Hanchen Luo
Research Subjects
Hanchen Luo has a total of 8 co-authors. Here are the top co-authors by number of shared publications:
- Chunlei Wang
China Information Security Research Institute Company Ltd., Beijing, China
3 shared publications - Shaopu Wang
China Information Security Research Institute Company Ltd., Beijing, China
2 shared publications - Yike Gao
North China Institute of Computer Systems Engineering, Beijing, China
2 shared publications - Bang Wang
Huazhong University of Science and Technology, Wuhan, China
1 shared publication - Enjun Qiao
Beijing Institute of Technology, Beijing, China and North China Institute of Computer Systems Engineering, Beijing, China
1 shared publication - Mengyuan Li
China Information Security Research Institute Company Ltd., Beijing, China
1 shared publication
Latest Publications
- Open Access
- 19 July 2026
- Published by Association for Computing Machinery (ACM)
- p. 4261-4266
EvoMem: Timeline-Grounded Validity Adjudication for Non-Monotonic Multi-Agent Memory
- 19 July 2026
- Published by Association for Computing Machinery (ACM)
- p. 4261-4266
Open AccessEvoMem: Timeline-Grounded Validity Adjudication for Non-Monotonic Multi-Agent Memory
- Open Access
- 23 January 2026
- Published by Association for Computing Machinery (ACM)
- p. 436-443
An NTK-Aware Multi-Round Iterative Moving Collocation Point Framework for Physics-Informed Neural Networks
Physics-Informed Neural Networks (PINNs) have emerged as a mesh-free alternative for solving Partial Differential Equations (PDEs), bypassing the computationally expensive mesh generation required by traditional numerical methods. However, this flexibility comes at a cost: by replacing a fixed mesh with a set of collocation points, PINNs shift the core challenge of spatial discretization to the strategic distribution of these points. Consequently, the arrangement of these collocation points directly impacts the efficacy and convergence speed of PINNs. Conventional sampling strategies, such as uniform distribution or residual-based sampling methods, essentially treat the PDE residual field as a static target and typically rely solely on residual magnitudes, overlooking the intrinsic training dynamics of the neural network. Unlike methods that rely solely on residual magnitudes, our approach leverages NTK-aware information to explicitly account for the network’s training behavior. To address these challenges, this paper proposes a Neural Tangent Kernel (NTK)-aware iterative moving collocation point framework. Specifically, we design an NTK-aware point selection strategy that uses training-dynamics-related signals to guide the redistribution of collocation points, prioritizing candidate points that better facilitate the learning of critical solution components. This selection strategy is embedded within a multi-round iterative framework, allowing collocation points to dynamically relocate and track evolving stiff regions. Experimental results on test cases featuring localized stiff regions demonstrate that the proposed framework exhibits superior convergence performance and achieves higher accuracy compared to both standard PINNs and residual-based adaptive methods.
- 23 January 2026
- Published by Association for Computing Machinery (ACM)
- p. 436-443
Open AccessAn NTK-Aware Multi-Round Iterative Moving Collocation Point Framework for Physics-Informed Neural Networks
Physics-Informed Neural Networks (PINNs) have emerged as a mesh-free alternative for solving Partial Differential Equations (PDEs), bypassing the computationally expensive mesh generation required by traditional numerical methods. However, this flexibility comes at a cost: by replacing a fixed mesh with a set of collocation points, PINNs shift the core challenge of spatial discretization to the strategic distribution of these points. Consequently, the arrangement of these collocation points directly impacts the efficacy and convergence speed of PINNs. Conventional sampling strategies, such as uniform distribution or residual-based sampling methods, essentially treat the PDE residual field as a static target and typically rely solely on residual magnitudes, overlooking the intrinsic training dynamics of the neural network. Unlike methods that rely solely on residual magnitudes, our approach leverages NTK-aware information to explicitly account for the network’s training behavior. To address these challenges, this paper proposes a Neural Tangent Kernel (NTK)-aware iterative moving collocation point framework. Specifically, we design an NTK-aware point selection strategy that uses training-dynamics-related signals to guide the redistribution of collocation points, prioritizing candidate points that better facilitate the learning of critical solution components. This selection strategy is embedded within a multi-round iterative framework, allowing collocation points to dynamically relocate and track evolving stiff regions. Experimental results on test cases featuring localized stiff regions demonstrate that the proposed framework exhibits superior convergence performance and achieves higher accuracy compared to both standard PINNs and residual-based adaptive methods.
- Open Access
- 23 January 2026
- Published by Association for Computing Machinery (ACM)
- p. 477-482
CR-DDPM: A Correction-Refined Diffusion Model for Sampling-Based Radio Map Construction
The Radio Map (RM) is a crucial tool for electromagnetic space perception and intelligent spectrum management, which is very important for enabling dynamic resource allocation, mitigating interference conflicts, and optimizing network performance. Currently, several RM construction methods based on convolutional neural network have been developed. Nevertheless, in scenarios where radiation source positions are unascertained, constructing high-precision radio maps (RMs) while resolving interference conflicts and optimizing network efficiency remains challenging. In recent years, a variety of RM construction approaches built on convolutional neural networks (CNNs) have been put forward. To tackle these problems, this study presents a two-stage correction-refined denoising diffusion probabilistic model (CR-DDPM), which leverages a diffusion model to produce RMs with sharpened edge features and integrates a correction-refinement module to offset accuracy degradation. Experiment results demonstrate that CR-DDPM achieves state-of-the-art performance in both reconstruction accuracy and edge feature preservation.
- 23 January 2026
- Published by Association for Computing Machinery (ACM)
- p. 477-482
Open AccessCR-DDPM: A Correction-Refined Diffusion Model for Sampling-Based Radio Map Construction
The Radio Map (RM) is a crucial tool for electromagnetic space perception and intelligent spectrum management, which is very important for enabling dynamic resource allocation, mitigating interference conflicts, and optimizing network performance. Currently, several RM construction methods based on convolutional neural network have been developed. Nevertheless, in scenarios where radiation source positions are unascertained, constructing high-precision radio maps (RMs) while resolving interference conflicts and optimizing network efficiency remains challenging. In recent years, a variety of RM construction approaches built on convolutional neural networks (CNNs) have been put forward. To tackle these problems, this study presents a two-stage correction-refined denoising diffusion probabilistic model (CR-DDPM), which leverages a diffusion model to produce RMs with sharpened edge features and integrates a correction-refinement module to offset accuracy degradation. Experiment results demonstrate that CR-DDPM achieves state-of-the-art performance in both reconstruction accuracy and edge feature preservation.