PowerPoint Presentation by N6Tc7ce

VIEWS: 0 PAGES: 14

									         Calibrating Function Points
        Using Neuro-Fuzzy Technique

  Vivian Xia             Danny Ho                 Luiz F Capretz

   IT Department                              Department of Electrical and
                        NFA Estimation Inc.
    HSBC Bank                                    Computer Engineering
                      London, Ontario, Canada
   Vancouver, BC                              University of Western Ontario
                     danny@nfa-estimation.com London, Ontario, Canada
      Canada
Vivian_xia@hsbc.ca                                  lcapretz@eng.uwo.ca




                                                                       1
Roadmap
    Concepts of Calibration
    Neuro-Fuzzy Function Points Calibration Model
    Validation Result
    Conclusions




                                                2
Calibration Concept
     Internal Logical File (ILF) Complexity Matrix
                                      Data Element Types (DET)

        Record Element Types (RET)    1-19      20-50     51+

                    1                 Low       Low      Average
                   2-5                Low      Average    High
                   6+                Average    High      High

        DET, RET --- Component Associated Files
     Same methodology for all FP 5 components
       External Input, External Output, External Inquiry
          Internal Logical File, External Interface File           3
Calibration Concept Cont’d
    e.g. One project has 3 Internal Logical Files (ILF)
                             ILF A           ILF B           ILF C
  DET                         50               20                19
  RET                         3                3                 3
  Original Classification   Average         Average          Low

  Original Weight Value       10               10                7
  Observation 1             Ambiguous Classification
  Observation 2                                 Crisp Boundary


    Calibrate complexity degree by fully utilizing the number
     of component associated files
    Calibrate to fit specific application                     4
Calibration Concept Cont’d
           Unadjusted Function Points Weight Values
           UFP weight values are determined in 1979 based on
           Albrecht’s study of 22 IBM Data Processing projects

                    Component         Low   Average   High
            External Input             3      4        6
            External Output            4      5        7
            Internal Logical File      7      10       15
            External Interface File    5      7        10
            External Inquiry           3      4        6

  .
     Calibrate UFP weight values to reflect global software
      industry trend                                             5
Neural Networks Basics
 Learning from Data Source
      Adapting capability
      Modeling any complex nonlinear
       relationships
      Lack of explanation: “black box”
      Cannot take linguistic information
       directly




                                            6
Neuro-Fuzzy Function Points
Calibration Model Overview

                     Estimation
    ISBSG 8           Equation

     Project Data
                                  Validated for better
                                      estimation

                     Calibrated
                         by
      Calibrated      Neural      MMRE, PRED
    by Fuzzy Logic    Network




                                                         7
Calibrating by Fuzzy Logic
              Fuzzy Logic System




            Fuzzy Set        Fuzzy Rule
   Input                                  Output



                  Fuzzy Inference




                                                   8
Calibrating by Neural Network
    Learn UFP weight
                                               X1
     values by effort                 NINLOW          w1
        the values should reflect
         complexity                                  w2          Y   v1
                                    NINAVG     X2                         Z   Effort
        complexity proportioned to
         effort
    15 UFP inputs as                                 w15
     neurons
                                               X15          v2
                                      NIXHGH
    Back-propagation
     algorithm
                                          1    X16

                                                                              9
Data Source --- ISBSG Release 8
    ISBSG
        International Software Benchmarking Standards Group
        Non-profit organization
    Release 8 Contains 2,027 projects
    75% built in recent 5 years
    Filter on ISBSG 8 data set
        Filter Criteria:
            Quality, Counting method, Resource level,
              Development Types, UFP breakdowns
        Shrink to 184 projects
                                                               10
Validation Methodology
    Developed a calibration tool
    Randomly split data set
        totally 184 data points
        100 training points
        84 testing points for validation
    Repeat 5 times
    Using estimation equation for comparison


                                                11
Validation Results (MMRE)
              Exp.1   Exp.2   Exp.3   Exp.4   Exp.5
                                                         MMRE:
MMRE          1.38    1.58    1.57    1.39    1.42           Mean Magnitude
Original
                                                              of Relative Error
MMRE          1.10    1.28    1.17    1.03    1.11
Neuro-Fuzzy                                              Criteria to assess
IMPRV %                                                   estimation error
              20%     19%     25%     26%     22%        The lower the
Avg.                          22%                         better
IMPRV %



                     1 n Esti  Acti
               MMRE  
                     n i 1 Acti
                                                                            12
Validation Results (PRED)
             80.0
                                                                             PRED:
             70.0
                                                                                  Prediction at level p
             60.0
                                                                             Criteria to assess
             50.0                                                             estimation ability
Percentage




             40.0
                                                             Original
                                                             Calibrated
                                                                             The higher the better
             30.0

                                                                                             k
             20.0
                                                                                  PRED( p) 
             10.0                                                                            n
              0.0
                    Pred 25   Pred 50   Pred 75   Pred 100


                                                                                                           13
Conclusions
   Neuro-Fuzzy Function Points model improves
    software cost estimation by an average of 22%.
   Fuzzy logic calibration part improves UFP
    complexity classification.
   Neural network calibration part overcomes
    problems with UFP weight values.



                                                 14

								
To top