GOALI: Machine Learning Approaches for Supply Chain Decision-Making

Lehigh University, Bethlehem PA

Investigators

Abstract

Supply chain refers to the system that moves goods from where they are produced to where they are consumed. This system includes manufacturing, assembly, warehousing, transportation, and retailing processes. This project will study how to improve efficiency and robustness of the supply chain by using novel machine learning techniques to control the supply chain automatically. The project specifically focuses on the development of decision-making strategies to deal with uncertainties and correct disruptions in the supply chain system. In collaboration with an industrial partner, Siemens, the project will focus on a real supply chain application related to the production and distribution of radiopharmaceuticals (an important component in health diagnostics). The advancements in supply chain and machine learning resulting from this project have potential to benefit a wide range of industries. This project studies a new approach for using machine learning (ML) as a tool for optimizing, analyzing and controlling supply chains. Current approach to supply chain operations makes assumptions about statistical distributions of uncertain factors in supply chain and uses predictive techniques or forecasts to estimate the few parameters required to characterize these distributions. These estimated distributions are then used in the analysis and control of the supply chain. This project has a novel approach of combining the data-analysis and supply-chain optimization stages into a single ML algorithm. The project focuses on two main classes of supply chain problems: production and distribution of radiopharmaceuticals (a core problem for the industrial partner, Siemens), and early warnings and corrective actions for stochastic supply chains. In addition, the project will make methodological contributions to the theory and implementation of ML algorithms, including new loss functions, techniques for using deep learning as a preprocessing step for reinforcement learning, adaptive strategies to handle non-stationary data, and improved initialization methods for ML. The tools and concepts to be develop will be generalizable more broadly, both within and beyond supply chain.