Genetic algorithms to support software engineering experimentation

Empirical software engineering is concerned with running experimental studies in order to establish a broad knowledge base to assist software developers in evaluating models, methods and techniques. Running multiple experimental studies is mandatory, but complex and the cost is high. Besides, replications may impose constraints difficult to meet in real contexts. Researchers face additional problems and cost restrictions when conducting meta-analysis on combined data from multiple experiments. In this paper we are concerned with both issues, of assisting users in carrying out meta-analysis tasks and gathering a meaningful body of data from experimental studies. We show how the genetic algorithms optimization model can effectively handle a specific meta-analysis problem that is not amenable to standard statistical approaches. We also introduce an approach to expand the universe of data by mapping the experimental design and known results into a suitable genetic algorithm model that simulates new results. The simulation allows researchers to prospect how the variation of different experimental parameters affects the results, without incurring in the cost of actually running additional experiments. We show that it is possible to simulate statistically valid data, expanding the universe of data for analysis and opening up some interesting possibilities for replicators.