An Investigation of Decomposition-Based Metaheuristics for Resource-Constrained Multi-objective Feature Selection in Software Product Lines

The multi-objective feature selection from software product lines is a problem that has attracted increasing attention in recent years. However, current studies on this problem suffer from two main limitations: (1) resource constraints are naturally adhered to the feature selection process, but they are completely ignored or inadequately handled, and (2) there is a strong preference to the use of evolutionary algorithms for the feature selection problem, and the suitability of other multi-objective metaheuristics remains to be fully explored. To address the above two limitations, this paper proposes the multi-objective feature selection with multiple linear and non-linear resource constraints, and investigates the performance of decomposition-based metaheuristics on the proposed problem. We construct a number of problem instances using both artificial and real-world software product lines, considering 2, 3 and 4 objectives. Experimental results show that, within the decomposition-based framework, reproduction operators that are based on probabilistic models (PM) perform better than genetic operators. Moreover, we demonstrate that adaptation of weight vectors can further improve the performance. Finally, we show that PMAD (the combination of PM-based reproduction operators, adaptation of weight vectors and decomposition-based framework) is better than several state-of-the-art algorithms when handling this problem.