My dissertation focuses on the use of event log knowledge, i.e. process mining, to support the development of business process simulation models. Despite the fact that the Process Mining Manifesto highlights this topic as a key research challenge, prior research e orts tend to have a proof-of-concept nature. To this end, this dissertation contributes towards fundamentally bridging the gap between these domains by (i) providing the required conceptualization and (ii) developing a set of methods that extract knowledge from event logs to support speci c business process simulation modeling tasks.
Every organization is comprised of a set of business processes such as the production process, and the transportation process. When decision-makers analyze these processes and particular issues appear, several ideas for potential process changes are likely to be generated.
To evaluate the e ects of policy measures, managers can experiment with
the real-life process by, e.g., changing the sta ng policy and measure its
operational e ect. However, this is a high-risk approach as the implementation of
a measure does not guarantee the desired outcomes. This is a setting in which
business process simulation can be a valuable instrument. Business process
simulation (BPS) refers to the imitation of business process behavior through the
use of a simulation model. Using a BPS model, an organization can verify the
consequences of proposed process modi cations prior to implementation [
The use of BPS requires the construction of a simulation model, which, in
its turn, necessitates a profound insight in the business process. To this end,
extensive information needs to be collected. Typical information sources include
business documents, interviews with business experts and observations of the
process. Despite the valuable insights that can stem from these information
sources and their common use, their limitations should also be recognized.
Business documents can contain information deviating from real-life process behavior
[
The aforementioned limitations stress the need for information sources that are more readily available and less in uenced by human perception. In this respect, an important trend is that business processes are increasingly supported by process-aware information systems such as Enterprise Resource Planning systems. These systems record process execution information in an event log, of which the structure is illustrated in Table 1. Each row in the event log represents `something' that happens in the process. For instance: the rst row indicates that Sue started to register application 143 on April 3rd, 2018 at 08:52:41. She completes this registration at 09:04:04, as shown in the second row. case id timestamp activity transaction type resource ... ... ... ... ... ... ... 143 03/04/2018 08:52:41 Register application start Sue ... 143 03/04/2018 09:04:04 Register application complete Sue ... 144 03/04/2018 09:04:04 Register application start Sue ... 107 03/04/2018 09:06:22 Judge application start Mike ... ... ... ... ... ... ...
As an event log is automatically recorded and captures real-life process
behavior, it is well-suited as an additional information source for the construction
of a BPS model. My dissertation focuses on the use of event log knowledge, i.e.
process mining, to support simulation model construction. This is marked as one
of the key research challenges in the Process Mining Manifesto [
Literature on this topic includes [
From a thorough analysis of existing literature, it follows that, despite the intrinsic potential of process mining to support BPS model construction, limited insights exist in the systematic use of event log knowledge within this context. A signi cant research gap is present as literature on this matter tends to have a proof-of-concept nature. This implies that important simplifying assumptions are made and that various modeling tasks that need to be performed when building a real-life BPS model are not taken into consideration. Moreover, the required conceptual foundation to fundamentally integrate event log knowledge in BPS model development is not present, marking another research challenge.
Consistent with these research gaps, the two main objectives of this work are (i) providing the required conceptualization to fundamentally integrate event log knowledge in BPS model construction and (ii) developing a set of methods that extract knowledge from event logs to support speci c BPS modeling tasks. 2
Given the limited research insights in the systematic use of event log knowledge within this context, the rst part of the dissertation is situated at a conceptual level (Section 2.1). Taking this conceptualization as an input, the second part of the thesis develops methods to mine relevant insights related to speci c BPS modeling tasks from an even log (Section 2.2). 2.1
A rst stage of conceptualization relates to de ning the key steps of a simulation study. This enables positioning the use of event log knowledge in the wider picture of a BPS study. Simulation literature proposes a multitude of methods describing the key steps in a simulation study. Even though key components such as computer modeling are included in most of them, di erences can also be observed. However, existing methods tend to be de ned in isolation. Given this observation, a new method for conducting a simulation study is developed, based on a critical analysis of 14 existing methods. The developed method consists of 9 steps with continuous assessment as a central feedback mechanism.
After elaborating on the wider perspective of a BPS study, the second stage of conceptualization uses a generic representation of a BPS model to de ne a series of modeling tasks. For each of them, the potential of event log knowledge to support their speci cation is outlined. These insights are compared to the state of the art in process mining literature in general and research on the use of process mining for BPS purposes in particular. From this comparison, it follows that few process mining algorithms are directly applicable to support BPS model construction. This can be attributed to the di erences between the underlying paradigms in process mining and simulation. While many process mining techniques treat an event log as a series of independent cases, simulation mimics the behavior of an operational business process in which the interaction between cases that are simultaneously present in the process in uences process behavior. Hence, tailored methods need to be developed for a wide range of BPS modeling tasks, which demonstrates that additional research e orts are required. The structured set of research challenges provided in the dissertation provides clear directions for future research and, in this way, is a solid foundation for the future development of the domain. 2.2
Using the conceptualization outlined in Section 2.1 as an input, the dissertation also develops a series of new algorithms to retrieve event log insights to support speci c BPS modeling tasks. Three modeling tasks are selected based on criteria including the potential consequences of inaccurately performing it on the simulation study results and the absence of process mining algorithms which are directly applicable or easily adjustable. Based on this assessment, new methods are developed to support the speci cation of (i) the entity arrival rate, (ii) batch processing, and (iii) resource schedules. All algorithms are rigorously evaluated using arti cial data and, whenever relevant, real-life event logs.
Entity arrival rate An entity is a dynamic object (e.g., a patient, a parcel) that moves through the process. The entity arrival rate models the pace at which new entities arrive, which is often expressed using a parameterized probability distribution for the time between two consecutive arrivals. As it de nes the in ow at the beginning of the process, inaccurately modeling the arrival rate can have a major in uence on simulation outputs.
Process mining literature on this topic implicitly assumes that an entity's arrival time corresponds to its rst recorded timestamp. However, the developed taxonomy shows that this assumption is only appropriate when no queues are formed at the start activity. Given this observation, an Arrival Rate Parameter Retrieval Algorithm (ARPRA) is created, which is the rst algorithm that explicitly takes queue formation into account during arrival rate mining. Batch processing Batch processing refers to a resource's tendency to accumulate entities in order to process them simultaneously, concurrently, or sequentially. Regarding batch processing, two distinct topics are treated: batch identi cation and batch activation rules. For both topics, a new mining algorithm is developed, implemented and evaluated. Firstly, for batch identi cation, the Batch Organization of Work Identi cation algorithm (BOWI) is introduced, which is based on a distinction between simultaneous, sequential and concurrent batching. It is the rst algorithm that systematically identi es batches in an event log and calculates a set of batch processing metrics.
Secondly, related to batch activation rules, the Batch Activation Rule Identi cation algorithm (BARI) is created. A batch activation rule captures the circumstances under which a resource starts processing a batch, e.g. when a particular number of entities has been collected. Using explicit and implicit event log information, BARI transforms the problem into a classi cation problem, which is solved using decision tree analysis to obtain batch activation rules. Resource schedules A resource schedule expresses the availability of a resource for the process under consideration. When focusing on human resources, sta members might not work full-time or might be involved in multiple processes. Consequently, it can be di cult to deduce their availability for one speci c process from generic schedules provided by the HR-department. This is especially the case when work organization is less rigid, leaving signi cant freedom for resources to divide their attention between processes. As resource availability is critical information when building a simulation model, a Resource Schedule Identi cation Method (RSIM) is developed, which supports the speci cation of resource schedules. More speci cally, daily availability records are mined from the event log, which express a resource's allocation to a particular process on a particular day. It is the rst algorithm in its kind as the daily availability records take into account (i) the temporal dimension of availability, i.e. the time of day at which a resource is available, and (ii) intermediate availability interruptions. 3
My dissertation presents an important step towards the integration of process
mining in BPS model construction. Despite the fact that the Process Mining
Manifesto [