An enhanced salp swarm optimizer boosted by local search algorithm for modelling prediction problems in software engineering

Scientific communities are still motivated to create novel approaches and methodologies for early estimation of software project development efforts and testing efforts in soft computing environments due to scheduling and budgetary concerns. Therefore, the software engineering prediction problems (SEPPs) are formulated as machine learning (ML) models with the aim of addressing these issues. In such methodologies that may exhibit significant limitations and drawbacks, efficient metaheuristic approaches are essential to improving prediction performance. Accordingly, this study aims to address software test effort prediction (STP) and software development effort prediction (SEP) with the aim of maximizing prediction accuracy, which in turn minimizes overall project costs and optimizes resource allocation. To achieve this goal, we developed several ML models composed of a backpropagation neural network (BPNN). The proposed models contain the Salp Swarm Algorithm (SSA), which is utilized to replace the traditional network training method and tackle its limitations. The models also contain the great deluge (GD) local search algorithm, which is hybridized with the SSA algorithm to enhance optimization capabilities by finding more balance between exploration and exploitation. During the validation stage of this study, fourteen benchmark datasets were utilized to evaluate the developed models for each of the respective problems. The obtained results were quantified using eight performance metrics and compared across two sections. In the first section, a comparison was made between the results of the hybrid-developed model (HSSA) and those of the standard SSA algorithm and BPNN. In the second comparison, the performance of the HSSA model was compared with several contemporary techniques that are considered state-of-the-art. The evaluation shows that the HSSA performs better than related approaches in most cases for both problems. Finally, additional analysis was performed on the collected results, including examinations of statistical significance, distribution through box plots, and model convergence behavior.