Parametrization of a demand-driven operating model using reinforcement learning

Abstract

Nowadays, production and supply planning are more complex than ever before with low customer tolerance, complicated bill-of-materials, and high product variety. Most of the time, conventional approaches such as MRP and Lean approaches are not efficient. To overcome these issues, Ptak and Smith (2019) introduced Demand Driven Material Requirements Planning (DDMRP). This methodology relies on a Demand-Driven Operating Model (DDOM) which uses actual demand in a combination of strategic buffers to protect critical parts. For the past decade, research around DDMRP has been focused on proving and advancing the methodology in different industrial environments, while neglecting its parametrization. Indeed, the authors suggested general rules to set the DDOM’s key parameters, yet no learning approach has been developed to set them. This present paper is the first that proposes to use machine learning to parametrize a DDOM facing unknown demand, and particularly to adjust dynamically the order spike threshold and the order spike horizon. A reinforcement learning algorithm with three different reward functions is coupled to a DDMRP flowshop simulation model facing an atypical demand including spikes. Besides studying the learning ability of the algorithm, we evaluate the performance of the model which is compared to a DDOM without parameter adjustment. The analysis shows that it is possible to drive the order spike thresholds to increase the performance of the production system, regarding customer satisfaction and stock level optimization. The findings of this paper point out the possibility to drive DDOM parameters with an automatic method using reinforcement learning.