Latest Issue
Vol. 53,
No. 2,
Feb.
2026
All Issues
A Study on Network-Based Storage Systems: Analysis and Trends
http://doi.org/10.5626/JOK.2026.53.2.91
Large-scale AI and datacenter workloads require storage systems that offer both high performance and scalability. To meet this demand, networked block storage access technologies have evolved, with iSCSI, NVMe-over-TCP, and NVMe-over-RDMA as key protocols. This paper examines the performance and characteristics of these three protocols, focusing on performance bottlenecks and CPU overhead in real server-based storage environments. We also review recent research trends and discuss software optimizations, hardware offload techniques, and server architecture challenges related to NUMA, PCIe, and the adoption of DPUs. Using experimental data and recent studies, we compare the performance and resource consumption of NVMe/TCP and NVMe/RDMA, highlighting their advantages and disadvantages in relation to iSCSI. This study, along with the latest developments, offers valuable insights for protocol selection and architectural strategies for next-generation storage systems.
Design of a Deep Reinforcement Learning Algorithm forSpatially Adaptive UAV Autonomous Navigation based onTransfer Learning
Sungjoon Lee, Gyu Seon Kim, Taejin Woo, Soohyun Park
http://doi.org/10.5626/JOK.2026.53.2.101
This study proposes a deep reinforcement learning algorithm for spatially adaptive unmanned aerial vehicle (UAV) autonomous navigation, utilizing transfer learning to enhance exploration efficiency across various environments. UAVs are vital for both military and civilian missions but face challenges when operating in diverse and dynamic settings. Traditional reinforcement learning methods are inefficient as they necessitate relearning from scratch in new environments. To overcome this limitation, the study implements transfer learning, which allows knowledge gained in one environment to be applied in another, thus improving learning speed and energy efficiency. By integrating Deep Q-Networks (DQN) with transfer learning, UAVs can effectively explore and adapt to different mission areas. Experimental results indicate that the proposed method achieves faster convergence and superior exploration performance compared to existing reinforcement learning techniques, highlighting its potential for practical applications.
Enhancing Stability and Performance of ReinforcementLearning Algorithms through Q Function-based LyapunovStability Constraints
http://doi.org/10.5626/JOK.2026.53.2.109
We introduce a lightweight Q-based stability regularizer for Actor–Critic methods, specifically Soft Actor-Critic (SAC) and TD3. Inspired by Lyapunov-style intuition—though without formal guarantees—this regularizer incorporates a one-sided hinge penalty into the policy loss. This penalty discourages updates that reduce the critic value at on-policy states. The training loop, which includes the replay buffer, target networks, and delayed policy updates, remains unchanged; the only additional computation required is extra forward passes through the target critic(s) during policy updates (one for SAC and two for TD3), resulting in a modest wall-clock time increase of approximately 1-2% and negligible memory overhead in our experiments. We evaluate our approach on MuJoCo tasks, including InvertedPendulum, InvertedDoublePendulum, and HumanoidStandup, using an identical environment-step budget. Performance is assessed along two dimensions: MeanRegret@K (lower is better) for early learning speed and MeanCost@K (lower is better) for safety, with an optional balanced composite measure of 0.5. Across tasks, the regularizer generally improves the trade-off between speed and safety, with particularly consistent enhancements when applied to TD3. Results for the HumanoidStandup task show higher variance due to sensitivity to contact; we present aggregate trends in the main text and detailed distributions in the appendix. Overall, this method should be considered a practical regulation mechanism that complements constraint-based approaches such as CPO and PPO-Lag. Limitations include sensitivity to the penalty weight and reliance on the accuracy of critic estimations.
Fine-Tuning BGE-M3 for Defense Language EmbeddingModel: The Impact of Negative Sample Selection inContrastive Learning
Junsub Kim, Dongnyeok Choi, Sung Gu Kim, Deuk Hwa Kim
http://doi.org/10.5626/JOK.2026.53.2.117
Korean language models specifically designed for the defense sector are still limited, even with the rapid advancements in text embeddings. In this study, we fine-tune the multilingual BGE-M3 model to better understand military terminology and investigate how negative sampling in contrastive learning impacts downstream performance. We evaluate three strategies: Easy (random negatives), Hard (lexicographic adjacency), and Harder (similarity-mined negatives). Our analysis, based on clustering metrics such as Accuracy, NMI, and ARI using a defense news dataset, reveals that the similarity-based Harder strategy consistently outperforms the others. Further evaluations on the KorSTS dataset demonstrate that the Harder approach maintains strong Spearman and Pearson correlations, indicating successful domain adaptation without compromising overall semantic competence. Interestingly, the three Harder variants—negatives mined with BGE-M3, ko-sroberta, and multilingual-e5—produce nearly identical similarity distributions and comparable improvements, while the Easy strategy plateaus and the Hard strategy shows only moderate performance. These findings suggest that mining sufficiently similar negatives, as opposed to using random or adjacent ones, is crucial for effective, domain-specific fine-tuning of multilingual embedding models.
VR Headset User’s Blendshape-driven Facial ExpressionTracking for Parametric Model and Photo-realistic AvatarReconstruction
http://doi.org/10.5626/JOK.2026.53.2.124
This study introduces a novel method for representing the facial expressions of VR headset users through a parametric mesh model and a photo-realistic avatar. While previous research has concentrated on optimizing mesh parameters using video input, such methods are not feasible for VR users due to the absence of accessible RGB imagery. To overcome this limitation, our approach indirectly tracks user expressions by utilizing the blendshape-based avatar expressions provided by the VR headset. Additionally, we create a proxy avatar that resembles the user`s appearance and remove loss terms sensitive to facial geometry, which enhances expression modeling. This method is particularly suitable for VR environments, as it relies solely on headset input, while achieving expression representations comparable to those derived from video-based methods. Ultimately, this approach bridges the gap between realistic mesh-based facial avatars and the expressions of users wearing VR headsets.
TACO: Transformer for Attentive Codon Optimization
http://doi.org/10.5626/JOK.2026.53.2.131
Codon optimization is a crucial technique for enhancing protein expression, particularly important in fields like gene therapy and vaccine development. However, traditional approaches based on codon frequency often overlook the structural stability of mRNA. In practice, optimizing mRNA codons is a multi-objective challenge that must balance expression efficiency and structural stability, while also considering long-range contextual dependencies within extended sequences, making it computationally complex. To tackle this issue, we propose TACO, a deep learning model that combines 1D-CNN and Transformer architectures. TACO optimizes both the Codon Adaptation Index (CAI) and Minimum Free Energy (MFE) through contrastive learning and a joint loss function. We evaluated the model using datasets from Homo sapiens, Escherichia coli, and Saccharomyces cerevisiae. The results show that TACO consistently outperforms existing BiLSTM-based models and popular commercial optimization tools. These findings underscore the potential of our approach as an AI-driven framework for sequence optimization under biologically constrained multi-objective conditions.
Diffusion-based Self-supervised Learning forMulti-behavior Recommendation
http://doi.org/10.5626/JOK.2026.53.2.139
Recently, graph augmentation methods utilizing random dropout have been introduced in graph learning-based recommender systems. However, these methods can dimish recommendation performance by losing crucial user-item interaction signals. To address this issue, we propose a self-supervised recommendation model that leverages diffusion model to generate new user-item interactions. This approach maintains meaningful target behavior signals while effectively reflecting user preferences. Furthermore, we integrate multi-behavior learning to capture informative signals from auxiliary behaviors and mitigate data sparsity. Our experiments on two real-world e-commerce datasets demonstrate that our proposed model outperforms previous studies that generate augmented graphs through random edge dropping, as measured by Hit@10 and Recall@10. This indicates that interaction augmentation via diffusion models can amplify important signals and enhance recommendation performance.
Domain-Adaptive Uncertainty-Aware Temporal BaggingEnsemble for Gait Phase Estimation
Gisu Heo, Yong Ki Son, YoungJun Kim, Changmok Oh, Dong-Woo Lee
http://doi.org/10.5626/JOK.2026.53.2.148
This study introduces an uncertainty-aware Temporal Bagging Ensemble (UA-TBE) designed to robustly estimate continuous gait phases (0-100%) across various walking environments using wearable multi-sensor data, which includes an IMU, FSRs, and soft bending sensors. The proposed model achieves this by: (1) determining environment probabilities and domain uncertainty through a walking-environment classifier, (2) integrating predictions from sub-models trained with different temporal window ratios, and (3) quantifying both aleatoric and epistemic uncertainties, which are then combined using adaptive ensemble weighting. Utilizing a dataset collected at 100 Hz with 72-frame inputs across six walking environments (standing, level walking, ramp ascent/descent, and stair ascent/descent), UA-TBE demonstrated improvements of 69.8% in MAE, 71.1% in RMSE, and 2.5% in R² compared to a simple averaging ensemble. Furthermore, UA-TBE effectively mitigated overconfidence during transitions between environments and enhanced the contribution of high-performing models in high-certainty regions.
Dual Storage Engines toward Multi-Model Analytic Workloads
Kyoseung Koo, Yoojin Choi, Bogyeong Kim, Bongki Moon
http://doi.org/10.5626/JOK.2026.53.2.158
Modern data analytic workloads increasingly require handling multiple data models simultaneously. Two primary approaches meet this need: polyglot persistence and multi-model database systems. However, they are limited by high communication costs due to the physical disaggregation of the system or inefficient query processing stemming from reliance on a single engine. To address these limitations, we present DSE, a multi-model analytic system with integrated storage engines optimized for each model. DSE treats all data models as first-class entities, composing query plans that incorporate operations across models. To effectively combine data from different models, the system introduces a specialized inter-model join algorithm called multi-stage hash join. Our evaluation demonstrates that DSE outperforms existing approaches on multi-model analytics, confirming the effectiveness of our proposed techniques.
Anomaly Detection in Robot Operating System(ROS) TopicData for Unmanned Ground Vehicles(UGV) Using a Transformer
http://doi.org/10.5626/JOK.2026.53.2.172
The Robot Operating System (ROS) is a fundamental framework for unmanned ground vehicles (UGVs), particularly in facilitating autonomous driving by enabling data exchange among various sensors and control topics. As the development and deployment of autonomous driving technologies advance, UGVs face increased risks from threats such as denial-of-service (DoS) and spoofing attacks. In this study, we propose a method to detect these threats by collecting topic data from the ROS running on a UGV and identifying abnormal patterns within this data. To capture the complex relationships present, we utilize a transformer-based autoencoder. This model is trained on normal operating data to learn both temporal and inter-topic relationships, and detects anomalies by analyzing reconstruction errors that arise when abnormal patterns occur. We validate the model’s effectiveness by training it with normal data collected from a multi-purpose UGV and testing it against attack data generated from cyberattacks on the same vehicle.
Search
Journal of KIISE
- ISSN : 2383-630X(Print)
- ISSN : 2383-6296(Electronic)
- KCI Accredited Journal
Editorial Office
- Tel. +82-2-588-9240
- Fax. +82-2-521-1352
- E-mail. chwoo@kiise.or.kr