What Makes Free/Libre Open Source Software (FLOSS) Projects Successful?: An Agent-Based Model of FLOSS Projects

Background

There have been a limited number of attempts to simulate various parts of the open source development process (Dalle & David, 2004). For example, Dalle and David (2004) use agent-based modeling to create SimCode, a simulator that attempts to model where developers will focus their contributions within a single project. However, in order to predict the success/failure of a single FLOSS project, other existing FLOSS projects, which are vying for a limited pool of developers and users, may need to be considered. This is especially true when multiple FLOSS projects are competing for a limited market share (e.g., two driver projects for the same piece of hardware or rival desktop environments such as GNOME and the KDE). Wagstrom, Herbsleb, and Carley (2005) created OSSim, an agent-based model containing users, developers, and projects that is driven by social networks. While this model allows for multiple competing projects, the published experiments include a maximum of only four projects (Wagstrom et al., 2005). Preliminary work on modeling competition among projects is currently being explored by Katsamakas and Georgantzas (2007) using a system dynamics framework. By using a population of projects, it is possible to consider factors between the projects, e.g., the relative popularity of a project with respect to other projects as a factor that attracts developers and users to a particular project. Therefore, our model pioneers new territory by attempting to simulate across a large landscape of FLOSS with agent-based modeling.

Gao, Madey, and Freeh (2005) approach modeling and simulating the FLOSS community via social network theory, focusing on the relationships between FLOSS developers. While they also use empirical data from the online FLOSS repository SourceForge to calibrate their model, they are mostly interested in replicating the network structure and use network metrics for validation purposes (e.g. network diameter and degree). Our model attempts to replicate other emergent properties of FLOSS development without including the complexities of social networking. However, both teams consider some similar indicators, such as the number of developers working on a project, when evaluating the performance of the models.

In addition, there have been attempts to identify factors that influence FLOSS. These have ranged from pure speculation (Raymond’s (2000) gift giving culture postulates) to surveys of developers (Rossi, 2004) to case studies using data mined from SourceForge (Michlmayr, 2005). Wang (2007) demonstrates specific factors can be used for predicting the success of FLOSS projects via K-Means clustering. However, this form of machine learning offers no insight into the actual underlying process that causes projects to succeed. Therefore, the research presented here approaches simulating the FLOSS development process using agent-based modeling instead of machine learning.

To encourage more simulation of the FLOSS development process, Antoniades, Samoladas, Stamelos, Angelis, and Bleris (2005) created a general framework for FLOSS models. The model presented here follows some of the recommendations and best practices suggested in this framework. In addition, Antoniades et al. (2005) developed an initial dynamical simulation model of FLOSS. Although the model presented here is agent-based, many of the techniques, including calibration, validation, and addressing the stochastic nature of the modeling process, are similar between the two models. One difference is the empirical data used for validation: Antoniades et al.’s (2005) model uses mostly code-level metrics from specific projects while the model presented here uses higher project-level statistics gathered across many projects.

Identifying And Selecting Influential Factors

Factors which are most likely to influence the success/failure of FLOSS must first be identified and then incorporated into the model. Many papers have been published in regards to this, but most of the literature simply speculates on what factors might affect the success and offers reasons why. Note that measuring the success of a FLOSS project is still an open problem: some metrics have been proposed and used but unlike for commercial software, no standards have been established. Some possible success indicators are: