https://scholarworks.bwise.kr/erica/handle/2021.sw.erica/41 2026-04-04T17:30:44Z https://scholarworks.bwise.kr/erica/handle/2021.sw.erica/127199 Title: A meta-learning enhanced deep reinforcement learning approach for generalizing across orienteering problem with time windows Authors: BARDE STEPHANE; 김현준 Abstract: The Orienteering Problem with Time Windows (OPTW) is a complex combinatorial optimization problem with applications in logistics, tourist route planning, and emergency services. Traditional methods for solving OPTW, including metaheuristics, often struggle with scalability, adaptability, and generalization to new instances. Recently, deep reinforcement learning (DRL) has shown promise in tackling routing problems. However, existing DRL methods typically rely on non-Markovian state representations and handcrafted masking rules, which limit their adaptability and generalization. This paper presents Meta Pointer Network for OPTW (MetaPNet-OPTW), a meta-learning-enhanced DRL framework that combines a Markovian state formulation with OR-based feasibility rules within a pointer network model. We introduce the Meta-Learning enhanced REINFORCE algorithm, which learns across diverse problem instances and enables rapid adaptation to unseen configurations with minimal fine-tuning. During inference, active search with beam search is used to refine solutions dynamically. Extensive experiments show that MetaPNet-OPTW outperforms existing DRL approaches in efficiency and generalization, and notably improves 20 of 33 best-known solutions on the Gavalas benchmark. We further provide a t-SNE analysis of the learned latent space, enriched with spatio-temporal statistics, which explains why the model excels on Gavalas instances while identifying harder clusters such as r2 and c2. This study contributes a scalable DRL framework for OPTW that not only achieves state-of-the-art performance but also provides new interpretability into benchmark difficulty and model adaptability. 2026-02-01T00:00:00Z https://scholarworks.bwise.kr/erica/handle/2021.sw.erica/126464 Title: Edge-compatible SOH estimation for Li-ion batteries via hybrid knowledge distillation and model compression Authors: BARDE STEPHANE Abstract: Accurate and efficient State of Health (SOH) estimation is essential for the reliability of lithium-ion batteries in electric vehicles (EVs). However, deploying deep learning models on a real-world Battery Management System (BMS) remains challenging due to edge device constraints. In this study, we propose a lightweight SOH estimation framework integrating knowledge distillation (KD), structured pruning, and dynamic quantization. Our KD approach employs a hybrid strategy, combining response-based loss with two relation-based losses (pairwise squared Euclidean distance and cosine similarity) in the latent feature space. This ensures the student model mimics not only the teacher’s outputs but also its internal data representation structure. Comprehensive experiments on the NASA and CALCE datasets demonstrate the framework’s effectiveness. Our compressed models achieve over 99% model compression while consistently outperforming a representative MobileNetV1-based lightweight baseline in both accuracy and compactness. The practical feasibility of our framework is further validated through on-device performance tests on a Raspberry Pi 4B, robustness analysis under various noise conditions, and an investigation showing a strong correlation between the learned latent space and physical degradation indicators. These results confirm that our framework produces highly efficient, robust, and physically meaningful models suitable for real-time, on-device battery health monitoring. 2025-11-01T00:00:00Z https://scholarworks.bwise.kr/erica/handle/2021.sw.erica/126562 Title: Enhancing battery SOH prediction with Butler–Volmer informed neural networks in data-scarce environments Authors: Seo, Younggeon; Kim, Taeyi; Barde, Stephane Abstract: Accurate and robust estimation of Lithium-Ion Battery (LIB) state of health (SOH) is critical for the safety and reliability of electric vehicles, yet conventional machine learning models often suffer from limited interpretability and poor generalization under data-scarce conditions. In this work, we propose Butler–Volmer Informed Neural Network (BVINN), a physics-informed machine learning framework that directly embeds the closed-form analytical solution of the classical Butler–Volmer equation into the network loss function as a regularization term. By penalizing deviations from fundamental electrochemical kinetics during training, BVINN enforces physically consistent representations of degradation and narrows the solution space to regions that are both statistically and mechanistically plausible. The overall loss combines five components, namely data loss, Butler–Volmer loss, initial condition loss, boundary condition loss and regeneration loss, to ensure adherence to governing physical laws throughout the learning process. We validate BVINN on two well-known benchmark datasets, the NASA battery dataset and the BIT dataset, across various sequence length environments. Using BVINN yields improved RMSE, and MAE performance compared to models without the Butler–Volmer informed regularization. Furthermore, by exploiting the Butler–Volmer equation as an explicit regularizer, BVINN outperforms benchmark methods such as Deep Hidden Physics Model (DeepHPM) and other Physics-Informed Neural Network (PINN), maintaining higher performance even in data-scarce environments. The physical plausibility of the model is further corroborated by electrochemical analysis, which confirms that BVINN learns mechanistically consistent degradation behaviors. 2025-10-01T00:00:00Z https://scholarworks.bwise.kr/erica/handle/2021.sw.erica/126334 Title: Balancing yield and makespan in wafer fabrication: A two-stage data-driven scheduling approach Authors: Kim, Min-geol; Kim, Hyunjoon; Barde, Stephane R.A.; Lee, Chang-ho Abstract: In semiconductor manufacturing, achieving high quality and productivity remains a challenging task due to the complexity and variability of multistage production processes. This study addresses the hybrid flow shop scheduling problem (HFSP) in wafer fabrication, targeting the inherent trade-off between quality (yield) and productivity (makespan). We propose a two-stage data-driven scheduling framework that integrates historical manufacturing data. In the first stage, sequential patterns are mined using the PrefixSpan algorithm and are statistically validated. Based on their yield, patterns are classified and recombined via rule-based filtering to derive plausible high-quality (PHQ) paths. In the second stage, the PHQ path-based HFSP is formulated and solved using GAInS, a hybrid metaheuristic framework that incorporates Genetic Algorithm (GA), Iterated Local Search, and Simulated Annealing. Computational experiments across various wafer counts (N=5,15,25,50) demonstrate that GAInS consistently outperforms Mixed Integer Linear Programming, Constraint Programming models, and basic GA approaches in minimizing makespan while maintaining high yield. Compared to an existing method in the literature that combines regression-based yield prediction with GA-based scheduling, the proposed approach achieves superior Pareto solutions by better balancing quality and productivity. These findings highlight the potential of the proposed framework in balancing critical objectives in wafer fabrication. © 2025 Elsevier B.V., All rights reserved. 2025-10-01T00:00:00Z