Técnicas de comitês para a estimação de esforço na correção de software

Abstract

A well-defined planning of a software project, since the early stages, is indispensable to its success, whether the development refers to product’s creation or maintenance. Accordingly to the software life cycle, maintenance is continuously executed after the product’s building and delivery, in parallel to the tests execution by engineers and/or users. In this stage, User Stories and Issue Reports are the first documents to be presented. These documents describe, in natural language, business requirements, error scenarios found, expected corrections and enhancements for the system. Its objectives are, among other things, ranking the activities needed to be accomplish during the project. Therefore, in line with the available resources – human, financial and temporal -, it is possible to estimate the effort that will be necessary in the activities development and generate essential information for an effective and efficient planning. As these documents are written in natural texts, it raises the opportunity to use Natural Language Processing and Machine Learning (ML) to predict software effort. In practice, in the daily life of software factories, it is common to use experts’ and project staff’s opinion to judge the effort required by an activity during Planning Poker sessions. Usually, in this technique, the effort is measured in Story Points, which follow Fibonacci sequence. But this planning model requires the scaling of more resources to be executed. The application of ML causes in a system, after the learning phase, the ability to seize the team experience and replicate it quickly and automatically to estimate the activities effort. Thus, this work covers the ML field, proposing a PV-DM Ensemble approach to extract features of Issue Reports to estimate Story Points, the effort indicator. Compared to the two other approaches of BoW and simple PV-DM, the proposed technique has presented good results, about 80% of f-measure, in a supervised learning SVM classifier. The experiments results proved to be a starting point for further study of PV-DM Ensemble approach and its improvement.