Automated generation of process simulation scenarios from declarative control-flow changes

Abstract

Business process simulation is an established approach to estimate the potential impact of hypothetical changes on a process, particularly in terms of time and cost-related performance measures. To overcome the complexity associated with manually specifying and fine-tuning simulation models, data-driven simulation (DDS) methods enable users to discover accurate business process simulation models from event logs. However, in the pursuit of accuracy, DDS methods often generate overly complex models. This complexity can hinder analysts when attempting to manually adjust these models to represent what-if scenarios, especially those involving control-flow changes such as activity re-sequencing. This article addresses this limitation by proposing an approach that allows users to specify control-flow changes to a business process simulation model declaratively, and to automate the generation of what-if scenarios. The proposed approach employs a generative deep learning model to produce traces resembling those in the original log while implementing the user-specified control-flow changes. Subsequently, the technique generates a stochastic process model, and uses it as a basis to construct a modified simulation model for what-if analysis. Experiments show that the simulation models generated through this approach replicate the accuracy of models manually created by directly altering the original process model.