Application of Online Learning for the Dynamic Configuration of Kanban Systems

Kanban systems constitute a widely used pull control for inventory management in production systems. As a result of an increasingly volatile and individualized customer demand, Kanban systems have to be reconfigured dynamically to achieve minimal inventory levels while maintaining a stable production. This paper investigates the application of an incremental online learning platform called XELOPRO to optimize inventory levels using the current state of the production system, while including contextual information, e.g., time-related information. As the platform uses an incremental support vector machine to update its models during runtime without the need to store and reevaluate large amounts of historical data, it constitutes a suitable tool for a decentralized inventory management. Results show a good performance with drastic decreases in inventory levels compared to static configurations and a higher reliability compared to a dynamic application of standard Kanban rules.