Software Development Effort Estimations Through Neural Networks Guillermo

  • Slides: 14
Download presentation
Software Development Effort Estimations Through Neural Networks Guillermo Sebastián Donatti sdeetnn. weebly. com

Software Development Effort Estimations Through Neural Networks Guillermo Sebastián Donatti sdeetnn. weebly. com

Motivation Software to be Developed Software Size Estimation Software Effort Software Schedule

Motivation Software to be Developed Software Size Estimation Software Effort Software Schedule

Motivation Underestimate Overestimate Software Development Effort Incomplete Software Higher Costs Parkinson Law Poor Quality

Motivation Underestimate Overestimate Software Development Effort Incomplete Software Higher Costs Parkinson Law Poor Quality Missing Customers Missing Deadlines

Objectives l l Develop a neural network model to solve the software development effort

Objectives l l Develop a neural network model to solve the software development effort estimation problem using a large software development metrics database. Compare the performance of the developed neural network model with the ones of wellknown cost estimation tools.

Data Preprocessing Data Reconstruction Data Conversion Useful Data of 967 Software Development Projects (48%)

Data Preprocessing Data Reconstruction Data Conversion Useful Data of 967 Software Development Projects (48%) Data Scaling R E S U L T

Neural Network Model Input Layer Interface Variables Definition Output Layer

Neural Network Model Input Layer Interface Variables Definition Output Layer

Interface Variables Definition ISBSG Data Base Fields Domain Description Expertise Analysis RESULT Best. Output

Interface Variables Definition ISBSG Data Base Fields Domain Description Expertise Analysis RESULT Best. Output Subset Regression Variable Input Variables RESULT

Neural Network Model Function Points Input Layer Application Type Development Platform Resource Level Derived

Neural Network Model Function Points Input Layer Application Type Development Platform Resource Level Derived Count Approach Hidden Layers Output Layer Topology Search: Recording Method Language Type Feed Growth Neural Network … Forward Topology Refine: Neural Network Pruning. . . Summary Work Effort

Neural Network Growth Creation Training Judgment Neural Network Model RESULT

Neural Network Growth Creation Training Judgment Neural Network Model RESULT

Neural Network Pruning Initialization Error Analysis Calculation Pruning Calculation Refined Neural Network Model RESULT

Neural Network Pruning Initialization Error Analysis Calculation Pruning Calculation Refined Neural Network Model RESULT

Neural Network Model Input Layer Hidden Layers …. . . Output Layer Three Hidden

Neural Network Model Input Layer Hidden Layers …. . . Output Layer Three Hidden Layers

Comparative Analysis SLIM Algorithm Construx Estimate Monte Carlo Simulation Feed Forward Neural Network Test

Comparative Analysis SLIM Algorithm Construx Estimate Monte Carlo Simulation Feed Forward Neural Network Test Data Set Neural Network Model ISBSG Reality Checker Error Back Propagation Training Algorithm Linear Regression Algorithm

Conclusions Less Slightly Estimation Less Estimation Variability Higher Amount of Accurate Estimations More Limited

Conclusions Less Slightly Estimation Less Estimation Variability Higher Amount of Accurate Estimations More Limited Estimation Error Concentrated in Few Test Patterns Tendency to Overestimate Construx ISBSG Estimate Reality Checker Neural Network Model Tendency to Underestimate

Thank you! Questions? sdeetnn. weebly. com

Thank you! Questions? sdeetnn. weebly. com