Aplicando Aprendizado de Máquina para Estimativa de Esforço no Desenvolvimento de Software.

• Ano de defesa: 2020
• Autor(a) principal: CORRÊA, Weldson Amaral
• Orientador(a): SANTOS, Davi Viana dos
• Banca de defesa: SANTOS, Davi Viana dos , BRAZ JUNIOR, Geraldo , RIVERO CABREJOS, Luis Jorge Enrique , GRACIANO NETO, Valdemar Vicente
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal do Maranhão
• Programa de Pós-Graduação: PROGRAMA DE PÓS-GRADUAÇÃO EM CIÊNCIA DA COMPUTAÇÃO/CCET
• Departamento: DEPARTAMENTO DE INFORMÁTICA/CCET
• País: Brasil
• Palavras-chave em Português: Estimativa de Esforço; Aprendizado de Máquina; Aprendizado de Máquina Automatizado; Esforço de Desenvolvimento de Software
• Palavras-chave em Inglês: Effort Estimation; Machine Learning; Automatic Machine Learning; Software Effort Development
• Área do conhecimento CNPq: Engenharia de Software
• Link de acesso: https://tedebc.ufma.br/jspui/handle/tede/3289

Estimates in software projects aim to help practitioners predict more realistic values on software development, impacting the quality of software process activities regarding planning and execution. However, software companies have difficulties when carrying out estimations that represent adequately the real effort needed to execute the software project activities. Although, the literature presents techniques to estimate effort, this activity remains complex. Recently, Machine Learning (ML) techniques are been applied to solve this problem. Through ML techniques it is possible to use databases of finished projects (datasets) to help get more precisely estimations. However, the estimations depends on the dataset they are applied. This research propose a methodolody based on automatic machine learning to generalize the estimations through many datasets. To evaluate our methodology, we conducted tests with ten datasets using four metrics: Mean Absolute Error, Median Magnitude Relative Error, Mean Magnitude Relative Error and Percentage Relative Error Deviation. The results show that the proposed methodology has consistent estimations for sofware effort based on the employed metrics, which indicates that our methodology is promising and can generalize the results to other datasets.