LCLei Cheng

Lei Cheng

*This is an algorithmically generated author record based on publicly available information.

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article


Lei Cheng has a total of 12 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Split Learning Based Cloud-Edge-End Collaborative Model Training in Heterogeneous Networks

    by Jian Wang, Gang Feng, Yi-Jing Liu, Xinyi Xu, Lei Cheng, Wei Jiang, Li Ping Qian

    Large language models (LLMs) demonstrate significant potential for enabling intelligent endogenous networks owing to their amazing intelligence level. LLMs are currently deployed on cloud servers as their vast parameter scales introduce a substantial computational burden to pre-training and inference processes. This deployment paradigm faces severe challenges, including insufficient personalization, high inference latency, and privacy concerns, as an increasing number of mobile users enjoy the LLM services. To enhance the personalization of pre-trained LLMs, exploiting massive local datasets distributed across edge devices to fine-tune LLMs is essential. In this paper, we propose a split learning-based cloud-edge-end collaborative training framework (SCCT) to harness the abundant computational resources of cloud and edge servers while exploiting massive distributed datasets on edge devices for LLM fine-tuning. In SCCT, an LLM with a parameter-efficient fine-tuning module is deployed on the cloud server, while a small-scale language model (SLM) is deployed across each edge device participating in SCCT and its associated edge server in the manner of split learning. SLM and LLM collaborate in a serial manner for training and inference. To optimize the collaboration efficiency, we formulate a mixed integer nonlinear programming problem to minimize the training latency of SCCT, considering the resource heterogeneity of edge devices and network dynamics. To solve this problem, a two-timescale optimization algorithm is proposed to determine the optimal split points on the large timescale and the allocation of communication and computational resources of the edge server on the small timescale. Numerical results demonstrate that the proposed two-timescale optimization algorithm assisted SCCT outperforms the state-of-the-art LLMs deployment and training paradigm in terms of training efficiency and inference performance.


  • Access Control in Satellite–Terrestrial Backhaul Network: An Emergent Communication Based Multi-Agent Reinforcement Learning Approach

    by Youkun Peng, Gang Feng, Shuang Qin, Lei Cheng, Jian Wang

    The satellite-terrestrial backhaul network (STBN) has recently emerged as a promising architecture to provide cost-effective seamless communication services to mobile users, especially in remote areas. In such networks, limited backhaul capacity, dynamic channel states, and user mobility pose significant challenges to user access control, involving efficient channel allocation and interference coordination to maximize user quality of service and network resource utilization. This dynamic user access problem can be formulated as a sequential decision process, and thus multi-agent reinforcement learning (MARL) can be a potential solution. However, conventional MARL lacks cooperation mechanisms for accurate state inference, which is critical for making optimal and cooperative resource allocation and interference management actions in such a dynamic and resource-constrained STBN. Moreover, although centralized coordination may improve system performance, it raises privacy and scalability concerns. In this paper, we propose a decentralized cooperative user access strategy, by leveraging emergent communication based MARL (EC-MARL). EC-MARL incorporates an emergent communication mechanism into the MARL framework, enabling user equipments (UEs) to cooperatively infer the network state more accurately while preserving privacy and scalability. Specifically, each agent encodes its local observation into a cooperation message, and shares the message among other agents. Based on these shared messages and their own local observations, UEs cooperatively make access decisions. Simulation results demonstrate that the proposed EC-MARL approach significantly outperforms several benchmarks in various network settings.

  • Scheduling Model Transmission and Training for Contact-Based Decentralized Satellite Federated Learning

    by Lei Cheng, Gang Feng, Shuang Qin, Yao Sun, Jian Wang, Feng Wang, Tony Q. S. Quek

    The significant enhancement of satellite onboard processing capability and drastic proliferation of Artificial Intelligence (AI) applications have fostered decentralized satellite federated learning (DSFL), a transformative paradigm that exchanges and aggregates machine learning (ML) models in satellite clusters for collaborative learning. However, the limited model exchange opportunities caused by intermittent inter-satellite contacts, along with heterogeneous onboard datasets, can lead to ineffective and/or biased model aggregation. To address these issues, it is crucial yet challenging to design an effective DSFL scheduling strategy that determines whether and when to pull models from contact satellites and perform local training for optimizing DSFL performance. In this paper, we propose a contact-based DSFL framework and formulate the DSFL scheduling problem to maximize the accuracy of trained models. As the problem cannot be solved directly, we transform it into a hierarchical Markov game by introducing options for decision agents deployed on individual satellites. Under a learning-to-learn paradigm, we develop a Multi-agent Dueling Double Deep Q Network (MA3DQN)-based intelligent DSFL scheduling strategy. The agents, trained in a distributed and alternating manner, adaptively make scheduling decisions based on instantaneous partial observations of the environment. Simulation results demonstrate the efficiency and adaptability of the MA3DQN-based strategy over three baselines.


  • Cooperative Transmission for Space-Air-Ground Integrated Networks: A Multi-Agent Cooperation Method

    by Lei Cheng, Xiaoqian Li, Gang Feng, Youkun Peng, Shuang Qin, Tony Q.S. Quek

    Unmanned aerial vehicles (UAVs) have been widely applied as aerial relays for satellite-to-terrestrial transmission so that data packets can be delivered either directly from satellites to user equipments (UEs) or relayed via UAVs. To exploit the advantages of such space-aerial cooperative transmission, data transmission, and resource allocation should be jointly optimized under constrained resources in dynamic space-air-ground integrated networks (SAGINs). However, the tight coupling between discrete transmission strategies and continuous resource allocation among users leads to a challenging long-term, large-scale, and non-convex optimization problem. While existing multi-agent reinforcement learning (MARL) techniques have exhibited potential in addressing such problems, their limitations in handling discrete-continuous decisions and ensuring robust multi-agent cooperation often result in significant performance deterioration. In this paper, we propose a new MARL-based Cooperative Transmission Strategy (MCTS) to overcome these limitations. In MCTS, each agent employs a hybrid Actor-Critic (H-AC) algorithm, where the optimal discrete transmission strategy with its associated resource allocation is modeled as parameterized actions and determined by an improved AC structure. To enhance cooperation among agents, a multi-agent cooperation framework is designed based on QMIX, employing a centralized Critic to efficiently guide convergence toward optimal solutions while reducing computational complexity. Simulation results demonstrate the superiority of MCTS over a number of benchmark algorithms and efficiency under various environments, in terms of overall transmitted data and resource utilization.


  • Energy-Constrained Satellite Edge Computing for Satellite-Terrestrial Integrated Networks

    by Lei Cheng, Gang Feng, Yao Sun, Shuang Qin, Feng Wang, Tony Q. S. Quek

    Satellite edge computing (SEC) has emerged as an innovative paradigm for future satellite-terrestrial integrated networks (STINs), expanding computation services by sinking computing capabilities into Low-Earth-Orbit (LEO) satellites. However, the mobility of LEO satellites poses two key challenges to SEC: 1) constrained onboard computing and transmission capabilities caused by limited and dynamic energy supply, and 2) stochastic task arrivals within the satellites' coverage and timevarying channel conditions. To tackle these issues, it is imperative to design an optimal SEC offloading strategy that effectively exploits the available energy of LEO satellites to fulfill competing task demands for SEC. In this paper, we propose a dynamic offloading strategy (DOS) with the aim to minimize the overall completion time of arriving tasks in an SEC-assisted STIN, subject to the long-term energy constraints of the LEO satellite. Leveraging Lyapunov optimization theory, we first convert the original long-term stochastic problem into multiple deterministic one-slot problems parameterized by current system states. Then we use sub-problem decomposition to jointly optimize the task offloading, computing, and communication resource allocation strategies. We theoretically prove that DOS achieves near-optimal performance. Numerical results demonstrate that DOS significantly outperforms the other four baseline approaches in terms of task completion time and dropping rate.