Iterative Improvement of Simulation Models Focussing on Time and Resources


MSc Thesis Project

Title: Iterative Improvement of Simulation Models Focussing on Time and Resources

Author: Daniel Tacke genannt Unterberg

Supervisor: Prof. Dr. Wil van der Aalst

1st examiner: Prof. Dr. Wil van der Aalst

2nd examiner: Prof. Dr. Bernhard Rumpe


One of the main goals of process mining is improving the conformance and/or performance of real-life processes. To convince process stakeholders of the quality of proposed improve-ments, data-based predictions are an important tool. These predictions can be generated via business process simulation (BPS) which is instantiated with historical data. That is, a simulation model is automatically derived from the given data. It can then be manually reconfigured with the proposed changes to simulate what would happen if the proposed changes were made. The reliability of these predictions is obviously directly linked to the model quality. Recently, commercial business process management (BPM) software by market leader Celonis, as well as Signavio have started o˙ering these capabilities which highlights their relevance and market demand. For maximal convincing power, human-interpretable models are an important avenue because they can be verified by domain experts. They can mimic how data is generated and not just what data. To make high quality predictions, an accurate and detailed model is necessary. It e˙ectively possesses a high information content and high “resolution” in terms of the main drivers of process dynamics. However, a lot of recorded data is insuÿcient for the purpose of automatically deriving such models. Manual configuration by domain experts is not always feasible. Making use of the available data and still providing a detailed model is our main goal. To achieve it, we propose an iterative fitting approach that bridges this gap via variable parameterizations. Essentially, we try multiple guesses of model parameters that relate to what we view as main drivers of business process dynamics: time and resources. More concretely, activity execution lifecycles and resource contention. We also develop a scor-ing metric and evaluate the approach on synthetic and real-life data. In the end, we also critically discuss our results in light of providing a future avenue for this direction of research.