Background: Distributed manufacturing (DM) is a decentralized production system that spreads resources and production processes across multiple locations and suppliers. This leads to efficient resource utilization and cost-effective production. However, efficient management of capacity sharing and inventory transshipment within DM facilities poses significant challenges due to the complex interdependencies among entities and dynamic operational conditions.

Research Objectives: In this project, we will propose a novel approach based on lookahead reinforcement learning (LRL) to tackle the capacity sharing and inventory transshipment problem in DM. The proposed framework will leverage RL agents to learn optimal decision-making policies for capacity allocation and inventory transshipment taking into consideration demand fluctuations, production constraints, and transportation costs. We will formulate the problem as a Markov Decision Process (MDP) and design a reward function that incentivizes efficient resource utilization and inventory management. Through extensive simulations and a case study from computer server industry, we demonstrate the effectiveness of the proposed LRL approach in improving system performance metrics such as throughput, resource utilization, and inventory turnover.

Given is a set of independent manufacturing sites (A, B, and C) of a computer server manufacturing company with different geographical locations (see Figure 1). Customer orders are allocated to the three sites based on cost and other planning factors. Each site serves customers in its respective geographical areas. The manufacturing environment is based on build-to-plan, make-to-order production strategies. Parts are received from suppliers and tested according to a predefined fabrication plan which is based on customer order forecasts. The tested parts are either used in customer orders or shipped to sister sites. Shipping the parts to sister sites can be based on a predefined shipping plan or an immediate shipping request sent by the sister plant. The immediate shipping requests arise from inherent uncertainties in demand forecasts, quality issues, and machine failures. Orders can also be offloaded from one site to another to avoid supply and demand shortages. Customer order is characterized by order quantity, type, configuration, destination, and ship date. A manufacturing site may not be able fulfill a customer order due to one or more of the following factors: 1) shortage of raw materials, 2) capacity limitations, 3) policy factors. For raw material shortage, the site can obtain parts from a sister site or an external supplier. For the other four factors (capacity, time, customer requirements, and policy factors), the site can offload the order to a sister site that can fulfill the order.

Research Plan: The REU student will develop a reinforcement learning environment that represents the manufacturing company with the three sites based on the details given above. The student will then conduct experimental analysis and prepare a paper draft.

REU Student Outcomes: This project fits a student with coding skills (e.g., python) and knowledge of production scheduling (e.g., Industrial Engineering). REU student will gain hands-on experience in reinforcement learning and system simulation. Student will be supervised by Dr. Faisal Aqlan (Industrial Engineering) and participate in writing conference and journal papers.

 References

Aqlan, F. and Lam, S.S., 2016. Supply chain optimization under risk and uncertainty: A case study for high-end server manufacturing. Computers & Industrial Engineering, 93, pp.78-87.

Nezamoddini, N., Gholami, A. and Aqlan, F., 2020. A risk-based optimization framework for integrated supply chains using genetic algorithm and artificial neural networks. International Journal of Production Economics, 225, p.107569.

Aqlan, F., Lam, S.S. and Ramakrishnan, S., 2016. A framework for inventory transhipment in integrated supply chains. International Journal of Supply Chain and Inventory Management, 1(2), pp.118-132.