Container Retrieval Sequence Uncertainty Quantification Using Calibrated Machine Learning Models^*

The Container Relocation Problem (CRP) has garnered significant attention from scholars. However, most studies addressing the CRP assume a deterministic container retrieval sequence, which deviates from real-world terminal operations and limits the practical applicability of such research. To address the uncertainty in container retrieval sequences, this study proposes a novel uncertainty quantification framework that integrates multi-class classification models with probabilistic calibration techniques. Experiments are conducted to compare two multi-class classification models and several probabilistic calibration strategies. In this study, we also propose a novel calibration strategy. The results demonstrate that the output of models calibrated with Isotonic Regression closely approximates true probabilities. Meanwhile, our newly proposed calibration strategy demonstrates potential as a complementary approach to Platt Scaling.