Optimizing Decision Making in Large-Scale Constrained Scheduling Problems using Simulation

Created by W.Langdon from gp-bibliography.bib Revision:1.9039

@PhdThesis{Benjamin_Kovacs:thesis,

• author = "Benjamin Kovacs",
• title = "Optimizing Decision Making in Large-Scale Constrained Scheduling Problems using Simulation",
• school = "Faculty of Technical Sciences, University of Klagenfurt",
• year = "2023",
• address = "Austria",
• month = sep,
• keywords = "genetic algorithms, genetic programming, production systems, artificial intelligence, AI, machine learning, reinforcement learning, simulation, semiconductor manufacturing, constraint programming, greedy algorithm",
• URL = "https://permalink.obvsg.at/UKL/AC17193213",
• URL = "https://netlibrary.aau.at/obvuklhs/content/titleinfo/9992946",
• URL = "https://netlibrary.aau.at/obvuklhs/download/pdf/9992946",
• size = "120 pages",
• abstract = "Production processes have been undergoing fundamental transformations in the past decade. On the hardware side, there is an enormous increase in automation. Robots gradually replace the human workforce in many industries, and new factories or production lines are usually built in a fully automated manner. Thus, the number of human operators is minimal; most staff focuses on equipment maintenance and process engineering. The transformation implies further challenges on the software side, with the need for advanced scheduling techniques capable of efficiently handling possible problems in the production process. However, automation also opens up further opportunities to maximize efficiency in the production process. Human factors, such as overriding algorithmic decisions based on experience, delays due to inattention, or staff unavailability, can mostly be factored out.

  This thesis develops and analyzes algorithms to optimize the operations of real-world production plants of various scales. First, we present the use of mathematical programming to solve mid-scale scheduling problems in the machine industry. A scalability analysis is performed on the approach with larger instance sizes, comparing performance to greedy and constraint programming-based algorithms and demonstrating limitations of the introduced methods.

  Next, we focus on the semiconductor industry, where both the scale of the problem and the automation level of the factories are significantly higher. In our collaborative project with Infineon Technologies Austria AG, we develop a performant Python-based simulator, which is the first open-source tool for the simulation of large-scale problem instances with industry-specific characteristics. The tool is evaluated and validated on a large factory model. Its performance allows the parallelized training of machine learning algorithms with high sample complexity. We use industry-standard dispatching heuristics to control the virtual factory and assess key performance indicators such as throughput, tardiness, and machine utilization. Additionally, we developed several visualization tools and user interfaces to set up and monitor the simulation and understand the behavior of the deployed dispatching agents. The simulator is ready for customization and augmentation using the built-in plugin infrastructure, which also allows for maximizing performance by limiting the data collection to the specific use case.

  Finally, we discuss the development of improved dispatching and planning tools for the simulated environment. Our goal is to design algorithms ready for real-world deployment and applicable in an industrial context. Therefore, we put emphasis on the explainability and verifiability of methods without endangering the production process. First, to reduce rework and waste incurred by scrapping faulty products, we introduce a deep learning-based predictive dispatching strategy to reduce the violations of time constraints among coupled operations. Second, we develop dispatching strategies customized for all machine groups using genetic programming. We show how the introduced techniques improve the factory performance indicators. Lastly, we discuss the ways to seamlessly integrate the methods with the dispatching systems of real-world fabs without endangering their continuous operation.",

• notes = "In english

  urn:nbn:at:at-ubk:1-52944

  Supervisors: Martin Gebser and Konstantin Schekotihin",

}

Genetic Programming entries for Benjamin Kovacs

Citations