Explainable AI for reinforcement learning based dynamic scheduling solutions in semiconductor manufacturing

Abstract The scheduling of complex job shop manufacturing environments is a difficult and NP-hard optimization problem. It is tackled increasingly with AI-based methods, often by reinforcement learning approaches. So far, this is mostly an academic endeavor as the trust in such systems is limited. The neural networks representing the policy are often seen as a black box by domain experts. By proposing a holistic approach to make such policies more interpretable and explainable, we hope to bridge the gap between academia and industry. With increasing trust in the solution, it becomes closer to being deployed in a real manufacturing environment. We propose a combination of different statistical and well-established ML-based methodologies to analyze single state-action explanations as well as the overall strategy and apply them to pretrained agents on open-source benchmark simulation models. This allows us to deliver humanly understandable explanations for the policies of the analyzed agents.