Artificial intelligence in online scheduling of dynamically interconnected assembly systems

• Künstliche Intelligenz in der Online-Ablaufplanung von frei verketteten Montagesystemen

Göppert, Amon Mirko Robin; Schmitt, Robert H. (Thesis advisor); Gries, Thomas (Thesis advisor)

Aachen : RWTH Aachen University (2022)
Dissertation / PhD Thesis

Dissertation, Rheinisch-Westfälische Technische Hochschule Aachen, 2022

Abstract

Increasing product individuality, globalization of markets, shorter product life cycles, and disrupted supply chains are global trends that demand adaptable industrial assembly systems. Dynamically interconnected assembly systems (DIAS) fulfill this demand by enabling individual job routes and by detaching from cycle time and linear transfer. A central component of DIAS is the control system that requires new online scheduling algorithms for efficient operation. In the computer science domain, the artificial intelligence (AI) algorithm AlphaZero showed groundbreaking results in playing strategy board games. Since online scheduling in DIAS is comparable to decision-making in board games, AlphaZero is a transferable AI solution that could significantly increase scheduling performance. This dissertation presents an AlphaZero online scheduling agent to investigate the performance potential. The agent uses AlphaZero's Monte-Carlo tree search and deep artificial neural networks trained by reinforcement learning. Various auxiliary software components and models were created for enabling AlphaZero online scheduling. An automated scenario analysis workflow, that incorporates a simulation environment for modeling DIAS, was created for training the agent. For evaluating DIAS states during training, an additional simulation software module was specifically developed. The training included hyperparameter optimization and was conducted in multiple cycles with large data sets based on two industrial use cases. It resulted in significant improvements of the Monte-Carlo tree search and the neural network. A modular online scheduling architecture was elaborated for the adaptive communication of the agent with simulation models and control systems via standardized interfaces and data models. This architecture facilitates the seamless deployment and updating of the AlphaZero agent into an operating production system. These newly created auxiliary components and models were successfully verified and validated with established techniques. The validation of the AlphaZero agent was performed with sensitivity analyses showing plausible algorithm behavior. Comparing the AlphaZero agent with heuristic rule-based reference scheduling agents and a mathematical mixed-integer linear programming model revealed heterogeneous performance improvements, depending on the DIAS scenario characteristics. On average, the AlphaZero agent could improve the scheduling performance with AI methods in DIAS.

Institutions

• Laboratory for Machine Tools and Production Engineering (WZL) of RWTH Aachen University [417200]
• Chair of Production Metrology and Quality Management [417510]