Ex-post and Ex-ante Predictive Models for Software Effort Estimation

Transcription

1 Ex-post and Ex-ante Predictive Models for Software Effort Estimation 1

2 Outline + The effort estimation problem + Our approach + The organizational context + The technical context + The data collected + The ex-post estimation equations + The validation process + The ex-ante estimation equation + Conclusions and further research 2

3 The effort estimation problem + Type of model + Type of sample + Type of scales + Type of approach + Estimation moment 3

4 The effort estimation problem : Type of model? + analogy based model experts' opinion based model + decomposition based model + prediction based model - statistical : generally use regression technics - experimental : use past data to predict effort - theoretical : use hypothesis on the resources - composite : mixed approach (experts' opinion & statistical models) 4

5 The effort estimation problem : Type of sample? + heterogeneous sample : different organizations, different technologies, etc + homogeneous sample : same organization, same technology, etc 0 ESSEC J.Akoka, D.Briolat, 1.Comyn-Wattiau 5

6 Th'e effort estimation problem : Type of scales? + mixed : from nominal to absolute scales + homogeneous : absolute scale of variables 6

7 he effort estimation problem : Type of approach? + top-down : the project as a whole - many cost drivers + bottom-up : the project as a sum of stages - less cost drivers for each stage + complementary approaches 7

8 Estimation moment? ex-post estimation + previous software effort as a function of project observed characteristics at time T >> end of project (top down : the project as a whole) >> end of stage (bottom up : the project as a sum of stages) 0 ESSEC J.Akoka, D.Briolat, 1.Comyn-Wattiau 8

9 The effort estimation problem : Estimation moment? ex-post estimation Step n-1 U Step n W Step n+l Y Step n+2 9

10 The effort estimation problem : Estimation moment? ex-ante estimation + At the beginning of the project (top down) + At the beginning of each stage (bottom up) + remaining software effort estimated from project estimated characteristics at time T >> expert guessing from + previous effort + specifications already known + remaining software effort estimated from project observed characteristics at time T 10

11 ÿ he problem : Estimation moment? - ex-post estimation T Step n-1 U Step n W Step n+l Y Step n+2 Effort (n) = f n (V, W) Effort (n+l) = fn+,(x, Y) ex-ante estimation T v X Step n-1 U Step n Stepn+l step n+2 f \ \ J X = gl (T, U, V, W) O ESSEC J.Akoka, D.Briolat, LComyn-Wattiau 11

12 Our approach : + the model : statistical + the sample : homogeneous + the variables : measured on absolute scales + the approach : bottom-up + the estimation moment : ex-post & ex-ante 12

13 The organizational context : + French insurance company employees + Registered capital : 30 MF + 22 regional branches 13

14 The technical context : + relational database + fourth generation language and lower case + design method : MERSE supported by PACBASE + project management method : SDMS methodology 14

15 he technical context : SDWS methodology stages + nformation Systems Planning (P) + Definition Study (DS) + System Design (SD) + Detailed System Design - External Detailed Design (EDD) - nternal Detailed Design (DD) + mplementation - Programming (PRG) - Testing (TST) + nstallation (NS) 15

16 The data collected : + Effort in maddays collected by project managers + At the end of each stage of project life cycle (update of a database) + Original data set (303 projects) + Reduced to 110 by considering only new projects development and main evolution of existing projects + Reduced to a dozen : >> homogoneous organizational and technical context >> all values for independent variables 16

17 The ex-post estimation equations + stepwise regressions + sensitivity analysis - 1% - 5% - 10% + no equation modification 17

18 The ex-post estimation equations : : n Number Variables used Variables in the Stepwise 1 retained by the stepwise Phases observations regressions regressions UPTOT, NT, EFF, NT OBJ, REL, DON, PSS, OP SCE, SORT, ENT, REC, ENT EDD 8 DD TST NS EFF, DON, OBJ, REL, PF, REC PF, REC, SORT, ENT, FC, EFF, OBJ. REL TAB, FC, SORT, ENT, MESS, MOD, EFF MOD, LG, COMP PF MOD LG, MOD LG, MOD, UT 1 LG PST. UT. SESS PST Final equations (1) E = NT 6mm+m REC R (4) E = MOD 0.97 (5) E = LG MOD (6) E = LG (7) E = PST 0.78 O ESSEC J.Akoka, D.Briolat, LComyn-Wattiau

19 The validation process : SDWS Phases DS Number of observations Final equations (1 ) E = NT MRE Average Error Prediction SD (2) E = REL +l,23 REC EDD DD (4) E = MOD PRG TST NS - (5) E = LG MOD (6) E = LG (7) E = PST O ESSEC J.Akoka, D.Briolat, LComyn-Wattiau 19

20 ~efinition Study Effort f (Number of nterviews) E = 9,31 + 1,32 NT Observed Effort Computed Effort "Computed Effort %" 0! Number of nterviews "Computed Effort +25%" 20

21 ~xternal Detailed Design Effort f (Number of Functional Procedures) - - Observed Effort Computed Effort Computed Effort +25% Computed Effort -25% Number of Functional Procedures O ESSEC J.Akoka, D.Briolat, LComyn-Wattiau

22 f (Number of Modules) E = -1,23 + 0,33 MOD e+ Observed Effort Computed Effort Computed Effort +25% Computed Effort -25% Number of Modules 0 ESSEC J.Akoka, D.Briolat, 1.Comyn-Wattiau

23 Testing Effort f (Number of Lines) E = 1,55 + 0,0023 LG + Observed Effort Computed Effort Computed Effort +25% Computed Effort -25% 23

24 nstallation Effort f (Number of Workstations) E = 0,86 + 0,32 PST + Observed Effort Computed Effort Computed Effort +25% Computed Effort - 25% 0 * Number of Workstations 24

25 The ex-ante estimation equation + not all steps for the dozen remaining projects + less successful regressions + number of Records REC (SD Phase) function of the number of Objects OB J(DS Phase) REC = 4.37"OB J

26 The ex-ante estimation equation RECord (SD) = f (OBJect) (DS) REC REC estimation REC estimation + 25% REC estimation - 25% Number of OBJects O ESSEC J.Akoka, D.Briolat, LComyn-Wattiau

27 Further research 4 Same variables for ex-post estimation equations in other firms? 4 Larger sample to improve ex-ante estimation equations O ESSEC J.Akoka, D.Briolat, LComyn-Wattiau 28