Differentiable neural architecture search (NAS) greatly accelerates the architecture search while re-taining enough search space. However, existing differentiable NAS is vague in distinguishing candidate operations using the relative magnitude of architectural parameters and suffers from instability and low performance. In this paper, we propose a novel probabilistic framework for differentiable NAS, named variational dropout for neural architecture search (VDNAS), that leverages variational dropout to achieve reformulated super-net sparsification for differentiable NAS. We propose a hierarchical structure to simultaneously enable operation sampling and explicit topology optimization via variational dropout. Specifically, for operation sampling, we develop semi-implicit variational dropout to enable selection of multiple operations and suppress the over-selection of skip-connect operation. We introduce embedded sigmoid relaxation to alleviate the biased gradient estimation in semi-implicit variational dropout to ensure the stability in sampling of architectures and optimization of architectural parameters. Furthermore, we design operation reparameterization to aggregate multiple sampling operations on the same edge to improve the shallow and wide architectures induced by multiple-operation sampling and enhance the transferring ability to large-scale datasets. Experimental results demonstrate that the proposed approaches achieve state-of-the-art performance with top-1 error rates of 2.45% and 15.76% on CIFAR-10/100. Remarkably, when transferred to ImageNet, the proposed approaches searched on CIFAR-10 outperform existing methods searched directly on ImageNet with only 10% of the search cost.
{"title":"Variational Dropout for Differentiable Neural Architecture Search","authors":"Yaoming Wang;Yuchen Liu;Wenrui Dai;Chenglin Li;Xiaopeng Zhang;Junni Zou;Hongkai Xiong","doi":"10.23919/cje.2024.00.183","DOIUrl":"https://doi.org/10.23919/cje.2024.00.183","url":null,"abstract":"Differentiable neural architecture search (NAS) greatly accelerates the architecture search while re-taining enough search space. However, existing differentiable NAS is vague in distinguishing candidate operations using the relative magnitude of architectural parameters and suffers from instability and low performance. In this paper, we propose a novel probabilistic framework for differentiable NAS, named variational dropout for neural architecture search (VDNAS), that leverages variational dropout to achieve reformulated super-net sparsification for differentiable NAS. We propose a hierarchical structure to simultaneously enable operation sampling and explicit topology optimization via variational dropout. Specifically, for operation sampling, we develop semi-implicit variational dropout to enable selection of multiple operations and suppress the over-selection of skip-connect operation. We introduce embedded sigmoid relaxation to alleviate the biased gradient estimation in semi-implicit variational dropout to ensure the stability in sampling of architectures and optimization of architectural parameters. Furthermore, we design operation reparameterization to aggregate multiple sampling operations on the same edge to improve the shallow and wide architectures induced by multiple-operation sampling and enhance the transferring ability to large-scale datasets. Experimental results demonstrate that the proposed approaches achieve state-of-the-art performance with top-1 error rates of 2.45% and 15.76% on CIFAR-10/100. Remarkably, when transferred to ImageNet, the proposed approaches searched on CIFAR-10 outperform existing methods searched directly on ImageNet with only 10% of the search cost.","PeriodicalId":50701,"journal":{"name":"Chinese Journal of Electronics","volume":"34 4","pages":"1247-1264"},"PeriodicalIF":3.0,"publicationDate":"2025-07-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=11151177","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144990312","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":4,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Aiming at the problems of autonomous decision-making and local convergence that occur in traditional reinforcement learning in the unmanned aerial vehicle (UAV) target tracking, this paper proposes a parallel multi-demonstrations generative adversarial imitation reinforcement learning algorithm to achieve control of UAVs and allow them quickly track the target. First, we classify different expert demonstrations according to different tasks to maximize the model to learn all expert experience. In addition, we develop a parallel multi-demonstrations training framework based on generative adversarial imitation learning, and design strategy update methods for different types of generators, which ensures the generalization ability of imitation learning while improving training efficiency. Finally, we integrate deep reinforcement learning with imitation learning. During the initial training phase, our focus lies in imitation learning while periodically transferring expert knowledge to the pool of reinforcement learning experiences. In the later stages, we increase the proportion of reinforcement learning training and achieve effective UAV target tracking through fine-tuning the weights obtained from reinforcement learning. Experimental results demonstrate that compared to existing reinforcement learning algorithms, our algorithm effectively mitigates issues such as local convergence and completes training in a shorter time frame, ensuring stable target tracking by UAVs.
{"title":"A Parallel Multi-Demonstrations Generative Adversarial Imitation Learning Approach on UAV Target Tracking Decision","authors":"Haohui Zhang;Bo Li;Jingyi Huang;Chao Song;Pingkuan He;Evgeny Neretin","doi":"10.23919/cje.2024.00.082","DOIUrl":"https://doi.org/10.23919/cje.2024.00.082","url":null,"abstract":"Aiming at the problems of autonomous decision-making and local convergence that occur in traditional reinforcement learning in the unmanned aerial vehicle (UAV) target tracking, this paper proposes a parallel multi-demonstrations generative adversarial imitation reinforcement learning algorithm to achieve control of UAVs and allow them quickly track the target. First, we classify different expert demonstrations according to different tasks to maximize the model to learn all expert experience. In addition, we develop a parallel multi-demonstrations training framework based on generative adversarial imitation learning, and design strategy update methods for different types of generators, which ensures the generalization ability of imitation learning while improving training efficiency. Finally, we integrate deep reinforcement learning with imitation learning. During the initial training phase, our focus lies in imitation learning while periodically transferring expert knowledge to the pool of reinforcement learning experiences. In the later stages, we increase the proportion of reinforcement learning training and achieve effective UAV target tracking through fine-tuning the weights obtained from reinforcement learning. Experimental results demonstrate that compared to existing reinforcement learning algorithms, our algorithm effectively mitigates issues such as local convergence and completes training in a shorter time frame, ensuring stable target tracking by UAVs.","PeriodicalId":50701,"journal":{"name":"Chinese Journal of Electronics","volume":"34 4","pages":"1185-1198"},"PeriodicalIF":3.0,"publicationDate":"2025-07-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=11151173","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144990198","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":4,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2025-07-01DOI: 10.23919/cje.2024.00.260
Wentao Duan;Min Zhi;Ping Ping;Yuening Zhang;Xuanhao Qi;Wei Hu;Zhe Lian
Efficient decoupling of rich facial features is crucial in the realm of cross-age face recognition. A novel strategy for cross-age facial recognition is proposed, focusing on the dual decoupling of multilevel features to optimize the extraction and processing of identity-related features. The method begins with multilevel feature extraction on facial images through convolutional neural networks, acquiring a series of low-dimensional and high-dimensional hybrid features, which are then effectively integrated. Subsequently, these fused features are introduced into both linear and nonlinear decomposition units. Under the supervision of multitask training, features related to individual identities are decoupled. Finally, the extracted identity features are utilized to perform cross-age facial recognition tasks. When evaluated on multiple standard cross-age facial recognition datasets and standard universal facial recognition datasets, the method demonstrates high accuracy, highlighting its significant advantages in effectiveness and generalizability.
{"title":"Dual-Decoupling and Multi-Level Feature Integration for Cross-Age Face Recognition","authors":"Wentao Duan;Min Zhi;Ping Ping;Yuening Zhang;Xuanhao Qi;Wei Hu;Zhe Lian","doi":"10.23919/cje.2024.00.260","DOIUrl":"https://doi.org/10.23919/cje.2024.00.260","url":null,"abstract":"Efficient decoupling of rich facial features is crucial in the realm of cross-age face recognition. A novel strategy for cross-age facial recognition is proposed, focusing on the dual decoupling of multilevel features to optimize the extraction and processing of identity-related features. The method begins with multilevel feature extraction on facial images through convolutional neural networks, acquiring a series of low-dimensional and high-dimensional hybrid features, which are then effectively integrated. Subsequently, these fused features are introduced into both linear and nonlinear decomposition units. Under the supervision of multitask training, features related to individual identities are decoupled. Finally, the extracted identity features are utilized to perform cross-age facial recognition tasks. When evaluated on multiple standard cross-age facial recognition datasets and standard universal facial recognition datasets, the method demonstrates high accuracy, highlighting its significant advantages in effectiveness and generalizability.","PeriodicalId":50701,"journal":{"name":"Chinese Journal of Electronics","volume":"34 4","pages":"1321-1330"},"PeriodicalIF":3.0,"publicationDate":"2025-07-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=11151250","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144990279","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":4,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2025-07-01DOI: 10.23919/cje.2024.00.014
Bosong Huang;Weiting Zhang;Ruibin Guo;Nian Tang;Wenhao Ye;Jian Jin
In this paper, a dueling double deep Q network (D3QN)-based resource scheduling algorithm is proposed for industrial Internet of things (IoT) to achieve the flexible adaptation of network resources. In the considered network scenario, the time-sensitive networking (TSN)-fifth generation (TSN-5G) network architecture, primarily composed of TSN switches and 5G base stations, is designed accordingly. Simulation results show that when network resources are limited, the D3QN-based resource scheduling algorithm can significantly improve the efficiency of task allocation, making it an ideal solution for reducing latency and optimizing resource utilization in industrial IoT.
{"title":"Industrial Deterministic Computation and Networking Resource Scheduling via Deep Reinforcement Learning","authors":"Bosong Huang;Weiting Zhang;Ruibin Guo;Nian Tang;Wenhao Ye;Jian Jin","doi":"10.23919/cje.2024.00.014","DOIUrl":"https://doi.org/10.23919/cje.2024.00.014","url":null,"abstract":"In this paper, a dueling double deep Q network (D3QN)-based resource scheduling algorithm is proposed for industrial Internet of things (IoT) to achieve the flexible adaptation of network resources. In the considered network scenario, the time-sensitive networking (TSN)-fifth generation (TSN-5G) network architecture, primarily composed of TSN switches and 5G base stations, is designed accordingly. Simulation results show that when network resources are limited, the D3QN-based resource scheduling algorithm can significantly improve the efficiency of task allocation, making it an ideal solution for reducing latency and optimizing resource utilization in industrial IoT.","PeriodicalId":50701,"journal":{"name":"Chinese Journal of Electronics","volume":"34 4","pages":"1275-1283"},"PeriodicalIF":3.0,"publicationDate":"2025-07-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=11151185","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144990096","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":4,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2025-07-01DOI: 10.23919/cje.2024.00.256
Xiao-Xue Zhang;Ri-Gui Zhou;Wen-Shan Xu
Semi-quantum communication, serving as transitional technology between quantum communication and classical communication, bridges fully quantum-capable users with “classical” users who have limited quantum capabilities. It provides practical solution for application scenarios that struggle to bear high costs of quantum resources. This paper designs a novel deterministic secure semi-quantum communication protocol that significantly enhances communication efficiency by utilizing Greenberger-Horne-Zeilinger (GHZ) states for entanglement compression. The protocol consists of two core components: eavesdropping detection mechanism and transmission process for compressed and encrypted information sequences. During the eavesdropping detection phase, the protocol incorporates decoy photon technology to effectively expose and prevent potential eavesdropping attempts. In the secret information transmission phase, the protocol combines the advantages of a pseudo-random number generator driven by one-way hash function and GHZ-state-based entanglement compression technology to randomly rearrange and compress-encrypt the secret information, ensuring its high security and integrity during transmission. Ultimately, the receiver can accurately decrypt and restore the original secret information using a pre-agreed key. The protocol not only successfully integrates multiple advanced technologies to resist various attacks and ensure the absolute secure transmission of secret information, but also provides strong support for practical applications with its communication efficiency of up to 50% and high practicality.
{"title":"Novel Deterministic Secure Semi-Quantum Communication Based on GHZ-State Entanglement Compression Technology","authors":"Xiao-Xue Zhang;Ri-Gui Zhou;Wen-Shan Xu","doi":"10.23919/cje.2024.00.256","DOIUrl":"https://doi.org/10.23919/cje.2024.00.256","url":null,"abstract":"Semi-quantum communication, serving as transitional technology between quantum communication and classical communication, bridges fully quantum-capable users with “classical” users who have limited quantum capabilities. It provides practical solution for application scenarios that struggle to bear high costs of quantum resources. This paper designs a novel deterministic secure semi-quantum communication protocol that significantly enhances communication efficiency by utilizing Greenberger-Horne-Zeilinger (GHZ) states for entanglement compression. The protocol consists of two core components: eavesdropping detection mechanism and transmission process for compressed and encrypted information sequences. During the eavesdropping detection phase, the protocol incorporates decoy photon technology to effectively expose and prevent potential eavesdropping attempts. In the secret information transmission phase, the protocol combines the advantages of a pseudo-random number generator driven by one-way hash function and GHZ-state-based entanglement compression technology to randomly rearrange and compress-encrypt the secret information, ensuring its high security and integrity during transmission. Ultimately, the receiver can accurately decrypt and restore the original secret information using a pre-agreed key. The protocol not only successfully integrates multiple advanced technologies to resist various attacks and ensure the absolute secure transmission of secret information, but also provides strong support for practical applications with its communication efficiency of up to 50% and high practicality.","PeriodicalId":50701,"journal":{"name":"Chinese Journal of Electronics","volume":"34 4","pages":"1172-1184"},"PeriodicalIF":3.0,"publicationDate":"2025-07-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=11151174","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144990043","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":4,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2025-07-01DOI: 10.23919/cje.2024.00.218
Jian Huang;Yanbo Li;Hao Han
Software vulnerabilities, particularly memory corruption, are significant sources of security breaches. Traditional security measures like data-execution prevention, address space layout randomization, control-flow integrity, code-pointer integrity/separation, and data-flow integrity provide insufficient protection or lead to considerable performance degradation. This research introduces, develops, and scrutinizes FastDIM, a novel approach designed to safeguarding user applications from memory corruption threats. FastDIM encompasses an low-level virtual machine (LLVM) instrumentation mechanism and a distinct memory monitoring module. This system modifies applications in user space into a more secure variant, proactively reporting vital memory operations to a memory monitoring component within the kernel to ensure data integrity. Distinctive features of FastDIM compared to prior methodologies are twofold: FastDIM's integrated out-of-band monitoring system that secures both control-flow and non-control data within program memory, and the creation of a dedicated shared memory space to enhance monitoring efficiency. Testing a prototype of FastDIM with a broad spectrum of real-life applications and standard benchmarks indicates that FastDIM's runtime overhead is acceptable, at 4.4% for the SPEC CPU 2017 benchmarks, while providing the defense against memory corruption attacks.
软件漏洞，特别是内存损坏，是安全漏洞的重要来源。传统的安全措施，如数据执行预防、地址空间布局随机化、控制流完整性、代码指针完整性/分离和数据流完整性，不能提供足够的保护或导致相当大的性能下降。本研究介绍、开发并详细分析了FastDIM，这是一种旨在保护用户应用程序免受内存损坏威胁的新方法。FastDIM包含一个低级虚拟机（LLVM）检测机制和一个独特的内存监视模块。该系统将用户空间中的应用程序修改为更安全的变体，主动向内核中的内存监视组件报告重要的内存操作，以确保数据完整性。与之前的方法相比，FastDIM的独特之处在于两个方面：FastDIM的集成带外监控系统，可以在程序内存中保护控制流和非控制数据，并创建专用共享内存空间，以提高监控效率。在广泛的实际应用程序和标准基准测试中测试FastDIM的原型表明，FastDIM的运行时开销是可以接受的，在SPEC CPU 2017基准测试中为4.4%，同时提供对内存损坏攻击的防御。
{"title":"Design, Realization, and Evaluation of FastDIM to Prevent Memory Corruption Attacks","authors":"Jian Huang;Yanbo Li;Hao Han","doi":"10.23919/cje.2024.00.218","DOIUrl":"https://doi.org/10.23919/cje.2024.00.218","url":null,"abstract":"Software vulnerabilities, particularly memory corruption, are significant sources of security breaches. Traditional security measures like data-execution prevention, address space layout randomization, control-flow integrity, code-pointer integrity/separation, and data-flow integrity provide insufficient protection or lead to considerable performance degradation. This research introduces, develops, and scrutinizes FastDIM, a novel approach designed to safeguarding user applications from memory corruption threats. FastDIM encompasses an low-level virtual machine (LLVM) instrumentation mechanism and a distinct memory monitoring module. This system modifies applications in user space into a more secure variant, proactively reporting vital memory operations to a memory monitoring component within the kernel to ensure data integrity. Distinctive features of FastDIM compared to prior methodologies are twofold: FastDIM's integrated out-of-band monitoring system that secures both control-flow and non-control data within program memory, and the creation of a dedicated shared memory space to enhance monitoring efficiency. Testing a prototype of FastDIM with a broad spectrum of real-life applications and standard benchmarks indicates that FastDIM's runtime overhead is acceptable, at 4.4% for the SPEC CPU 2017 benchmarks, while providing the defense against memory corruption attacks.","PeriodicalId":50701,"journal":{"name":"Chinese Journal of Electronics","volume":"34 4","pages":"1233-1246"},"PeriodicalIF":3.0,"publicationDate":"2025-07-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=11151235","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144990196","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":4,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2025-07-01DOI: 10.23919/cje.2024.00.150
Chi Zhang;Zehua Wang;Dongyang Lu;Qian Lu;Yongji Du;Ruozhou Li;Jing Yan;Ying Yu
In this paper, we propose a 1-bit reconfigurable bidirectional transmission/reflection array (TRA) antenna. The TRA adopts a two-layer compact unit structure to achieve transmission and reflection modes simultaneously. Two PIN diodes have been applied to the unit to complement 1-bit quantification for beam scanning. Then a reconfigurable array antenna composed of 14 × 14 elements is designed, manufactured, and measured at 9.2 GHz. The peak gain of transmission and reflection reaches 17.41 dBi and 16.74 dBi, respectively, when the beam scans within the range of ±60°. The aperture efficiency of 10.57% in transmission and 9.06% in reflection have been attained. Based on this, a 1-bit reconfigurable bidirectional dual-beam TRA has been attempted with the same unit. The phase synthesis method is utilized in dual-beam TRA design. The transmission and reflection peak gains for symmetrical dual-beam of equal amplitude reach 13.64 dBi and 12.63 dBi, respectively, with a scanning angle of ±10°, and 13.86 dBi and 13.01 dBi when the scanning angle is ±20°. The results demonstrate that the above designs have potential applications in communication and multi-target radar systems.
{"title":"A Beam Reconfigurable Bidirectional Reflection/Transmission Array Antenna","authors":"Chi Zhang;Zehua Wang;Dongyang Lu;Qian Lu;Yongji Du;Ruozhou Li;Jing Yan;Ying Yu","doi":"10.23919/cje.2024.00.150","DOIUrl":"https://doi.org/10.23919/cje.2024.00.150","url":null,"abstract":"In this paper, we propose a 1-bit reconfigurable bidirectional transmission/reflection array (TRA) antenna. The TRA adopts a two-layer compact unit structure to achieve transmission and reflection modes simultaneously. Two PIN diodes have been applied to the unit to complement 1-bit quantification for beam scanning. Then a reconfigurable array antenna composed of 14 × 14 elements is designed, manufactured, and measured at 9.2 GHz. The peak gain of transmission and reflection reaches 17.41 dBi and 16.74 dBi, respectively, when the beam scans within the range of ±60°. The aperture efficiency of 10.57% in transmission and 9.06% in reflection have been attained. Based on this, a 1-bit reconfigurable bidirectional dual-beam TRA has been attempted with the same unit. The phase synthesis method is utilized in dual-beam TRA design. The transmission and reflection peak gains for symmetrical dual-beam of equal amplitude reach 13.64 dBi and 12.63 dBi, respectively, with a scanning angle of ±10°, and 13.86 dBi and 13.01 dBi when the scanning angle is ±20°. The results demonstrate that the above designs have potential applications in communication and multi-target radar systems.","PeriodicalId":50701,"journal":{"name":"Chinese Journal of Electronics","volume":"34 4","pages":"1100-1110"},"PeriodicalIF":3.0,"publicationDate":"2025-07-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=11151184","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144990314","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":4,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2025-07-01DOI: 10.23919/cje.2024.00.202
Ling Yi;Yongbin Qin
Adaptive bitrate (ABR) algorithms are crucial for video streaming services by dynamically adjusting video bitrate based on current network conditions to ensure better quality of experience (QoE). However, traditional ABR algorithms often face challenges in adapting to diverse network environments and fail to fully utilize expert knowledge. In this study, we propose a novel approach using adversarial inverse reinforcement learning (AIRL) to learn ABR algorithms. Unlike traditional methods, AIRL can effectively leverage expert demonstrations to learn robust reward functions and generate stable ABR policies. Simultaneously, the learned ABR policy adjusts based on the updated reward function, aiming to closely emulate the video bitrate decision-making behavior of experts. Moreover, by decoupling the reward function, we can develop a more robust ABR strategy that can effectively adapt video bitrates to significant fluctuations in network conditions, while also optimizing different video QoE objectives. We conducted experiments across various network conditions, demonstrating that the proposed method exhibits stable and superior performance.
{"title":"Learning Robust Adaptive Bitrate Algorithms with Adversarial Inverse Reinforcement Learning","authors":"Ling Yi;Yongbin Qin","doi":"10.23919/cje.2024.00.202","DOIUrl":"https://doi.org/10.23919/cje.2024.00.202","url":null,"abstract":"Adaptive bitrate (ABR) algorithms are crucial for video streaming services by dynamically adjusting video bitrate based on current network conditions to ensure better quality of experience (QoE). However, traditional ABR algorithms often face challenges in adapting to diverse network environments and fail to fully utilize expert knowledge. In this study, we propose a novel approach using adversarial inverse reinforcement learning (AIRL) to learn ABR algorithms. Unlike traditional methods, AIRL can effectively leverage expert demonstrations to learn robust reward functions and generate stable ABR policies. Simultaneously, the learned ABR policy adjusts based on the updated reward function, aiming to closely emulate the video bitrate decision-making behavior of experts. Moreover, by decoupling the reward function, we can develop a more robust ABR strategy that can effectively adapt video bitrates to significant fluctuations in network conditions, while also optimizing different video QoE objectives. We conducted experiments across various network conditions, demonstrating that the proposed method exhibits stable and superior performance.","PeriodicalId":50701,"journal":{"name":"Chinese Journal of Electronics","volume":"34 4","pages":"1309-1320"},"PeriodicalIF":3.0,"publicationDate":"2025-07-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=11151225","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144990280","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":4,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2025-07-01DOI: 10.23919/cje.2024.00.228
Xin Zhang;Haoyang Zhang;Ruizhe Yang;Lilin Li;Donglin Su
High-dimensional electromagnetic optimization problems, such as array antenna design, pose significant challenges due to their complexity and high dimensionality. The Maxwell's equations derived optimization (MEDO) algorithm, a novel optimization method with strong performance in electromagnetics, experiences a decline in efficiency as the problem dimensionality increases. To address these challenges, graphics processing unit (GPU)-accelerated MEDO algorithm with differential grouping (DG-GMEDO) is proposed in this paper, which builds on the MEDO algorithm through the integration of an enhanced differential grouping strategy and GPU-based parallel acceleration. This approach allows for more effective management of variable interactions while leveraging high computational speeds. Comparative evaluations with traditional algorithms like particle swarm optimization and genetic algorithm, as well as state-of-the-art methods such as MAES2-EDG, GTDE, RCI-PSO, and CCFR2-IRRG, highlight its competitive performance in terms of both accuracy and efficiency. Furthermore, DG-GMEDO demonstrates significant runtime acceleration and achieves promising results in high-dimensional settings, as validated through its application in array antenna radiation patterns optimization.
{"title":"GPU-Accelerated MEDO Algorithm with Differential Grouping (DG-GMEDO) for High-Dimensional Electromagnetic Optimization Problems","authors":"Xin Zhang;Haoyang Zhang;Ruizhe Yang;Lilin Li;Donglin Su","doi":"10.23919/cje.2024.00.228","DOIUrl":"https://doi.org/10.23919/cje.2024.00.228","url":null,"abstract":"High-dimensional electromagnetic optimization problems, such as array antenna design, pose significant challenges due to their complexity and high dimensionality. The Maxwell's equations derived optimization (MEDO) algorithm, a novel optimization method with strong performance in electromagnetics, experiences a decline in efficiency as the problem dimensionality increases. To address these challenges, graphics processing unit (GPU)-accelerated MEDO algorithm with differential grouping (DG-GMEDO) is proposed in this paper, which builds on the MEDO algorithm through the integration of an enhanced differential grouping strategy and GPU-based parallel acceleration. This approach allows for more effective management of variable interactions while leveraging high computational speeds. Comparative evaluations with traditional algorithms like particle swarm optimization and genetic algorithm, as well as state-of-the-art methods such as MAES2-EDG, GTDE, RCI-PSO, and CCFR2-IRRG, highlight its competitive performance in terms of both accuracy and efficiency. Furthermore, DG-GMEDO demonstrates significant runtime acceleration and achieves promising results in high-dimensional settings, as validated through its application in array antenna radiation patterns optimization.","PeriodicalId":50701,"journal":{"name":"Chinese Journal of Electronics","volume":"34 4","pages":"1052-1063"},"PeriodicalIF":3.0,"publicationDate":"2025-07-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=11151224","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144989944","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":4,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2025-07-01DOI: 10.23919/cje.2024.00.220
Yiwen Wang;Qianfan Wang;Jifan Liang;Xiao Ma
In this paper, we propose an algorithm to transform a generator matrix of a linear block code into a minimum weight staircase generator matrix (MWSGM). This allows us to apply the representative ordered statistics decoding with local constraints (LC-ROSD) algorithm to cyclic redundancy check (CRC) aided polar (CA-polar) codes. Distinguished from the conventional ordered statistics decoding (OSD), the LC-ROSD implements parallel Gaussian elimination (GE) for MWSGM instead of serial GE for a general matrix, potentially reducing the decoding latency. Numerical results show that the performance of 5G CA-polar codes under LC-ROSD is better than that of CRC aided successive cancellation list (CA-SCL) decoding and can approach the corresponding maximum likelihood (ML) lower bounds at different code rates. The numerical results also show that the LC-ROSD with MWSGM has lower decoding complexity than the CA-SCL decoding in the high signal-to-noise ratio (SNR) region.
{"title":"Representative OSD with Local Constraints of CA-Polar Codes","authors":"Yiwen Wang;Qianfan Wang;Jifan Liang;Xiao Ma","doi":"10.23919/cje.2024.00.220","DOIUrl":"https://doi.org/10.23919/cje.2024.00.220","url":null,"abstract":"In this paper, we propose an algorithm to transform a generator matrix of a linear block code into a minimum weight staircase generator matrix (MWSGM). This allows us to apply the representative ordered statistics decoding with local constraints (LC-ROSD) algorithm to cyclic redundancy check (CRC) aided polar (CA-polar) codes. Distinguished from the conventional ordered statistics decoding (OSD), the LC-ROSD implements parallel Gaussian elimination (GE) for MWSGM instead of serial GE for a general matrix, potentially reducing the decoding latency. Numerical results show that the performance of 5G CA-polar codes under LC-ROSD is better than that of CRC aided successive cancellation list (CA-SCL) decoding and can approach the corresponding maximum likelihood (ML) lower bounds at different code rates. The numerical results also show that the LC-ROSD with MWSGM has lower decoding complexity than the CA-SCL decoding in the high signal-to-noise ratio (SNR) region.","PeriodicalId":50701,"journal":{"name":"Chinese Journal of Electronics","volume":"34 4","pages":"1111-1119"},"PeriodicalIF":3.0,"publicationDate":"2025-07-01","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=11151187","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"144990059","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":4,"RegionCategory":"计算机科学","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"OA","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
京公网安备 11010802042870号
京ICP备2023020795号-1
扫码关注我们
求助内容：
应助结果提醒方式：