Cross Trees- Visualizing Estimations using Decision Trees

Oct1 '14

with Rahul Krishna, NC State

Optimization has been the goal of almost every human thought and action. With growing computational capabilities, solutions to problems are also exponentially increasing. Literature proves that with rising demand for data and analytics on this large data, solutions to problems are multiplied. These solutions are supported with strong statistical and analytical reasoning that does help experts narrow down solutions. Optimizing these large set of solutions can get tricky to experts who seek knowledge of how these solutions have improved and what decisions need to be taken for remaining solutions to improve. This problem persists in software engineering with managers and experts taking decisions which decide the course of project.

This thesis proposes a method for optimizing solutions along with providing decisions that help improve a solution. Literature supports that landscape visualization of data gives an inside scoop of data behaviour. A method is proposed which takes benefit of visualizing data and improving solutions based on their position in landscape. Cross Trees are built by grouping data based on their similarities. Traversing from bad group of solutions to better group of solutions require few decisions to taken. These decisions are proposed by CrossTree and are tested for how valid they are. CrossTree is tested with two models which simulate software projects-POM3 and XOMO. Also, models are simluated with one of the best genetic algortihms NSGA-II, to generate set of optimized solutions. CrossTree is compared against results from NSGA-II to validate performance.