Reparation in Evolutionary Algorithms for Multi-objective Feature Selection in Large Software Product Lines

Software Product Lines Engineering is the area of software engineering that aims to systematise the modelling, creation and improvement of groups of interconnected software systems by formally expressing possible alternative products in the form of Feature Models. Deriving a software product/system from a feature model is called Feature Configuration. Engineers select the subset of features (software components) from a feature model that suits their needs, while respecting the underlying relationships/constraints of the system–which is challenging on its own. Since there exist several (and often antagonistic) perspectives on which the quality of software could be assessed, the problem is even more challenging as it becomes a multi-objective optimisation problem. Current multi-objective feature selection in software product line approaches (e.g., SATIBEA) combine the scalability of a genetic algorithm (IBEA) with a solution reparation approach based on a SAT solver or one of its derivatives. In this paper, we propose MILPIBEA, a novel hybrid algorithm which combines IBEA with the accuracy of a mixed-integer linear programming (MILP) reparation. We show that the MILP reparation modifies fewer features from the original infeasible solutions than the SAT reparation and in a shorter time. We also demonstrate that MILPIBEA outperforms SATIBEA on average on various multi-objective performance metrics, especially on the largest feature models. The other major challenge in software engineering in general and in software product lines, in particular, is evolution. While the change in software components is common in the software engineering industry, the particular case of multi-objective optimisation of evolving software product lines is not well-tackled yet. We show that MILPIBEA is not only able to better take advantage of the evolution than SATIBEA, but it is also the one that continues to improve the quality of the solutions when SATIBEA stagnates. Overall, IBEA performs better when combined with MILP instead of SAT reparation when optimising the multi-objective feature selection in large and evolving software product lines.