On the configuration of multi-objective evolutionary algorithms for PLA design optimization

Abstract

Search-based algorithms have been successfully applied in the Product Line Architecture (PLA) optimization using the seminal approach called Multi-Objective Approach for Product-Line Architecture Design (MOA4PLA). This approach produces a set of alternative PLA designs intending to improve the different factors being optimized. Currently, the MOA4PLA uses the NSGA-II algorithm, a multi-objective evolutionary algorithm (MOEA) that can optimize several architectural properties simultaneously. Despite the promising results, studying the best values for the algorithm parameters is essential to obtain even better results. This is also crucial to ease the adoption of MOA4PLA by newcomers or non-expert companies willing to start using search-based software engineering to PLA design. Three crossover operators for the PLA design optimization were proposed recently. However, reference values for parameters have not been defined for PLA design optimization using crossover operators. In this context, the objective of this work is conducting an experimental study to discover which are the most effective crossover operators and the best values to configure the MOEA parameters, such as population size, number of generations, and mutation and crossover rates. A quantitative analysis based on quality indicators and statistical tests was performed using four PLA designs to determine the most suitable parameter values to the search-based algorithm. Empirical results pointed out the best combination of crossover operators and the most suitable values to configure MOA4PLA.