Investigation and prediction of open source software evolution using automated parameter mining for agent-based simulation

To guide software development, the estimation of the impact of decision making on the development process can be helpful in planning. For this estimation, often prediction models are used which can be learned from project data. In this paper, an approach for the usage of agent-based simulation for the prediction of software evolution trends is presented. The specialty of the proposed approach lies in the automated parameter estimation for the instantiation of project-specific simulation models. We want to assess how well a baseline model using average (commit) behavior of the agents (i.e., the developers) performs compared to models where different amount of project-specific data is fed into the simulation model. The approach involves the interplay between the mining framework and simulation framework. Parameters to be estimated include, e.g., file change probabilities of developers and the team constellation reflecting different developer roles. The structural evolution of software projects is observed using change coupling graphs based on common file changes. For the validation of simulation results, we compare empirical with simulated results. Our results showed that an average simulation model can mimic general project growth trends like the number of commits and files well and thus, can help project managers in, e.g., controlling the onboarding of developers. Besides, the simulated co-change evolution could be improved significantly using project-specific data.