Dual-View Stack State Learning Network for Attribute-Based Container Location Assignment

Rong Yang, Zihan Feng, Linfa Pu, Yajun Yang, Chenya Liu, Qinghua Hu

Abstract: With the increasing demand for global trade transportation, the shipping container market plays a pivotal role in maritime terminals. Container location assignment is one of the most critical issues, as it significantly impacts both transportation efficiency and production safety. Recent efforts by industry and academia introduce various methods to address this issue such as Expert Assignment and Dynamic Assignment. However, two main limitations remain: First, current practices rely on rule-based methods supplemented by manual intervention, leading to inefficiencies and high costs. Second, a significant gap exists in understanding the implicit relationships between containers within a stack, limiting the robustness in real-world scenarios. To address these limitations, we propose the Dual-view Stack State Learning Network for Attribute-based Container Location Assignment, named DSLA. DSLA aims to capture and exploit the implicit relationships between containers within stacks. On the one hand, a sequential-view stack state learning module is introduced to extract the transition patterns between containers within a stack. On the other hand, a pairwise-view stack state learning module is designed to extract feature commonalities between any two containers. By integrating these dual-view representations, DSLA can effectively evaluate candidate containers for each available stack. Additionally, DSLA employs adversarial training to generate high-quality negative samples, further enhancing the robustness in practical scenarios. Extensive experiments on two real-world datasets validate that DSLA can significantly improve container assignment performance.