Zhongguo Li
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Previous Organizations / Institutes
The organizations / institutes from the scholar's other articles in the past ten years
Research Subjects
Zhongguo Li has a total of 61 co-authors. Here are the top co-authors by number of shared publications:
- Zhengtao Ding
Department of Electrical and Electronic Engineering, The University of Manchester, Manchester, U.K.
22 shared publications - Jun Yang
The Hong Kong University of Science and Technology (Guangzhou)
13 shared publications - Wen-Hua Chen
The Hong Kong Polytechnic University
9 shared publications - Zhen Dong
Harbin Institute of Technology
6 shared publications - Zhongchao Liang
Northeastern University
5 shared publications - Hanbo Ma
Department of Electrical and Electronic Engineering, The University of Manchester, Manchester, U.K.
4 shared publications
Latest Publications
- 3 September 2026
- Published by Elsevier in Neurocomputing
- Vol. 706
Load balancing for large language model inference serving: A control and learning oriented survey
- 3 September 2026
- Published by Elsevier in Neurocomputing
- Vol. 706
Load balancing for large language model inference serving: A control and learning oriented survey
- 1 July 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Internet of Things Journal
- Vol. 13 (13) , 29818-29829
A Hierarchical Control Framework for Autonomous Subsea Pipeline Leak Localization With a Crewed Submersible Vehicle
- 1 July 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Internet of Things Journal
- Vol. 13 (13) , 29818-29829
A Hierarchical Control Framework for Autonomous Subsea Pipeline Leak Localization With a Crewed Submersible Vehicle
- Open Access
- 18 June 2026
- Published by Elsevier in SSRN Electronic Journal
Load Balancing for Large Language Model Inference Serving: A Control and Learning Oriented Survey
Large language models (LLMs) are widely deployed as interactive neural inference services, where performance depends on how compute, memory, communication bandwidth, and persistent state are allocated under non-stationary demand. This survey studies load balancing for LLM inference serving as an intelligent scheduling and resource-allocation problem for large-scale learning systems. The analysis focuses on imbalance induced by prefill–decode phase asymmetry, key–value (KV) cache residency, heavy-tailed request variability, heterogeneous hardware, and service-level objectives (SLOs). A structured review protocol codes representative studies by imbalance source, observable signals, control knobs, assumptions, limitations, and evaluation evidence. We introduce a state– action–constraint abstraction, compare six strategy families, and synthesize evaluation practice around tail latency, time to first token (TTFT), token cadence, goodput, robustness, and energy per SLO-compliant token. Open directions are discussed for long-context and multimodal serving, decentralized deployment, green inference, and autonomous self-adaptive serving.
- 18 June 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessLoad Balancing for Large Language Model Inference Serving: A Control and Learning Oriented Survey
Large language models (LLMs) are widely deployed as interactive neural inference services, where performance depends on how compute, memory, communication bandwidth, and persistent state are allocated under non-stationary demand. This survey studies load balancing for LLM inference serving as an intelligent scheduling and resource-allocation problem for large-scale learning systems. The analysis focuses on imbalance induced by prefill–decode phase asymmetry, key–value (KV) cache residency, heavy-tailed request variability, heterogeneous hardware, and service-level objectives (SLOs). A structured review protocol codes representative studies by imbalance source, observable signals, control knobs, assumptions, limitations, and evaluation evidence. We introduce a state– action–constraint abstraction, compare six strategy families, and synthesize evaluation practice around tail latency, time to first token (TTFT), token cadence, goodput, robustness, and energy per SLO-compliant token. Open directions are discussed for long-context and multimodal serving, decentralized deployment, green inference, and autonomous self-adaptive serving.
- Open Access
- 3 April 2026
- Published by Elsevier in SSRN Electronic Journal
A Control-Oriented Survey of Load Balancing for LLM Inference Serving
Large Language Models (LLMs) have evolved into always-on services whose performance is governed by system-level decisions that integrate workload, state, and infrastructure signals rather than kernel-level peak performance alone. Prior surveys largely emphasize inference acceleration; in contrast, load balancing for LLM serving transcends request routing and becomes a coupled control problem shaped by phase asymmetry (prefill vs. decode), persistent and growing KV-cache state, and heavy-tailed workload variability. This survey provides a systematic review of load balancing for LLM inference serving from a full-stack control perspective. We first formalize LLM-specific load formation and interference mechanisms, highlighting how continuous batching, KV paging/residency, and parallelism choices reshape resource contention and tail latency. We then propose a mechanism-driven taxonomy that connects architectures (disaggregation and pooling), algorithms (phase-aware scheduling, variance-aware queueing, and KV-centric feasibility control), and deployment realities (heterogeneity/topology, multi-tenancy, and energy constraints) under a unified full-stack control perspective. Furthermore, we synthesize evaluation methodology emphasizing SLO-compliant goodput, streaming responsiveness (e.g., TTFT and token cadence), and robustness under bursty workloads, and we critique the limitations of simplistic throughput metrics. Finally, we discuss emerging frontiers—including long-context serving, decentralized and multi-tenant deployments, and energy-efficient control—offering a roadmap for next-generation inference systems.
- 3 April 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessA Control-Oriented Survey of Load Balancing for LLM Inference Serving
Large Language Models (LLMs) have evolved into always-on services whose performance is governed by system-level decisions that integrate workload, state, and infrastructure signals rather than kernel-level peak performance alone. Prior surveys largely emphasize inference acceleration; in contrast, load balancing for LLM serving transcends request routing and becomes a coupled control problem shaped by phase asymmetry (prefill vs. decode), persistent and growing KV-cache state, and heavy-tailed workload variability. This survey provides a systematic review of load balancing for LLM inference serving from a full-stack control perspective. We first formalize LLM-specific load formation and interference mechanisms, highlighting how continuous batching, KV paging/residency, and parallelism choices reshape resource contention and tail latency. We then propose a mechanism-driven taxonomy that connects architectures (disaggregation and pooling), algorithms (phase-aware scheduling, variance-aware queueing, and KV-centric feasibility control), and deployment realities (heterogeneity/topology, multi-tenancy, and energy constraints) under a unified full-stack control perspective. Furthermore, we synthesize evaluation methodology emphasizing SLO-compliant goodput, streaming responsiveness (e.g., TTFT and token cadence), and robustness under bursty workloads, and we critique the limitations of simplistic throughput metrics. Finally, we discuss emerging frontiers—including long-context serving, decentralized and multi-tenant deployments, and energy-efficient control—offering a roadmap for next-generation inference systems.
- 12 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Industrial Informatics
- Vol. 22 (4) , 3483-3494
A Transformer-Initialized Dual-Population Evolution for Large-Scale Task Scheduling in Heterogeneous Distributed Systems
Task scheduling in heterogeneous distributed systems is critical for industrial platforms, where decisions must be made under strict time constraints while resource states evolve dynamically. Existing approaches face significant limitations: classical heuristics yield suboptimal solutions; metaheuristics scale poorly; learning-based methods require extensive training with limited generalization. This article proposes a transformer-initialized dual-population evolution (TIDE), integrating three innovations: first, enhanced graph coloring preprocessing for enriched task representation, second, Transformer-based cross-modal attention for intelligent initialization of feasible solutions without offline pretraining, supported by an online adaptation mechanism, and third, asymmetric dual-population cooperative optimization with adaptive dimensionality reduction. Comprehensive experiments demonstrate that TIDE consistently outperforms state-of-the-art metaheuristics by 8%–13% in makespan while achieving an 80%–85% reduction in algorithm computing time compared to the metaheuristic average. On real scientific workflows, TIDE improves resource utilization by 4%–6% and maintains load balance above 94%, while maintaining response times within industrial deadlines. These results establish TIDE as a scalable solution for real-time scheduling in large-scale industrial systems.
- 12 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Industrial Informatics
- Vol. 22 (4) , 3483-3494
A Transformer-Initialized Dual-Population Evolution for Large-Scale Task Scheduling in Heterogeneous Distributed Systems
Task scheduling in heterogeneous distributed systems is critical for industrial platforms, where decisions must be made under strict time constraints while resource states evolve dynamically. Existing approaches face significant limitations: classical heuristics yield suboptimal solutions; metaheuristics scale poorly; learning-based methods require extensive training with limited generalization. This article proposes a transformer-initialized dual-population evolution (TIDE), integrating three innovations: first, enhanced graph coloring preprocessing for enriched task representation, second, Transformer-based cross-modal attention for intelligent initialization of feasible solutions without offline pretraining, supported by an online adaptation mechanism, and third, asymmetric dual-population cooperative optimization with adaptive dimensionality reduction. Comprehensive experiments demonstrate that TIDE consistently outperforms state-of-the-art metaheuristics by 8%–13% in makespan while achieving an 80%–85% reduction in algorithm computing time compared to the metaheuristic average. On real scientific workflows, TIDE improves resource utilization by 4%–6% and maintains load balance above 94%, while maintaining response times within industrial deadlines. These results establish TIDE as a scalable solution for real-time scheduling in large-scale industrial systems.
- Open Access
Auto-Optimization with Active Learning in Uncertain Environment: A Predictive Control Approach
This paper presents an auto-optimal model predictive control (MPC) framework enhanced with active learning, designed to autonomously track optimal operational conditions in an unknown environment,where the conditions may dynamically adjust to environmental changes. First, an exploitation-oriented MPC (EO-MPC) is proposed, integrating real-time sampling data with robust set-based parameter estimation techniques to address the critical challenge of parameter identification. By introducing virtual excitation signals into the terminal constraint and establishing a validation mechanism for persistent excitation condition, the EO-MPC effectively resolves the issue of insufficient persistent excitation in parameter identification. Building upon this foundation, an active learning MPC (AL-MPC) approach is developed to integrate both available and virtual future data to resolve the fundamental conflict between tracking an unknown optimal operational condition and parameter identification. The recursive feasibility and convergence of the proposed methods are rigorously established, and numerous examples substantiate the reliability and effectiveness of the approach in practical applications.
Open AccessAuto-Optimization with Active Learning in Uncertain Environment: A Predictive Control Approach
This paper presents an auto-optimal model predictive control (MPC) framework enhanced with active learning, designed to autonomously track optimal operational conditions in an unknown environment,where the conditions may dynamically adjust to environmental changes. First, an exploitation-oriented MPC (EO-MPC) is proposed, integrating real-time sampling data with robust set-based parameter estimation techniques to address the critical challenge of parameter identification. By introducing virtual excitation signals into the terminal constraint and establishing a validation mechanism for persistent excitation condition, the EO-MPC effectively resolves the issue of insufficient persistent excitation in parameter identification. Building upon this foundation, an active learning MPC (AL-MPC) approach is developed to integrate both available and virtual future data to resolve the fundamental conflict between tracking an unknown optimal operational condition and parameter identification. The recursive feasibility and convergence of the proposed methods are rigorously established, and numerous examples substantiate the reliability and effectiveness of the approach in practical applications.