, the characteristics of time-varying channels are modeled, and then both communication and computation models under three different offloading decisions are constructed. Second, the constraints on task offloading decisions, on remaining available computing resources, and on the power control of LEO satellites as well as the cloud server are formulated, followed by the maximization problem of satisfied task number over the long run. Third, using the parameterized action Markov decision process (PAMDP) and P-DQN, the joint computing offloading, resource allocation, and power control are made in real time, to accommodate dynamics in LEO satellite edge networks and dispose of the discrete-continuous hybrid action space. Simulation results show that the proposed P-DQN method could approach the optimal control, and outperforms other reinforcement learning (RL) methods for merely either discrete or continuous action space, in terms of the long-term rate of satisfied tasks.

Transportation big data (TBD) are increasingly combined with artificial intelligence to mine novel patterns and information due to the powerful representational capabilities of deep neural networks (DNNs), especially for anti-COVID19 applications. The distributed cloud-edge-vehicle training architecture has been applied to accelerate DNNs training while ensuring low latency and high privacy for TBD processing. However, multiple intelligent devices (e.g., intelligent vehicles, edge computing chips at base stations) and different networks in intelligent transportation systems lead to computing power and communication heterogeneity among distributed nodes. Existing parallel training mechanisms perform poorly on heterogeneous cloud-edge-vehicle clusters. The synchronous parallel mechanism may force fast workers to wait for the slowest worker for synchronization, thus wasting their computing power. The asynchronous mechanism has communication bottlenecks and can exacerbate the straggler problem, causing increased training iterations and even incorrect convergence. In this paper, we introduce a distributed training framework, Heter-Train. First, a communication-efficient semi-asynchronous parallel mechanism (SAP-SGD) is proposed, which can take full advantage of acceleration effect of asynchronous strategy on heterogeneous training and constrain the straggler problem by using global interval synchronization. Second, Considering the difference in node bandwidth, we design a solution for heterogeneous communication. Moreover, a novel weighted aggregation strategy is proposed to aggregate the model parameters with different versions. Finally, experimental results show that our proposed strategy can achieve up to $6.74 \times$ speedups on training time, with almost no accuracy decrease.

Although convolutional neural networks (CNNs) have made significant progress, their deployment onboard is still challenging because of their complexity and high processing cost. Tensors provide a natural and compact representation of CNN weights via suitable low-rank approximations. A novel decomposed module called DecomResnet based on Tucker decomposition was proposed to deploy a CNN object detection model on a satellite. We proposed a remote sensing image object detection model compression framework based on low-rank decomposition which consisted of four steps, namely (1) model initialization, (2) initial training, (3) decomposition of the trained model and reconstruction of the decomposed model, and (4) fine-tuning. To validate the performance of the decomposed model in our real mission, we constructed a dataset containing only two classes of objects based on the DOTA and HRSC2016. The proposed method was comprehensively evaluated on the NWPU VHR-10 dataset and the CAST-RS2 dataset created in this work. The experimental results demonstrated that the proposed method, which was based on Resnet-50, could achieve up to 4.44 times the compression ratio and 5.71 times the speedup ratio with merely a 1.9% decrease in the mAP (mean average precision) of the CAST-RS2 dataset and a 5.3% decrease the mAP of the NWPU VHR-10 dataset.

Multi-beam low earth orbit (LEO) satellite communications, as an essential component for 6G systems, may encounter challenges from severe Doppler shifts and co-channel interference. This paper addresses a realistic problem in 6G-LEO systems, that is, how to meet the high-reliability demands of massive high-mobility terminals. We propose an integrated framework to exploit the synergy of non-orthogonal multiple access (NOMA) and orthogonal time frequency space (OTFS). OTFS modulation is employed to achieve full time-frequency diversity to combat Doppler shifts, while NOMA is used to accommodate more access requests. Specifically, within each beam, power domain superposition is applied to the delay-Doppler domain, enabling multiple terminals to share delay-Doppler grid resources. We analyze the performance of reliability, outage probability and ergodic capacity. Notably, we derive a novel closed-form expression to characterize the distribution of multi-beam interference with varying beam gains. Theoretical analysis and simulation results confirm that the proposed framework achieves a substantially lower outage probability compared to conventional OFDM schemes, with a system capacity improvement exceeding 11.9%.