Reinforcement Learning for Optimizing Retail Inventory Management

Modern retail supply chains face complexity in operations with thousands of Stock Keeping Unit (SKU), volatile demand and multi echelon distribution networks. In this setup, conventional rule based policies perform poorly and resulting in costly stock outs or overstocking. This study develops a scalable, competition aware, cooperation enabled, multi agent reinforcement learning system for real world inventory management.This study integrates Graph Neural Networks (GNNs) with attention based message passing to enable inter agent communication and coordination across multiple echelons. Using parameter sharing, policy learning is generalized among similar SKUs to stabilize training. It supports both actor-critic and value-based learning paradigms. Capacity constraints are added directly into the reward structure to penalize overflow and ensure balanced utilization across different echelons. A structured coordination mechanism synchronizes replenishment decisions through relational representations of demand, supply and lead time dynamics. The approach was evaluated based on the multi-agent benchmark for inventory management (MABIM) (X. Yang et al., 2023) and a simulated national retail chain. The results indicated improvements in profit and model convergence. These findings advance reinforcement learning based supply chain optimization in large scale dynamic environments.

Author

• Devireddy, Anish Reddy

Language

• en

Keyword

• MAPPO
• Supply Chain Management
• QTRAN
• Graph Neural Networks
• Inventory Management
• Reinforcement Learning

Date created

• 2025

Type of Work

• Dissertation

License

• All rights reserved

Rights statement

• In Copyright

GW Unit

• Data Science

Degree

• D.Engr.

Advisor

• Islam, Muhammad M

Committee Member(s)

• Yu, Jeffrey
• Etemadi, Amir

Persistent URL

•  https://scholarspace.library.gwu.edu/etd/sq87bv750