Integrating Probabilistic Suffix Prediction into a Business Process Simulation Model
Provided and advised by: Henryk Mustroph
Business Process Simulation (BPS) aims to construct, based on historical data captured in information systems, a digital twin of a real-world process. This digital twin incorporates activity control flows, processing times and durations, correct resource assignments, and potentially all additional data relevant to the business scenario. One of the goals of BPS is to create simulations as similar as real world executions could happen to e.g., evaluate multiple optimization techniques and analyze their effects on both the process and the underlying business case, prior to potential adoption in the real world. This makes BPS a powerful and widely applicable approach for organizations, largely independent of a specific domain or use case.
Traditional BPS is typically based on discrete-event simulation, in which event sequences are replayed on a discovered process model. A BPS model usually consists of (i) a control-flow specification (e.g., BPMN or a Petri net), (ii) arrival processes for case generation, (iii) stochastic activity durations and routing rules, and (iv) resource models with scheduling and allocation policies [1]. A process engine repeatedly spawns process instances, replays activities on the process model, advances simulation time, and resolves resource contention, enabling analysts to estimate performance measures such as cycle time, throughput, waiting times, costs, and resource utilization under different scenarios. Classical BPS models can be improved by integrating prediction models [2].
Probabilistic Suffix Prediction (PSP) [3] originates from predictive process monitoring [4] and focuses on forecasting the future events of an ongoing process instance. Given a prefix, the sequence of events observed so far for a running case, a PSP model predicts a probability distribution over possible suffixes, that is, over the remaining sequence of events until completion. Unlike deterministic approaches that output a single most likely continuation, PSP models generate multiple plausible future traces, typically by sampling from an autoregressive sequence model (e.g., an encoder–decoder LSTM or a Transformer). These models often predict both future activities and temporal information, such as the time until the next event, and explicitly model uncertainty using techniques such as Monte Carlo dropout.
When integrated into a BPS model, PSP can serve as a data-driven generator of future process behavior. Instead of relying on static routing probabilities or predefined loop counts, the simulator can use the PSP model to obtain conditional forecasts of how cases are likely to evolve given their current state. The thesis should find out state of the art BPS model solutions and should explore possible options how to use PSP for BPS. Then a simulation model based on a PSP model should be implemented and evaluated.
Recommended Readings:
Simulation:
[1] Rozinat, A., Mans, R. S., Song, M., & van der Aalst, W. M. (2009). Discovering simulation models. Information systems, 34(3), 305-327.
[2] Meneghello, F., Di Francescomarino, C., Ghidini, C., & Ronzani, M. (2025). Runtime integration of machine learning and simulation for business processes: Time and decision mining predictions. Information Systems, 128, 102472.
Predictive Process Monitoring & Probabilistic Suffix Prediction:
[3] Mustroph, H., Kunkler, M., & Rinderle-Ma, S. (2025). An Uncertainty-Aware ED- LSTM for Probabilistic Suffix Prediction. arXiv preprint arXiv:2505.21339.
[4] Di Francescomarino, C., & Ghidini, C. (2022). Predictive process monitoring. In Process mining handbook (pp. 320-346). Cham: Springer International Publishing.
Topics: Artificial Intelligence, Process Mining, Business Process Prediction & Simulation.
The application must contain:
- Thesis topic you are interested in.
- Current Transcript of Records.
- Application form
- CV
Please send the application to master.i17(at)in.tum.de AND (in cc) henryk.mustroph(at)tum.de.