Enhancing Middle-Mile Inventory Management Policies Through Simulation and Reinforcement Learning

Abstract

This thesis explores approaches for enhancing middle-mile inventory management within the global supply chain of a large footwear and apparel company, referred to as "Atlas". The first part discusses the design and implementation of a high-performance, heuristic system to determine stock transfer order (STO) decisions between Atlas’s distribution centers. This system employs a greedy algorithm to match supply to demand while respecting resource constraints. As Atlas’s newly procured third-party solution proved insufficient for testing due to slow performance, this work develops an emulator of the production system that achieves a 30x speedup and integrates with Atlas’s end-to-end supply chain simulation framework. This emulator enabled Atlas to efficiently test different configurations and decision making rules on historical and theoretical data, providing valuable insights prior to deploying the production system. The second part investigates the potential of reinforcement learning (RL) to augment or replace Atlas’s middle-mile decision making. A simplified supply chain environment is modeled as a Markov Decision Process, and an RL agent is trained and benchmarked against optimization-based and heuristic approaches. While the RL policy does not outperform these alternatives in the simplified environment, this work provides a foundation for Atlas to explore RL applications as they scale to more realistic supply chain environments.