Software Engineering Process I Estimating Software Size

1
Software Engineering Process IEstimating
Software Size

• INFO 636
• Glenn Booker

2
Why Plan?

• As emphasized earlier, we need a good estimate of
  the amount of work to be performed, in order to
  predict effort and time accurately (per Boehm)
• Estimation is one of the most challenging aspects
  of managing software development, hence our
  substantial focus on it here

3
Estimation Example

• Other fields have well established formulas for
  estimating work
• Construction knows the cost per square foot of
  various types of construction
• More complex projects look at the linear amount
  of walls, and the areas of various parts (walls,
  ceilings, etc.) to develop good estimates

4
Size Estimation Process

• The framework, or process, for planning a project
  was covered last lecture
• Define system requirements
• Product conceptual design
• Estimate product size
• Estimate resources and schedule
• Develop the product
• Refine basis for later estimates

5
Estimation Tools

• Most software estimation tools have been
  calibrated to use software size as an input, and
  produce effort and schedule as outputs
• COCOMO, SLIM, PriceS, and McConnells tables in
  Rapid Development
• Often start at fairly large project sizes, e.g.
  10,000 LOC and up

6
Estimation Tools

• We need a basis for estimation which works for an
  individual (programmer)
• Most organizations use either no estimation
  methods, or use terribly unreliable ones
• 100 error is far too common

7
Desired Estimation Goals

• Criteria for a good estimation method include
• Use structured and trainable methods
• Should apply to both development and maintenance
• Should be able to handle all aspects of
  development, not just code

8
Desired Estimation Goals

• It should be suitable for statistical analysis
• It should be adaptable to future types of work
• It should be possible to judge the accuracy of
  your work (and hence refine the model)
• Well briefly cover four estimation methods, then
  explain the proxy-based PROBE approach

9
Estimation Methods

• Wideband-Delphi Method
• Fuzzy Logic Method
• Standard Component Method
• Function Point Method
• Proxy-based Estimating

10
Wideband-Delphi Method

• This method was developed by Rand Corporation
• It uses several people to estimate the same task,
  then applies a Delphi method to get a consensus
  estimate
• The process is
• Discuss the problem

11
Wideband-Delphi Method

• Get anonymous estimates, and hand them to a
  moderator
• Find the median estimate, and show everyone the
  set of estimates
• Discuss the results, to uncover different views
  of the project scope
• Repeat the process until estimates converge to
  within a predefined range

12
Fuzzy Logic Method

• This approach uses historic data to arrive at
  some meaningful estimates based on qualitative
  descriptions
• Size categories such as Very Small, Small,
  Medium, Large, and Very Large
• How data are divided into these categories
  depends on the type of data

13
Fuzzy Logic Method

• Data with a small range (say, a factor of five
  from very small to very large) can use a linear
  divisions
• Data with a large range can use a base 10
  logarithmic division (as shown in the text)

14
Fuzzy Logic Method

• Linear division breaks up sizes into evenly
  divided pieces
• Heres an example for the N track
• If your work to read the text involves chapters
  from 23 to 75 pages long (I made those numbers
  up), then the range of sizes is 75-2352 pages
• Divide that range into five pieces by dividing by
  four 52/4 13

15
Fuzzy Logic Method

• The midpoints of each size are just the lowest
  size, then add the 13 four times
• Very Small midpoint 23 pages
• Small midpoint 231336 pages
• Medium midpoint 2313249 pages
• Large midpoint 23 13362 pages
• Very Large midpoint 23 13475 pages (which
  equals the largest chapter size)

16
Fuzzy Logic Method

• Use half of 13, or 6.5, to find the ranges for
  each size
• Very Small range is up to 236.529.5 pages
• Small range is 29.5 to 366.542.5 pages
• Medium range is 42.5 to 496.555.5 pages
• Large range is 55.5 to 626.568.5 pages
• Very Large range is 68.5 pages and up
• Notice each categorys range is also 13 pages,
  since we have linear divisions

17
Fuzzy Logic Method

• The logarithmic version is messier, since we have
  to
• Convert the sizes to their logarithms
• Follow the linear approach using the logarithms
• Take everything to the power of 10 to convert it
  back to the original units

18
Fuzzy Logic Method

• The example in the book has LOC ranging from 173
  to 10,341 LOC
• The log10 of 173 is 2.238
• The log10 of 10,341 is 4.014
• The difference is 4.014 2.238 1.776
• Divide the difference by four to get the interval
  1.776/40.444
• Mimic slide 15 to find the midpoints

19
Fuzzy Logic Method

• The midpoints of each size are just the lowest
  size, then add the 0.444 four times
• Very Small midpoint 2.238
• Small midpoint 2.238 0.444 2.682
• Medium midpoint 2.238 0.4442 3.126
• Large midpoint 2.238 0.4443 3.570
• Very Large midpoint 2.238 0.4444 4.014
  (which equals the largest code size)
• Mimic slide 16 to find the ranges of each size
  category

20
Fuzzy Logic Method

• Use half of 0.444, or 0.222, to find the ranges
  for the first size (then just keep adding 0.444
  to each range boundary)
• Very Small range is up to 2.2380.2222.460
• Small range is 2.460 to 2.4600.4442.904
• Medium range is 2.904 to 2.9040.4443.348
• Large range is 3.348 to 3.3480.4443.792
• Very Large range is 3.792 and up

21
Fuzzy Logic Method

• Now take 10 to the power of the logarithms to
  find the actual LOC
• Very Small range is up to 102.460288 LOC
• Small range is 288 to 102.904802 LOC
• Medium range is 802 to 103.3482228 LOC
• Large range is 2228 to 103.7926194 LOC
• Very Large range is 6194 LOC and up
• This is the basis for the poorly labeled table at
  the bottom of page 104 in the text

22
Fuzzy Logic Method

• An asideTables 5.2 in the text divide each of
  the five basic categories (Very Small, etc.) into
  five more subranges
• This follows the same approach, just adding more
  detail to each category
• Its unlikely youll have enough data to worry
  about subranges

23
Standard Component Method

• The Standard Component Method, by Putnam,
  assumes you have a substantial database from
  which to make your estimates
• Make a realistic estimate of how many screens you
  think will be in your system
• Estimate the lowest and highest possible numbers
  of screens you could imagine will be in your
  system

24
Standard Component Method

• For actual estimation, usen (lowest number
  highest number 4realistic number)/6
• The idea is to try to account for possible error
  in your estimate
• Repeat this process for each type of component in
  your system

25
Function Point Method

• The function point approach uses function
  points as a proxy for the complexity of the
  system, independent of the programming language
  used
• See ISYS 420, lecture 8 for details of this
  approach

26
Function Point Method

• Each input or output function, interface, file,
  and inquiry is judged on a fixed complexity scale
  of small to large (not shown in the Humphrey
  text), and assigned some number of function
  points
• The total number of function points is adjusted
  for 14 influence factors, such as the
  developers expertise, business environment, etc.

27
Function Point Method

• While a great language-independent method for
  judging the complexity of a program, it isnt as
  reliable for estimating development effort
• See IFPUG for more details

28
Proxy-based Estimating

• We are trying to predict the final size of a
  software product
• Measuring or estimating that directly is tricky
  at best, so we use proxies to help get there
• A proxy is an intermediate concept or substitute
  for what we really want to predict

29
Proxy-based Estimating

• The overall process is like this
• We want to take the conceptual design, and break
  it into parts which correspond to the proxies
  available
• Estimate each part of the system, based on the
  proxies
• Add them up to get the overall product size

30
Choosing a Proxy

• The proxy size should correspond to the
  development effort size
• Proxy content should be countable and easy to
  visualize
• Proxy must be customizable
• The proxy should be sensitive to the same factors
  which affect development

31
Possible Proxies

• In a manner similar to function points, any
  characteristic of the system could be proxies
• Input screens, output reports, data files
• Objects or classes
• The fuzzy logic and function point concepts are
  essentially blended to produce the PROBE approach

32
PROBE Method

• PROxy-Based Estimation (PROBE) uses objects as
  proxies
• See also Appendix C, Tables C36 and C40
• First choose appropriate proxy categories (e.g.
  Table 5.7, p. 117)
• For code, calculation, data, I/O, control, print,
  etc. might be suitable proxies
• Reading, discussion, homework, (N track)

33
PROBE Method

• Choose reasonable size options for the proxies
• For class, you might only have enough data for
  three sizes instead of five
• Analyze your historic data to determine
  approximate sizes (LOC) for each proxy
• For N track, the amount of effort needed

34
PROBE Method

• Now start using your method for a given
  assignment
• Develop a conceptual design for the solution
• Use your proxies to estimate the amount of code
  or effort needed to develop them
• The example on page 120 is the first use of form
  C39 (p. 683)

35
A Course Note

• P track students will use the estimating pretty
  much as written in the text
• Our forms are slightly different
• N track students will develop their own proxies
  to correspond to their weekly activities, and
  create a custom form N39 to follow a similar
  process

36
PROBE Method

• The BASE PROGRAM section of C39 is a summary of
  the expected changes to the preexisting code
• Base Size (B) is the amount of code already
  present
• LOC Deleted (D) is how much existing code you
  plan to remove
• LOC Modified (M) is how much existing code you
  expect to change

37
PROBE Method

• The PROJECTED LOC section contains
• Base Additions (BA) are planned additions to
  existing code (new lines within existing modules)
• New Objects (NO) are new modules or classes which
  will need to be implemented
• Your proxy structure is used to describe the
  Type, Methods, and Relative Size of the changes
  to BA and NO

38
PROBE Method

• The REUSED OBJECTS (R) section of C39 is used to
  describe
• Code youll reuse from another preexisting
  source
• Code youll create during this assignment which
  will be reusable
• These tend to be rare during the course

39
PROBE Method

• Now comes the number crunching part
• The Projected LOC (P) is the total amount of new
  development for this assignment P BA NO
• The terms b0 (hereafter beta0) and b1 (beta1)
  are linear regression parameters from your work
  history
• By now you have a history of planned LOC or
  effort, and actual

40
PROBE Method

• What the flock are beta0 and beta1?
• The classic equation for a line is y mx b
• m is the slope, which corresponds to beta1
• b is the y-intercept, which is beta0
• Here the x axis is the planned LOC or effort,
  and the y axis has actual values

41
PROBE Method
42
PROBE Method

• See regression handout for an example of
  calculating beta0 and beta1
• Note that Sxi2 means S(xi2) not S(xi)2
• When you use this, make sure the formulas are
  correct
• n changes each week as new data is created

43
PROBE Method

• Incidentally, if your estimates are always
  perfect, youd have beta1 1, and beta0 0
  (why?)
• Once you have beta0 and beta1, find
• New and Changed LOC (N) beta0 beta1(P M)
• Its critical to note that later calculations
  for prediction interval use N, not P

44
PROBE Method

• The expected size of the application after this
  project is
• Total LOC (T) N B - D M R
• The Total New Reused is the sum of code flagged
  (with a ) in the New Objects section which are
  being reused
• Dont need to use this very often

45
PROBE Method

• Then we get to the Range calculation
• We have a refined estimate of the size of the
  system, but want to establish a prediction
  interval in which the real outcome is likely to
  fall
• See the PSP_Calculation_Example.xls spreadsheet

46
PROBE Method

• To find the Range, we start with a parameter from
  the t distribution
• Called t(a/2, n-2) where
• a/2 is the width of the prediction interval
  generally 70 or 90
• n-2 is the number of degrees of freedom again,
  n is the number of data pairs
• In Excel, use TINV(1 - a/2, n - 2)

47
PROBE Method

• Next we need the standard deviation, s
• Thats why column G adds up(Yi - b0 b1Xi)2
• s sqrt S(Yi - b0 b1 Xi)2 / (n-2)
• Now theres a new term, xk (xk)
• xk P M
• This is the same term used in the N formula the
  projected and modified LOC

48
PROBE Method

• Now use this to plug into formula 5.3 on page 124
• Im not going to copy it here ?
• Notice in the spreadsheet the column H
  calculation of(Xi - Xavg)2which is also used
  to find the Range

49
PROBE Method

• Finally, find the Upper and Lower Prediction
  Intervals (UPI and LPI)
• UPI N Range
• LPI N Range
• The Prediction Interval Percent is either 70 or
  90, the value used to find t

50
PROBE Method

• If Range is comparable to N in magnitude
• Choose a Prediction Interval Percent of 70 to
  keep Range smaller, and/or
• Look for data fliers which can have a strong
  influence on sigma (s)
• E.g. data points with relatively large value of
  (Yi - b0 b1Xi)2

51
Object Size Ranges

• The fuzzy logic method (starting on slide 12)
  summarizes the two most likely approaches for
  defining size ranges based on your historic data
• A Linear approach, generally best if the range of
  the data is well under a factor of 10
• A logarithmic approach for wider range data

52
Object Size Ranges

• If your work is following a true normal
  distribution, then your objects should have
• 6.68 each in Very Small and Very Large
  categories
• 24.17 each in Small and Large categories
• 38.30 in the Medium category
• Its good to see if this holds

53
Object Size Ranges

• If your object size distribution is really
  skewed, you could
• Reconsider the size categories
• Look for better proxies
• See if your design approach is leaning toward
  very large or very small objects, or very
  inconsistent object sizes

54
N Track Notes

• Youll use most of the preceding discussion
• Youll have different proxies instead of the
  Base Program, Projected LOC, and Reused
  Objects
• Youll have some equivalent of P and N, and
  still find beta0, beta1, and Range
• Your P and N will measure time instead of LOC
• Youll still find prediction intervals UPI, LPI

55
Improving Estimation

• We tend to try to estimate many small things for
  a large task
• The estimation errors tend to cancel each other
  somewhat
• The PSP allows you to know what your estimation
  errors have been, and hence improve later
  estimates
• Though thats hard to see during the term

56
Improving Estimation

• As you follow this consistently, your values for
  beta0 and beta1 will tend to stabilize
• Then you dont have to keep recalculating them!
• If you get really weird beta0 and beta1, or have
  no history yet, look at other options for
  refining your estimate, on page 679 (Table C35)

57
Improving Estimation

• On large projects, look for a consistent, and
  fairly low, level of abstraction
• The conceptual design might need to be refined to
  provide enough detail for a good estimate
• If a single object performs the work of many
  kinds of proxies, then it probably needs to be
  broken down

58
Improving Estimation

• Estimating products which have no precedent is
  really tough
• Make sure the level of uncertainty is clear to
  your customer
• Avoid overcompensating for your own history of
  errors
• Make small changes in your approach and try them
  for a while