ABSTRACT

Software effort estimation (SEE) is a software engineering problem that requires robust predictive models. To establish robust models, the most feasible configuration of hyperparameters of regression methods is searched. Although only a few works, which include hyperparameter optimisation (HO), have been done so far for SEE, there is not any comprehensive study including deep learning models. In this study, a feed-forward deep neural network algorithm (FFDNN) is proposed for software effort estimation. The algorithm relies on a binary-search-based method for finding hyperparameters. FFDNN outperforms five comparison algorithms in the experiment that uses two performance parameters. The results of the study suggest that: 1) Employing traditional methods such as grid and random search increases tuning time remarkably. Instead, sophisticated parameter search methods compatible with the structure of regression method should be developed; 2) The performance of SEE is enhanced when associated hyperparameter search method is devised according to the essentials of chosen deep learning approach; 3) Deep learning models achieve in competitive CPU time compared to the tree-based regression methods such as CART_DE8.

摘要

软件工作量估计（SEE）是一个软件工程问题，需要强大的预测模型。为了建立鲁棒的模型，搜索回归方法的超参数的最可行配置。尽管到目前为止，针对SEE仅完成了很少的工作，其中包括超参数优化（HO），但是还没有包括深度学习模型在内的任何全面研究。在这项研究中，提出了一种前馈深度神经网络算法（FFDNN）用于软件工作量估计。该算法依赖于基于二进制搜索的方法来查找超参数。FFDNN在使用两个性能参数的实验中胜过五种比较算法。研究结果表明：1）采用网格和随机搜索等传统方法会显着增加调谐时间。代替，应开发与回归方法结构兼容的复杂参数搜索方法；2）根据所选深度学习方法的要点设计相关的超参数搜索方法，可以提高SEE的性能；3）与基于树的回归方法（例如CART_DE8）相比，深度学习模型可在竞争的CPU时间上实现。