EURO 2024 Copenhagen
Abstract Submission

EURO-Online login

3549. Multi Agent Reinforcement Learning and Graph Neural Networks for Inventory Management

Invited abstract in session TB-45: Artificial Intelligence and Machine Learning for Decision Support, stream Decision Support Systems.

Tuesday, 10:30-12:00
Room: 30 (building: 324)

Authors (first author is the speaker)

1. Niki Kotecha
Imperial College London
2. Antonio del Rio Chanona
Imperial College London

Abstract

Current supply chains operate under uncertain environments causing risks in disruptions and sub-optimal performance due to operational failure or lack of coordination. The inventory control problem, a sequential decision-making problem, is challenged by stochastic and volatile factors such as lead times and seasonal demand patterns, often resulting in sub-optimal performance. Reinforcement learning (RL) remains a promising alternative to enhance decision-making in supply chains. However, as the supply chain grows in size, the complexity of decision-making can grow significantly which may hinder the performance of traditional RL algorithms as they struggle to scale efficiently. The extension of single-agent RL to multi-agent RL allows for scalability as well as a decentralised decision-making framework of individual entities. Our methodology leverages on the inherent graph structure of a supply chain, developing a multi agent RL framework with Graph Neural Networks for a multi-echelon multi-product inventory management system. We show the benefits of a collaborative approach by testing the policies on a series of disruptions. Additionally, the framework moves computational costs from online to offline, ensuring faster decision making than most efficient optimisation methods. As a result, the methodology proposed shows promising scalability with number of agents for a decentralised and online decision-making framework whilst still ensuring collaboration between entities.

Keywords

Status: accepted

Back to the list of papers