INU-TUMSAT International Symposium on Logistics and SCM

Incheon National University (INU) has been at the forefront of advancing education and research in the fields of logistics and supply chain management (SCM) through major government-funded initiatives, including the Regional Innovation Strategy in Education (RISE) Project and the Brain Korea 21 (BK21) Program. 

As part of its strategic commitment to building a robust global cooperation network, INU is co-organizing an international symposium with its long-standing partner institution, Tokyo University of Marine Science and Technology (TUMSAT). TUMSAT is one of Japan's leading universities in maritime science, transportation, and logistics, making it a natural partner for collaborative initiatives in these domains.

The symposium aims to serve as a platform for faculty members from both universities to share their latest research outcomes, exchange pedagogical insights, and explore actionable collaboration opportunities. Beyond research presentations, the event will facilitate in-depth discussions on cooperative strategies spanning undergraduate and graduate curricula, joint research projects, and long-term institutional partnership frameworks.

Event Details

• Event Title: INU-TUMSAT International Symposium on Logistics and Supply Chain Management

• Hosted by: Graduate School of Logistics, Incheon National University | RISE (Regional Innovation System & Education) Project Office, Incheon National University | BK21 (Brain Korea) Program, Incheon National University

• Date: Tuesday, April 7, 2026 (3pm to 6pm)

• Venue: Multiplex room at Library, Songdo Campus, Incheon National University

• Target Audience: Faculty members, graduate and undergraduate students

• Key Themes: AI and Mathematical Optimization in SCM, Maritime Shipping Operations and Supply Chain Contracts, Smart Port Operations and Logistics Innovation

SCHEDULE

Host and Facilitator

• Vice Dean and Associate Professor Minho Ha (INU Graduate School of Logistics)

Session 1. AI and Optimization (15:00 ~ 16:15)

• (15:00 ~ 15:15) Dean and Professor SangHwa Song, "AI-Augmented Optimization: Learning to Solve, Formulate, and Explain Complex Decision Problems "

• (15:15 ~ 15:45) Associate Professor Yuta Tanoue, "Bernstein's inequality with a proper cover and the upper bound of portfolio VaR and TVaR"

• (15:45 ~ 16:15) Assistant Professor Kazunori Matsui, "Universal Approximation Property of ODENet and ResNet: Toward Theoretical Foundations for Reliable AI"

Session 2. Logistics System Design and Technology (16:30 ~ 17:30)

• (16:30 ~ 17:00) Professor Daisuke Watanabe, "Strategic Optimization of Logistics Networks for Autonomous Trucks in Japan"

• (17:00 ~ 17:30) Assistant Professor Woojun Kim, "Smart Freight Transport Systems: Emerging Technologies, Current Progress, and Future Directions" 

ABSTRACTS

Session 1. AI and Optimization

"AI-Augmented Optimization: Learning to Solve, Formulate, and Explain Complex Decision Problems" by Professor Song

Classical optimization has long provided powerful tools for decision making, including mathematical programming, meta-heuristics, and local search. However, many real-world optimization problems remain difficult to use in practice. Model parameters are often uncertain, problem structures become increasingly complex, and the number of inputs and constraints grows rapidly. As a result, building, solving, and interpreting optimization models often requires substantial expert knowledge, while computation can also be too slow for time-sensitive applications. This seminar discusses how AI can complement traditional optimization to address these limitations. Rather than replacing optimization, AI can make it faster, more accessible, and easier to interpret. I will present three research directions based on my experience with real energy network optimization problems. First, I show how optimal solutions generated by conventional optimization algorithms can be used to train deep learning models that quickly produce high-quality approximate solutions for new but similar instances. Second, I introduce a fine-tuned text-to-formulation approach that converts natural-language problem descriptions into optimization models, lowering the barrier for non-experts to use advanced optimization tools. Third, I discuss an LLM-based framework that helps users explore and explain optimization outcomes through natural-language interaction, enabling root-cause analysis and more transparent decision support. Overall, this talk highlights a practical vision of AI-augmented optimization in which learning models help solve, formulate, and explain complex decision problems more effectively.

"Strategic Optimization of Logistics Networks for Autonomous Trucks in Japan" by Professor Watanabe

This research explores the reform of Japan's logistics by strategically optimizing logistics networks to incorporate emerging technologies such as truck platooning and autonomous trucks. The study formulates the optimal location of logistics hubs for line haul transport in Japan using the hub location problem, considering cost reductions from increased capacity and automation due to truck platooning implementation between hubs. The findings indicate that, under various automation scenarios, as the number of hubs increases, they tend to be located in areas with high traffic demand along the Pacific Ocean. Furthermore, despite technological advancements, the study found no significant differences in location results, suggesting that considering base allocation as a single allocation is approximately feasible.

"Bernstein's inequality with a proper cover and the upper bound of portfolio VaR and TVaR" by Professor Yanoue

This paper develops a new concentration inequality for random variables that satisfy Bernstein's condition in the presence of a "proper cover"([1]). The proper cover, a concept from graph dependency theory, provides a framework for analyzing variables with partial dependence by partitioning them into independent subsets. By leveraging this structure, the proposed inequality generalizes standard concentration results for various distributions, including Gamma, Exponential, and bounded random variables. This theoretical advancement bridges the gap between the sum of purely independent variables and those with complex dependency structures.The study extends these theoretical findings to the field of financial risk management, specifically addressing the challenge of "dependence uncertainty". Estimating the joint distribution of multiple risk factors is often difficult in practice, yet essential for accurate risk assessment. To address this, the paper applies the generalized Bernstein's inequality to derive analytical upper bounds for two critical risk measures: Value-at-Risk (VaR) and Tail Value-at-Risk (TVaR). These bounds are specifically designed for portfolios characterized by an "independent subgroup" structure ([2]), where risk factors are independent across groups but may exhibit internal dependencies. Numerical simulations were conducted to verify the practical effectiveness of these derived upper bounds. The results demonstrate that the proposed bounds for both VaR and TVaR are tight and offer significant practical and theoretical benefits for rigorous risk assessment. By integrating probabilistic concentration inequalities with financial risk measurement tools, this research provides a robust mathematical framework for managing risks in portfolios with dependent structures.

Session 2. Logistics System Design and Technology (16:30 ~ 17:30)

"Universal Approximation Property of ODENet and ResNet: Toward Theoretical Foundations for Reliable AI" by Professor Matsui

Deep learning models are increasingly used in real-world decision-making systems, including logistics, transportation, and supply chain management. For these applications, establishing theoretical foundations for reliable and trustworthy AI is essential.  In deep neural networks, when the depth is overly increased, the gradient vanishing problem occurs, and accuracy tends to stagnate or degrade. ResNet is a deep neural network architecture proposed to overcome these issues. ResNet can be interpreted as a time discretization of an ordinary differential equation (ODE) system. We study a universal approximation property of the ODENet and ResNet. We restrict the function f to the following form: f(x, t) = α(t) ⊙ σ(β(t) x + γ(t)), which is the most common choice in practical machine learning systems. Here, α is a weight vector, β is a weight matrix, and γ is a bias vector. The operator ⊙ denotes the Hadamard product (element-wise product) of two vectors. The function σ is called an activation function. Examples of activation functions include the sigmoid function, the hyperbolic tangent function, and the rectified linear unit (ReLU) function. Based on [1], we show that such ODENet and ResNet with the restricted vector field can uniformly approximate ODENet with a general vector field. These findings provide mathematical justification for why simplified deep learning architectures can be effective in practice, contributing to the theoretical understanding of why deep learning works well.

"Smart Freight Transport Systems: Emerging Technologies, Current Progress, and Future Directions" by Professor Woojung Kim

Freight transportation systems are facing growing structural pressure, particularly from human-related problems in the trucking industry, including driver shortages, an aging workforce, rising driver compensation, etc. In this context, emerging technologies are drawing increasing attention not merely because technology is advancing, but because the freight sector urgently needs new solutions. This presentation examines how smart freight transport systems are being transformed by autonomous trucks, drones, sidewalk delivery robots, and humanoid robots, with attention to both their current state of development and future implications. In addition, it discusses related academic research on the emerging freight transport systems.