Smart Inventory Optimization with Deep Reinforcement Learning: Challenges and Real-World Applications

Inventory management is a key part of running an efficient supply chain, but it's also tricky, especially when things are uncertain. Traditional methods work well when everything is stable and deterministic, but they often fall short in fast-changing situations where demand shifts unexpectedly, or information is incomplete shifting the demand patterns. That's where Deep Reinforcement Learning (DRL) comes in—a modern approach that helps systems learn how to make better decisions over time. This invited talk focuses on how DRL can be used to improve inventory control. The goal is to make smart decisions step-by-step to reduce costs like holding costs and maintaining service level. By modeling the system as Markov Decision Process techniques such as Deep Q-Networks (DQN) and Proximal Policy Optimization (PPO) can be deployed to learn robust policies that generalize across stochastic demands and lead time scenarios. Drawing from empirical research and experimental studies, we'll look at what makes DRL promising and the challenges it faces. These include issues related to reward design, sample inefficiency, exploration strategies, and the interpretability of learned policies.