Bayesian Contextual Bandits Optimize Warehouse Sorter Diversion in Real-Time

This study investigates real-time optimization for warehouse sorter diversion control, a critical aspect of operational efficiency in large e-commerce warehouses. Traditional systems often rely on static cost functions that struggle to adapt to dynamic conditions like volume changes, congestion, and equipment status. To address this, researchers compared three machine learning frameworks: Linear Regression with Gradient Descent Optimization (LR+GDO), XGBoost with Bayesian Optimization (XGB+BO), and Bayesian Contextual Bandits (BCB). Using a high-fidelity emulator for training and evaluation, the study found that while tree-based models offered slightly better predictive power, the BCB framework delivered the highest overall performance, achieving a 2.03% reward uplift over the heuristic baseline. BCB demonstrated several key advantages, including a decisive time-optimal policy, continuous online learning capabilities, a strategic balance between exploration and exploitation, and significantly lower inference latency. These findings strongly support BCB's potential for real-time control optimization in complex warehouse environments.