Hao Zhang
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Hao Zhang has a total of 54 co-authors. Here are the top co-authors by number of shared publications:
- Jie Liu
Harbin Institute of Technology, Shenzhen, China
18 shared publications Jie Liu
11 shared publications- Siyao Cheng
Faculty of Computing, Harbin Institute of Technology, Harbin, China
11 shared publications - Tingting Wu
Harbin Engineering University
10 shared publications - Zhifeng Ma
Faculty of Computing, Harbin Institute of Technology, Harbin, China
5 shared publications - Chenxin Cai
Harbin Institute of Technology
4 shared publications
Latest Publications
- Open Access
- 6 November 2026
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Sensor Networks
Path Refinement for Large-Scale Heterogeneous Multi-UAV Coverage
The emergence of Unmanned Aerial Vehicles (UAVs) has transformed static Wireless Sensor Networks (WSNs) into dynamic mobile sensing networks, enabling large-scale data acquisition for applications such as surveillance and precision agriculture. In these missions, multiple heterogeneous UAVs equipped with onboard sensors (e.g., cameras, multispectral sensors or LiDAR units) must collaboratively perform area coverage. The efficiency of the planned paths, governed by the Heterogeneous Multi-robot Coverage Path Planning (HMCPP) problem, directly impacts mission success. However, existing approaches to HMCPP are often limited by either prohibitive computational complexity or suboptimal path efficiency. To address this challenge, the HMCPP problem is modeled from a graph-based perspective, and a novel framework, termed Heterogeneous Cost Balancing (HCB), is introduced to efficiently generate high-quality solutions. HCB transforms existing solutions into HMCPP solutions by integrating fine-grained planning with cost-balancing strategies. Specifically, the proposed fine-grained planning approach refines traditional spiral routes, reducing path redundancy and enhancing coverage efficiency. Experimental results demonstrate that HCB outperforms baseline methods in most instances while maintaining high computation efficiency. Notably, on a large-scale real-world map (366 × 376 with 40 UAVs), HCB achieves improvements in makespan of 33.82%, 9.79%, 6.37%, and 12.34% over four heuristic-based baselines, and an 86.66% improvement over MIP-based approach.
- 6 November 2026
- Published by Association for Computing Machinery (ACM) in ACM Transactions on Sensor Networks
Open AccessPath Refinement for Large-Scale Heterogeneous Multi-UAV Coverage
The emergence of Unmanned Aerial Vehicles (UAVs) has transformed static Wireless Sensor Networks (WSNs) into dynamic mobile sensing networks, enabling large-scale data acquisition for applications such as surveillance and precision agriculture. In these missions, multiple heterogeneous UAVs equipped with onboard sensors (e.g., cameras, multispectral sensors or LiDAR units) must collaboratively perform area coverage. The efficiency of the planned paths, governed by the Heterogeneous Multi-robot Coverage Path Planning (HMCPP) problem, directly impacts mission success. However, existing approaches to HMCPP are often limited by either prohibitive computational complexity or suboptimal path efficiency. To address this challenge, the HMCPP problem is modeled from a graph-based perspective, and a novel framework, termed Heterogeneous Cost Balancing (HCB), is introduced to efficiently generate high-quality solutions. HCB transforms existing solutions into HMCPP solutions by integrating fine-grained planning with cost-balancing strategies. Specifically, the proposed fine-grained planning approach refines traditional spiral routes, reducing path redundancy and enhancing coverage efficiency. Experimental results demonstrate that HCB outperforms baseline methods in most instances while maintaining high computation efficiency. Notably, on a large-scale real-world map (366 × 376 with 40 UAVs), HCB achieves improvements in makespan of 33.82%, 9.79%, 6.37%, and 12.34% over four heuristic-based baselines, and an 86.66% improvement over MIP-based approach.
- 1 November 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Mobile Computing
- Vol. 25 (11) , 20982-20996
COSMIC: Channel Optimized Semantic Superpixel Intelligence Communication With NOMA for STINs
- 1 November 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Mobile Computing
- Vol. 25 (11) , 20982-20996
COSMIC: Channel Optimized Semantic Superpixel Intelligence Communication With NOMA for STINs
- Open Access
- 27 September 2026
- Published by Association for Computing Machinery (ACM)
- p. 866-876
Scalable RL-Based Online Scheduling Framework for Modular Process Systems
To meet the online scheduling requirements of large-scale parallel systems, traditional centralized path planning algorithms suffer from severe computational bottlenecks as system scale increases. This paper proposes FreeFlow, a fully decentralized, expert-free framework for online distributed scheduling. The core highlight of FreeFlow is to decouple decision complexity from system scale, enabling highly parallel execution on ultra-large platforms without the need for global state synchronization. Driven strictly by local observations, FreeFlow employs a guided reward mechanism that promotes proactive collision avoidance and cooperative spatial distribution without relying on expert knowledge or heuristic guidance. To support continuous task streams, we further introduce a dynamic agent pool mechanism based on object reuse, addressing the mismatch between dynamic task arrivals and fixed input dimensions in parallel neural networks. Extensive experiments show that FreeFlow achieves throughput comparable to state-of-the-art centralized methods in large-scale settings while significantly reducing runtime. It also demonstrates strong generalization to larger environments, maintaining over 92% success rate with millisecond-level inference latency in a 1024×1024 grid environment.
- 27 September 2026
- Published by Association for Computing Machinery (ACM)
- p. 866-876
Open AccessScalable RL-Based Online Scheduling Framework for Modular Process Systems
To meet the online scheduling requirements of large-scale parallel systems, traditional centralized path planning algorithms suffer from severe computational bottlenecks as system scale increases. This paper proposes FreeFlow, a fully decentralized, expert-free framework for online distributed scheduling. The core highlight of FreeFlow is to decouple decision complexity from system scale, enabling highly parallel execution on ultra-large platforms without the need for global state synchronization. Driven strictly by local observations, FreeFlow employs a guided reward mechanism that promotes proactive collision avoidance and cooperative spatial distribution without relying on expert knowledge or heuristic guidance. To support continuous task streams, we further introduce a dynamic agent pool mechanism based on object reuse, addressing the mismatch between dynamic task arrivals and fixed input dimensions in parallel neural networks. Extensive experiments show that FreeFlow achieves throughput comparable to state-of-the-art centralized methods in large-scale settings while significantly reducing runtime. It also demonstrates strong generalization to larger environments, maintaining over 92% success rate with millisecond-level inference latency in a 1024×1024 grid environment.
- Open Access
- 14 September 2026
- Published by Elsevier in SSRN Electronic Journal
Learning to Retrieve and Fuse: An Efficient Historical Knowledge Integration Approach for Federated Continual Learning
Federated Learning (FL) typically assumes that the underlying data distribution remains stationary throughout training, which is unrealistic in many real world applications. In practice, data distributions may evolve over time and vary substantially across clients, making it challenging to preserve previously acquired knowledge while adapting to newly emerging tasks. To address this issue, federated continual learning (FCL) has been proposed. However, existing FCL methods often perform coarse integration of historical knowledge, without explicitly considering the heterogeneous relationships among historical tasks and their relevance to the current task. In this work, we propose a Retrieval and Fusion based Federated Continual Learning (RFFCL) algorithm to selectively retrieve and effectively integrate relevant historical knowledge into current task learning, thereby mitigating catastrophic forgetting. Specifically, to reduce the storage overhead of local devices, we adopt an encoder-decoder architecture to decouple shared knowledge from task-specific knowledge. Contrastive learning is incorporated into encoder training to learn more generalizable and transferable knowledge representations from the shared knowledge, which enables the task-specific decoder to learn better decision boundaries. A locality-sensitive hashing mechanism is then developed to efficiently organize historical knowledge and enable fast retrieval of relevant knowledge. We further introduce a weighted regularization term into the local objective to guide the selective integration of historical and current task knowledge while preserving previously learned representations. Extensive experiments on multiple benchmark datasets demonstrate that RFFCL consistently improves current-task learning performance while effectively mitigating catastrophic forgetting compared with state-of-the-art FCL methods.
- 14 September 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessLearning to Retrieve and Fuse: An Efficient Historical Knowledge Integration Approach for Federated Continual Learning
Federated Learning (FL) typically assumes that the underlying data distribution remains stationary throughout training, which is unrealistic in many real world applications. In practice, data distributions may evolve over time and vary substantially across clients, making it challenging to preserve previously acquired knowledge while adapting to newly emerging tasks. To address this issue, federated continual learning (FCL) has been proposed. However, existing FCL methods often perform coarse integration of historical knowledge, without explicitly considering the heterogeneous relationships among historical tasks and their relevance to the current task. In this work, we propose a Retrieval and Fusion based Federated Continual Learning (RFFCL) algorithm to selectively retrieve and effectively integrate relevant historical knowledge into current task learning, thereby mitigating catastrophic forgetting. Specifically, to reduce the storage overhead of local devices, we adopt an encoder-decoder architecture to decouple shared knowledge from task-specific knowledge. Contrastive learning is incorporated into encoder training to learn more generalizable and transferable knowledge representations from the shared knowledge, which enables the task-specific decoder to learn better decision boundaries. A locality-sensitive hashing mechanism is then developed to efficiently organize historical knowledge and enable fast retrieval of relevant knowledge. We further introduce a weighted regularization term into the local objective to guide the selective integration of historical and current task knowledge while preserving previously learned representations. Extensive experiments on multiple benchmark datasets demonstrate that RFFCL consistently improves current-task learning performance while effectively mitigating catastrophic forgetting compared with state-of-the-art FCL methods.
- 28 May 2026
- Published by Elsevier in Knowledge-Based Systems
- Vol. 348
A Lyapunov-based client selection approach to handle system-induced heterogeneity in federated learning
- 28 May 2026
- Published by Elsevier in Knowledge-Based Systems
- Vol. 348
A Lyapunov-based client selection approach to handle system-induced heterogeneity in federated learning
- 16 April 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Geoscience and Remote Sensing
- Vol. 64 (01962892) , 1-18
S 3 NN: Exploring Stochastic Modeling in Precipitation Nowcasting Domain
Precipitation nowcasting based on artificial intelligence has gained significant attention in both meteorology and computer science communities in the last few years. While researchers continuously introduce new models to improve forecasting performance, it is worth noting that most of these models are deterministic. Unfortunately, it is difficult for deterministic models to model future uncertainty, visually shown as blurry. Stochastic generative models offer a feasible solution, they are dedicated to learning the data distribution and generating new samples from it. Differently, ensemble models deal with uncertainty through probability processing. Inspired by ensemble models, we view precipitation nowcasting as a stochastic process. We propose a new model named SNN, which samples high-quality frames from a collection of proposal frames to compose the final prediction. The ongoing debate about which model is better has not yet concluded, and the better approach is to integrate and absorb their respective strengths. By combining SNN with two existing stochastic generative models, the generative adversarial network and the variational auto-encoder, we propose S3NN to model uncertainty and mitigate ambiguity in precipitation nowcasting. S3NN can operate in member-mode and ensemble-mode simultaneously, requiring only one end-to-end training. This enables it not only to produce accurate precipitation forecasts but also to perform probabilistic forecasts, the latter being extremely useful in estimating probabilities of excessive precipitation. Finally, we perform extensive experiments on two radar datasets named HKO-7 and MeteoNet, and the results validate the superiority of S3NN.
- 16 April 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Geoscience and Remote Sensing
- Vol. 64 (01962892) , 1-18
S 3 NN: Exploring Stochastic Modeling in Precipitation Nowcasting Domain
Precipitation nowcasting based on artificial intelligence has gained significant attention in both meteorology and computer science communities in the last few years. While researchers continuously introduce new models to improve forecasting performance, it is worth noting that most of these models are deterministic. Unfortunately, it is difficult for deterministic models to model future uncertainty, visually shown as blurry. Stochastic generative models offer a feasible solution, they are dedicated to learning the data distribution and generating new samples from it. Differently, ensemble models deal with uncertainty through probability processing. Inspired by ensemble models, we view precipitation nowcasting as a stochastic process. We propose a new model named SNN, which samples high-quality frames from a collection of proposal frames to compose the final prediction. The ongoing debate about which model is better has not yet concluded, and the better approach is to integrate and absorb their respective strengths. By combining SNN with two existing stochastic generative models, the generative adversarial network and the variational auto-encoder, we propose S3NN to model uncertainty and mitigate ambiguity in precipitation nowcasting. S3NN can operate in member-mode and ensemble-mode simultaneously, requiring only one end-to-end training. This enables it not only to produce accurate precipitation forecasts but also to perform probabilistic forecasts, the latter being extremely useful in estimating probabilities of excessive precipitation. Finally, we perform extensive experiments on two radar datasets named HKO-7 and MeteoNet, and the results validate the superiority of S3NN.
- 20 November 2025
- Published by Elsevier in Neural Networks
- Vol. 196, 108217
Chain-of-Detection enables robust and efficient jailbreak defense
- 20 November 2025
- Published by Elsevier in Neural Networks
- Vol. 196, 108217
Chain-of-Detection enables robust and efficient jailbreak defense
- Open Access
- 27 October 2025
- Published by Association for Computing Machinery (ACM)
- p. 4058-4067
BOLT: Fewer Tokens but More Performance Retention for Efficient Vision-Language Models Inference
- 27 October 2025
- Published by Association for Computing Machinery (ACM)
- p. 4058-4067
Open AccessBOLT: Fewer Tokens but More Performance Retention for Efficient Vision-Language Models Inference
- 22 October 2025
- Published by Elsevier in Expert Systems with Applications
- Vol. 299
An online multi-agent path finding algorithm for large-scale puzzle-based conveyor system
- 22 October 2025
- Published by Elsevier in Expert Systems with Applications
- Vol. 299
An online multi-agent path finding algorithm for large-scale puzzle-based conveyor system