A systematic review of machine learning approaches in inventory control optimization

This systematic review investigates the applications of machine learning (ML) in inventory control, analyzing 122 articles to provide a comprehensive overview of the state of the art and identify future research directions. The study proposes a typology to classify the integration of ML into the inventory optimization framework, distinguishing three primary approaches: (1) separate estimation and optimization, where ML is applied to demand forecasting before optimization, (2) static ML-integrated optimization, where ML is directly embedded into optimization models, and (3) dynamic ML-integrated optimization, where reinforcement learning (RL) is employed to derive optimal inventory policies. The findings highlight that while RL applications are gaining prominence, significant research gaps remain, particularly in scaling algorithms to real-world problems, handling large action spaces, and developing RL algorithms that are tailored to inventory control. The review also assesses the operational dynamics of inventory systems addressed in the literature, such as single/multi-item models, lead time assumptions, and echelon structures. Underexplored areas include stochastic lead times, complementary items, quantity discounts, product obsolescence, and multi-echelon networks. The study concludes by outlining key research gaps and offering directions for future research to advance the integration of ML in inventory control.