Quentin NIVON

Abstract:
Modelling and designing business processes have become crucial activities for companies in the last decades.
Consequently, multiple workflow modelling notations emerged.
Among them, the Business Process Modelling Notation (BPMN) is now considered as the de facto standard for process modelling.
The BPMN notation requires a certain level of expertise to allow one to write correct and well-structured processes compliant with some expected requirements.
The BPMN modelling phase can thus become tedious, and even error-prone if carried out by non-experts.
The first part of this thesis consists in providing a solution to help users modelling BPMN processes.
To achieve this goal, the proposed approach takes as input the requirements of the user in a textual format, in which the tasks and their ordering constraints are informally described.
It then makes use of Large Language Models (LLMs) to translate these constraints into a machine-readable format.
From this internal format, the approach generates and returns a BPMN process satisfying these constraints.
As a side contribution, this thesis also presents techniques to verify that a process does not deviate from its expected behaviour.
Such verifications are usually performed using classical model checking techniques.
However, we thought that a user not familiar with the BPMN notation would struggle using such techniques, and more precisely, writing the expected behaviour of the process in the form of temporal logic properties.
Thus, the core of this side contribution consists in facilitating the writing of such properties by allowing the user to generate them directly from their textual description.

Recent studies suggested that, once built, business processes are subject to changes throughout their lifetime.
In companies, such changes may often lead to non-optimal processes, responsible of issues such that the increase of the execution time or the costs related to them.
To be able to optimise processes, there is a need to have at hand an explicit model of their behavior and quantitative features.
Thus, beside the processes themselves, usually modelled using workflow-based notations, one must provide, among others, an explicit description of the durations and the resources required by the tasks composing these processes.
The second part of this thesis consists in providing a solution based on refactoring techniques, whose goal is to change the structure of the process in order to optimise one or several criteria of interest, such as process execution time, resource usage, or total costs.
To do so, the proposed approaches consist of various ingredients.
For instance, we propose refactoring patterns, useful for moving the tasks of the process from one place to another while preserving some of its semantics.
We also make use of simulation techniques to compute metrics of interest from the execution of (one or several instances of) the process.
Similarly, we utilise several exploration algorithms in order to navigate through the space of solutions.
In the end, the multiple presented approaches all return an optimised version of the original process given as input.