A data-driven approach to solving the container relocation problem with uncertainties

Abstract

Container relocations are inevitable and reduce terminal efficiency, making their optimization a critical research focus. Scholars have extensively studied the Container Relocation Problem (CRP) with the goal of reducing relocations. However, most existing research assumes prior knowledge of retrieval sequences, which often does not reflect real-world conditions. Consequently, addressing the CRP with uncertain retrieval sequences necessitates overcoming challenges related to both uncertainty and complexity. To manage this uncertainty, we propose a novel concept: the Retrieval Probability Matrix (RPM). A data-driven model is developed to predict the RPM, utilizing real terminal operational records. Building on this foundation, this study extends the online CRP to the Probabilistic Container Relocation Problem (PCRP) and presents a decision tree-based algorithm for obtaining optimal solutions. To address the inherent complexity of the PCRP, we propose an Adapted Monte Carlo Tree Search algorithm. It minimizes the expected number of container relocations by integrating a novel heuristic: Local Safety and Global Flexibility. The proposed algorithms are validated through experiments, demonstrating their effectiveness and feasibility. Furthermore, sensitivity analysis is conducted to evaluate the impact of RPM prediction accuracy on algorithm performance.