Unsupervised Learning and Simulation for Complexity Management in Business Operations

Abstract

A key resource in data analytics projects is the data to be analyzed. What can be done in the middle of a project if this data is not available as planned? This chapter explores a potential solution based on a use case from the manufacturing industry where the drivers of production complexity (and thus costs) were supposed to be determined by analyzing raw data from the shop floor, with the goal of subsequently recommending measures to simplify production processes and reduce complexity costs. The unavailability of the data-often a major threat to the anticipated outcome of a project-has been alleviated in this case study by means of simulation and unsupervised machine learning: a physical model of the shop floor produced the necessary lower-level records from high-level descriptions of the facility. Then, neural autoencoders learned a measure of complexity regardless of any humancontributed labels. In contrast to conventional complexity measures based on business analysis done by consultants, our data-driven methodology measures production complexity in a fully automated way while maintaining a high correlation to the human-devised measures.