Machine Learning Guides the Solution of Blocks Relocation Problem in Container Terminals

Abstract

The blocks relocation problem (BRP) is a well known and important combinatorial optimization problem, in which the initial storage state and retrieval priority of containers are known, and the containers should be picked up in the retrieval order with the goal of minimizing container relocations. This paper studied how machine-learning techniques guide the solution of this Non-deterministic Polynomial problem(NP-hard problem) NP-hard problem. Through our self-developed data generator, we generated initial state stacking matrices and extracted 22 influencing factors for container relocation operations. The supervised learning method and attribution technique were used to verify the relationship between significant container relocated influence features and the number of container relocations using the unrestricted BRP and restricted BRP models. We characterize the potential patterns in the data based on 22 container relocated influence features using four supervised learning models: random forest (RF), extra trees (ET), support vector machine (SVM), and logistic regression (LR). The experimental results demonstrate that RF has a classification accuracy rate of up to 94% on the restricted BRP model. The attribution technique identifies the most sensitive features to the number of container relocations. We organically integrate machine learning into the BRP problem and propose an interactive iterative framework that may provide a novel method for studying the BRP problem.