Time: Wednesday, 04/Sept/2024: 11:00am - 12:00pm Session Chair: Philipp Willms |
Location: Wirtschaftswissenschaften 0544 Room Location at NavigaTUM |
A supervised machine learning approach for replenishment order decisions under transportation cost uncertainty
1Catholic University of Eichstätt-Ingolstadt, Germany; 2Université du Québec à Montréal, Canada
Companies worldwide seek a balance between risks and cost-efficiency in their supply chains. Due to the increase in extreme weather events, global inland waterway transport disruptions gained growing attention as shipment carriers enforce contractual surcharges to account for capacity losses. To improve efficiency and resilience, one open question is, therefore, when and how much to transport from which of the multiple suppliers considering their lead time differences and account for the transportation cost uncertainty driven by the enforced surcharges.
We formulate this problem as a stochastic Inventory Routing Problem with Direct Deliveries and introducing a new Cost Focused Machine Learning (CFML) framework. Compared to existing approaches, we perform a hyperparameter tuning where the costs of applying the resulting transportation decisions are optimized instead of the prediction score of the individual decisions of the machine learning model.
We evaluate our CFML in a case study, based on a chemical company at the border of the river Rhine. Relevant features for our CFML include the inventory position, historical water level, their trends, and predictions. While traditional learning approaches result in inefficient policies, we show that our CFML can reduce costs by 18% compared to classical machine learning frameworks and more than 20% compared to a standard (s,Q)-reorder policy representing industry standard.
Accelerated scheduling heuristics for reinforcement learning approaches applied to the Westenberger-Kallrath problem
1University of Kassel, Germany; 2Flensburg University of Applied Sciences, Germany
In the chemical process industry, production planning and scheduling (PPS) problems show a variety of characteristics which increases the complexity of corresponding modeling approaches and solution techniques. For recent years, the concept of reinforcement learning (RL) has been evaluated for its effectiveness to solve PPS problems for multiple case studies in discrete manufacturing. Our study delves into the utilization of RL to address the Westenberger-Kallrath (WK) problem which is a prominent benchmark for chemical production planning. The proposed solution approach decouples the lotsizing part from the scheduling activity via a custom MRP heuristic which creates chains of batches linked to customer orders. We train an RL agent to find the best sequence of scheduling those chains with the objective to minimize overall makespan. Building on our previous work, which used a custom forward scheduling heuristic with an internal inventory projection, we propose a new scheduling logic based on minimum waiting times between the multi-stage processes and their assigned resources. We provide a comprehensive performance comparison between two implementation approaches, evaluating criteria such as makespan, inventory and resource utilization. Apart from the effectiveness measures, a CPU time comparison proves the efficiency of our new method. Using the new logic in RL agent training, we detect further modeling challenges regarding reward function design, which provide future research perspectives on RL for PPS problems in chemical process industry.