Neuroevolution and Neuroswarm with Hooke-Jeeves Local Search Algorithm for Predicting the Construction Effort of Software Projects

Software construction corresponds to one of the software development life cycle (SDLC) activities. It consists of coding, unit testing, integration testing, and debugging activities. The SDLC effort (i.e. the number of person-hours) is an important variable predicted for budgeting the software projects. A common practice in software managers is to predict the effort by SDLC activity. Accordingly, we propose the incorporation of the Hooke-Jeeves (HJ) local search algorithm in particle swarm optimization (PSO), and genetic algorithms (GA) for tuning the hyperparameters of three types of neural networks, whose accuracies were compared to that of a statistical regression model. All of the models were applied for predicting the construction effort of software projects selected from an international public repository of software projects. Results allow concluding that a fully connected neural networks optimized with HJ, GA, and PSO can be used by software managers for predicting the construction effort of new or maintained software projects.软件构建是软件开发生命周期（Software Development Life Cycle，SDLC）的一项活动，涵盖编码、单元测试、集成测试与调试等工作环节。SDLC工作量（即人时数）是软件项目预算编制中至关重要的预测变量。软件项目经理通常会依据SDLC各活动环节预测项目工作量。据此，我们提出将Hooke-Jeeves（HJ）局部搜索算法嵌入粒子群优化（Particle Swarm Optimization，PSO）与遗传算法（Genetic Algorithm，GA）中，用于调优三类神经网络的超参数，并将上述模型的预测精度与统计回归模型的精度进行对比。所有模型均被应用于预测从国际公开软件项目仓库中遴选的软件项目的构建工作量。研究结果表明，经HJ、GA与PSO优化的全连接神经网络（Fully Connected Neural Network），可被软件项目经理用于预测全新或已维护软件项目的构建工作量。