Development of Agent-Based Simulation Models for Software Evolution
von Daniel Honsel
Datum der mündl. Prüfung:2019-12-12
Erschienen:2020-02-11
Betreuer:Prof. Dr. Stephan Waack
Gutachter:Prof. Dr. Stephan Waack
Gutachter:Prof. Dr. Jens Grabowski
Dateien
Name:PhDThesis_Daniel-Honsel_rev.pdf
Size:6.69Mb
Format:PDF
Zusammenfassung
Englisch
Software has become a part of everyday life for us. This is also associated with increasing requirements for adaptability to rapidly changing environments. This evolutionary process of software is being studied by a software engineering related research area, called software evolution. The changes to a software over time are caused by the work of the developers. For this reason, the developer contribution behavior is central for analyzing the evolution of a software project. For the analysis of real projects, a variety of open source projects is freely available. For the simulation of software projects, we use multiagent systems because this allows us to describe the behavior of the developers in detail. In this thesis, we develop several successive agent-based models that cover different aspects of software evolution. We start with a simple model with no dependencies between the agents that can simulative reproduce the growth of a real project solely based on the developer’s contribution behavior. Subsequent models were supplemented by additional agents, such as different developer types and bugs, as well as dependencies between the agents. These advanced models can then be used to answer different questions concerning software evolution simulative. For example, one of these questions answers what happens to the software in terms of quality when the core developer suddenly leaves the project. The most complex model can simulate software refactorings based on graph transformations. The simulation output is a graph which represents the software. The representative of the software is the change coupling graph, which is extended for the simulation of refactorings. In this thesis, this graph is denoted as \emph{software graph}. To parameterize these models, we have developed different mining tools. These tools allow us to instantiate a model with project-specific parameters, to instantiate a model with a snapshot of the analyzed project, or to parameterize the transformation rules required to model refactorings. The results of three case studies show, among other things, that our approach to use agent-based simulation is an appropriate choice for predicting the evolution of software projects. Furthermore, we were able to show that different growth trends of the real software can be reproduced simulative with a suitable selection of simulation parameters. The best results for the simulated software graph are obtained when we start the simulation after an initial phase with a snapshot of real software. Regarding refactorings, we were able to show that the model based on graph transformations is applicable and that it can slightly improve the simulated growth.
Keywords: Agent-Based simulation; Simulation; Software Evolution; Mining software repositories