Daehan Won

Large Neighborhood Search (LNS) is an effective primal-search strategy for mixed-integer linear programming (MILP), but its performance depends critically on both the size of the neighborhood and the variables selected for release. Existing learning-based approaches typically adapt neighborhood size or learn variable-selection policies, while often reusing a single variable ranking across different neighborhood sizes. This work investigates whetherthese two decisions should instead be explicitly co-adapted. We propose CoAda-LNS, a hierarchical graph-neural framework that first predicts a feasible neighborhood size and then conditions the variable-ranking policy on the selected size. A bipartite graph neural network encodes the MILP structure and search state, while true multi-radius Local Branching is used as an optimization-based expert to generate both neighborhood-size and destroy set supervision. The method is evaluated on an instance-disjoint MIPLIB 2017 benchmark and compared against an adaptive-k shared-ranking baseline that uses the same size-selection mechanism but a size-independent variable ranking. Across 102 valid paired trajectories, the shared-ranking baseline achieves a lower reference-normalized primal integral on 33 instances, CoAda-LNS on 25, while 44 instances are tied. The shared-ranking policy also attains lower mean and median primal integral (74.47 and 85.42) than CoAda-LNS (76.77 and 90.68). However, the paired Wilcoxon signed-rank test does not indicate a statistically significant difference between the methods (p = 0.223, rank-biserial effect = -0.184), and the exact sign test similarly shows no significant difference in win frequency (p = 0.358). Theseresults suggest that although the shared-ranking baseline exhibits somewhat better descriptive anytime performance, the observed differences are not statistically significant under the current primal-integral evaluation. Thus, the additional representational flexibility provided by size-conditioned ranking does not yield a consistent downstream LNS advantage under the current training and evaluation regime. More broadly, the findings highlight thedistinction between representing richer expert neighborhood structure and achieving stronger end-to-end optimization behavior.

Human and veterinary healthcare systems face many parallel challenges, yet opportunities for cross-sectoral learning remain underexplored. This scoping review aimed to compare healthcare delivery in both domains, highlighting operational inefficiencies, shared challenges, and innovative strategies that may enhance outcomes in both sectors. A search and review of 43,987 articles was conducted across multiple databases. After applying inclusion and exclusion criteria and conducting quality assessments, 100 studies were selected for detailed analysis. The review focused on operational elements such as patient scheduling, staffing, electronic health records, patient flow, satisfaction, wait times, length of stay, facility layout, and resource allocation. The review identified recurring inefficiencies in both sectors, notably in scheduling and resource planning. Common operational challenges in healthcare include poor electronic health records integration, suboptimal facility layout, and understaffing. Innovative solutions like improved protocols, Lean Six Sigma, smart scheduling, artificial intelligence, and ergonomics can improve efficiency and patient outcomes. The review findings underscored the potential for shared solutions between human and veterinary healthcare systems. Recognizing and applying successful strategies across both fields could improve operational efficiency, patient satisfaction, and overall quality of care. Veterinary healthcare lags in adopting industrial engineering solutions to solve operational challenges. This review provided invaluable information from the literature and could serve as a blueprint in multiple areas to expedite improvements in veterinary care. Human healthcare systems are much more standardized than veterinary systems and provide multiple potential solutions to improve veterinary care delivery.

Surface Mount Technology (SMT) has transformed modern electronics manufacturing, enabling the production of compact, reliable, and high-performance devices. Among the critical phenomena in SMT, the self-alignment effect during reflow soldering plays a pivotal role in ensuring precise component placement through the principles of energy minimization. Despite its importance, current research primarily focuses on numerical predictions of self-alignment metrics, leaving a significant gap in generating visual representations, such as Post-Automated Optical Inspection (AOI) images, to comprehensively depict alignment outcomes. In this study, we introduce a novel framework that combines physics-based modeling with advanced deep learning to address this gap. Leveraging data from Solder Paste Inspection (SPI), Pre-Automated Optical Inspection (Pre-AOI), and Post-Automated Optical Inspection (Post-AOI) stages, we developed an energy minimization-based physics model to produce displacement vectors capturing self-alignment dynamics. We integrated these inputs, along with key component attributes, into PhyViT-GAN, a Physics-Guided Variational Autoencoder, and MobileViT as the discriminator in adversarial training for generating high-fidelity Post-AOI images. Our approach effectively visualizes the self-alignment effect, achieving a Structural Similarity Index Measure (SSIM) of 0.8348 and a Learned Perceptual Image Patch Similarity (LPIPS) score of 0.0917, demonstrating high structural accuracy and perceptual fidelity. This work represents the first application of deep learning to generate Post-AOI images for visualizing self-alignment in SMT. This framework bridges the gap between physics-driven predictions and visual inspections, offering transformative potential for SMT manufacturing. It enables enhanced defect detection, virtual prototyping, and process optimization, laying the foundation for future advancements in automated and data-driven electronics assembly.