Documentos de Académico
Documentos de Profesional
Documentos de Cultura
ExpertFit Version 8
User’s Guide
by
1. Introduction ................................................................................................................ 1
1.1. Types of ExpertFit Analyses............................................................................... 3
1.2. Type and Amount of Data Required by ExpertFit............................................... 6
1.3. Installation Instructions ....................................................................................... 8
1.4. ExpertFit Help System........................................................................................ 9
1.5. ExpertFit Software Architecture........................................................................ 10
2. Data Analysis Module – Standard Mode .................................................................. 11
2.1. Confirmation of the Best Distribution ................................................................. 15
2.2. Examples .......................................................................................................... 16
Example 2.1: Customer Service Times............................................................. 17
Example 2.2: Ship-Loading Times .................................................................... 32
Example 2.3: Repair Times for a Machine ........................................................ 38
Example 2.4: Weekly Product Sales ................................................................. 43
3. Data Analysis Module – Advanced Mode................................................................. 50
3.1. Example ............................................................................................................ 56
Example 3.1: Testing the Homogeneity of Two Data Sets ............................... 57
4. Task-Time Models Module ....................................................................................... 62
4.1. Organization and Options ................................................................................. 63
4.2. Example ............................................................................................................ 65
Example 4.1: Modeling a Task Time................................................................. 66
5. Machine-Breakdown Models Module ....................................................................... 71
5.1. Organization and Options ................................................................................. 76
5.2. Examples .......................................................................................................... 78
Example 5.1: Modeling Machine Downtimes in the Absence of Data ............... 79
Example 5.2: Continuation of Previous Example .............................................. 83
6. Distribution Viewer ................................................................................................... 87
7. Batch Mode .............................................................................................................. 88
References .................................................................................................................... 89
Appendix A: Distributions Included in ExpertFit ........................................................... 90
Index ........................................................................................................................... 133
1. Introduction
1
be stored in a Project for later reuse. All ExpertFit results (i.e., charts and tables) can
be printed or copied to the Windows Clipboard for use in other applications (e.g.,
Microsoft Word or Excel).
2
1.1. Types of ExpertFit Analyses
ExpertFit can perform the three main types of analyses given in Table 1.1.
Module Description
Data Analysis Used to determine what probability distribution best
represents a data set. You can either have ExpertFit
determine the best distribution automatically or specify
the distributions for consideration manually. See
Chapters 2 and 3 for further discussion and examples.
3
The probability distributions available in ExpertFit are given in Table 1.2.
Continuous Discrete
beta log-Laplace Bernoulli
Cauchy log-logistic binomial
chi-square lognormal geometric
Erlang normal hypergeometric
error Pareto logarithmic series
exponential Pearson type 5 negative binomial
exponential power Pearson type 6 Poisson
extreme value (2 types) power function uniform
F random walk
gamma Rayleigh
inverse Gaussian Student’s t
inverted Weibull triangular
Johnson SB uniform
Johnson SU Wald
Laplace Weibull
logistic
4
If you are using a single-vendor version of ExpertFit, then the resulting probability
distributions are put into the format for that vendor’s software. The Analyst versions of
ExpertFit support Crystal Ball and @RISK.
5
1.2. Type and Amount of Data Required by ExpertFit
ExpertFit requires that your data set be in ASCII format and contain between 10
and 100,000 observations (larger data sets are truncated with an indication given).
There may be one or more data values per line. In the latter case, observations should
be separated by blanks and, if desired, by commas. In general, data files should follow
the simple ASCII file format created by editors like Notepad.
You can copy data from an active spreadsheet (e.g., Excel) to the Clipboard and
then paste it into ExpertFit (see the tutorial in general Help).
If all of the data values in the file are integers in the range –214,748,345 to
214,748,345, then the sample will be considered to be integer; otherwise, it will be
taken to be real. Real values must have magnitudes between -1.0E+99 and 1.0E+99.
If all of the data values are integers, then you will be asked whether the data should be
considered to be real valued.
We recommend that the following ideas be used in collecting or analyzing a data
set:
1. If, at all possible, collect at least 100 observations on the random phenomenon of
interest, with 200 observations providing more ability to discriminate between two
distributions. In general, the benefit from increasing the sample size from 200 to
300 will be less than that provided by increasing the sample size from 100 to 200,
etc.
2. If you are collecting observations on a continuous random variable (e.g., a service
time), then the data values should have enough resolution so that the sample will
have a “large” number of distinct values. Otherwise, it will be difficult, in general, to
find a continuous distribution that provides a good representation.
3. If the available data values are integer, then you may want to convert them to real
numbers. ExpertFit contains many more continuous distributions than discrete
distributions.
4. You should understand the process that produced the data, rather than treating the
observations as just abstract numbers. For example, suppose your data set
contains a few extremely large observations – these are called outliers. If you don't
6
understand the problem context, then it will be difficult to know whether these large
observations are really legitimate or, perhaps, the result of measuring or recording
errors.
5. If you have collected times of arrival of “customers,” then these can be converted to
interarrival times using the ExpertFit transformation “DIFF.”
7
1.3. Installation Instructions
8
1.4. ExpertFit Help System
There are two main types of help available in ExpertFit: context-dependent and
general. Context-dependent help is available for every options and results screen, and
is accessed by clicking on the displayed Help button.
General help is accessed by clicking on the Help pull-down menu in the Menu
Bar at the top of the screen. This menu contains the following entries:
• Contents
• Context-Dependent
• Glossary
• Search
• Tutorials
• User’s Guide
• About ExpertFit
9
1.5. ExpertFit Software Architecture
ExpertFit uses the concept of a Project, which is a file containing one or more
analysis items. In the case of a simulation study, a Project could contain several items
corresponding to different data sets and their corresponding ExpertFit analyses, and
other items corresponding to Task-Time Models or Machine-Breakdown Models. A
Project allows you to save the results of an ExpertFit analysis for future reuse.
When an ExpertFit Project is created or read from a file, a Project Window is
created to represent it. This window acts as a directory for the analysis items contained
in the Project. Buttons at the bottom of the window allow you to create new elements,
to change an element’s name or description, to begin a new (or return to an existing)
analysis, and to delete an element.
10
2. Data Analysis Module – Standard Mode
There are two modes of operation (see the Mode pull-down menu) for the Data
Analysis module: Standard and Advanced. Standard Mode (the default), which is
described in this chapter, is sufficient for 95 percent of all analyses and is easier to use.
It focuses the user on those features that are the most important at a particular point in
an analysis. Advanced Mode, which is discussed in Chapter 3, contains a large
number of additional features for the sophisticated user. A user can switch from one
mode to another at any time.
The use of the Data Analysis module to determine what probability distribution
best represents a data set is based on the sequential application of the four tabs shown
in Table 2.1.
Data Used to read in a data set from a file, to enter a data set at the
keyboard, or to paste in a sample from the Clipboard
The options available in these four tabs are shown in Tables 2.2 through 2.5,
respectively.
11
Table 2.2. Options for the Data tab.
12
Table 2.4. Options for the Comparisons tab.
Distribution-Function-Differences Plot
P-P Plot
13
Although there are different ways that these four tabs could be used to
determine the best distribution for a data set, the following are the explicit steps that we
recommend for real data (see Example 2.4 for integer data):
1. Obtain a data set using the Data tab – see Section 1.2 for a discussion of the type
and amount of data required.
2. View the resulting Data-Summary Table (at the Data tab) – provides information on
the center, shape, and range of the true density (mass) function.
3. Make a histogram of your data (used in Step 5) using the Data tab – see the
Constructing a Histogram from Your Data tutorial in the online help.
4. Select the distribution that is the best representation for your data using the
Automated Fitting option at the Models tab.
5. Confirm using the Comparisons tab that the best distribution as determined by
ExpertFit is, in fact, satisfactory in an absolute sense – See Section 2.1 for
recommendations.
6. If you are doing simulation modeling, then either represent the best-fitting
distribution (if good in an absolute sense) or an empirical distribution based on your
data (if the best distribution is not satisfactory) in your simulation software using the
Applications tab.
Four examples of the use of the Data Analysis module are given in Section 2.2.
14
2.1. Confirmation of the Best Distribution
Before actually using the best model, we recommend that some amount of
confirmation of this model be done using the options in the Comparisons tab. We
suggest that a Density-Histogram Plot and/or a Frequency-Comparison Plot be
made with an appropriately constructed histogram. When the maximum of the density
function of the best model is “far” from x = 0, then the former plot is probably preferable.
If, on the other hand, the maximum occurs “close” to x = 0, then the latter plot will often
be more useful. This is because it may be difficult in this case to determine from a
histogram whether the true underlying density function strictly decreases as x increases
(similar to an exponential density) or whether the density function has its mode (x-value
where the maximum occurs) close to x = 0 (similar to a lognormal density with shape
parameter = 3/2). Care must be taken when using these plots since the choice of the
histogram intervals is somewhat subjective.
We also recommend that a Distribution-Function-Differences Plot and a
P-P Plot be used to confirm the quality of the best model.
Finally, one might perform the Anderson-Darling Test, the Kolmogorov-
Smirnov Test, and/or the Chi-Square Test in order to get a formal evaluation of the
best-fitting model.
15
2.2. Examples
We now present four examples of the use of the Data Analysis module,
following the six-step approach outlined at the beginning of this chapter. For the first
example, we give the ExpertFit commands necessary to accomplish a particular part of
the analysis and then the actual results of the analysis. For the other examples, we
only discuss the results.
Examples 2.1 through 2.3 use High Precision (the default) – see the “Overview”
in the Precision pull-down menu. It doesn’t matter which option you use for the integer
data of Example 2.4, since the two options are identical in this case.
16
Example 2.1: Customer Service Times
At window: Do:
Project 1 Click on New.
Project-Element Editing Select Fit distributions to data.
In the Project-Element Name edit box, enter
Example 2.1.
Click on OK.
Project 1 Click on Analyze.
Data tab Click on Enter Data.
Enter-Data Options Click on Apply.
Open Select Example 21 in the ExpertFit folder.
A: The Data-Summary Table for this set of n = 450 service times (read in the Data
tab) is given in Table 2.6. The positive value of the sample skewness indicates that the
underlying distribution of the data is skewed to the right (i.e., it has a longer right tail
than left tail). This is supported by the sample mean being larger than the sample
median.
17
Table 2.6. Data summary for the service-time data.
18
Steps for Action B:
At window: Do:
Data tab Click on Histogram.
Histogram Options Click on Apply.
Click on Done.
Click on Done.
B: In Figure 2.1 we present the default ExpertFit histogram for the service-time data.
Note that the histogram is quite “ragged,” since the interval width is too small. Using a
trial-and-error approach discussed in the Constructing a Histogram from Your Data
tutorial, we determined that a better histogram is obtained by using an interval width of
0.25. The improved “smooth” histogram is shown in Figure 2.2. In general we
recommend that you construct your own histogram rather than rely on the ExpertFit
default. There is no definitive prescription for choosing histogram intervals!
Note that the histogram interval width can also be changed by using the
two buttons with arrowheads at the top of the histogram screen. The left (right)
19
Histogram
0.10
0.08
0.06
Proportion
0.04
0.02
0.00
0.13 0.53 0.93 1.33 1.73 2.13 2.52 2.92
Interval Midpoint
23 intervals of w idth 0.133
20
Histogram
0.17
0.13
0.10
Proportion
0.07
0.03
0.00
0.13 0.63 1.13 1.63 2.13 2.63 3.13
Interval Midpoint
13 intervals of w idth 0.25
Figure 2.2. Histogram of the service-time data with an interval width of 0.25.
21
button decreases (increases) the interval width by 5 percent, and can be applied
repeatedly.
22
Steps for Action C:
At window: Do:
Data tab Click on the Models tab.
Models tab Click on Automated Fitting.
for our data by selecting the Automated Fitting option at the Models tab. Based on
certain heuristics, ExpertFit determined that the “best” representation for the data is
provided by a Weibull distribution (see Table 2.7) with location, scale, and shape
parameters of 0, 1.334, and 2.183, respectively. This best model received a Relative
Score of 100.00 and its Absolute Evaluation message is “Good,” indicating no reason
for concern. (See the context-dependent online help for a discussion of the terms in
boldface.) Furthermore, the model mean and the sample mean are almost identical.
Note that the third-best fitting model is a Rayleigh distribution with an estimated
location parameter (denoted by “E”) of 0.062. (If we were to click on View/Delete
Models at the Models tab, we would see that the normal distribution was not
automatically fit to our non-negative service-time data. This is because the normal
distribution can take on negative values. However, the normal distribution could, if
desired, be fit to our data using the Fit Individual Models option at the Models tab.)
23
Table 2.7. Evaluation of the candidate models.
Evaluation: Good
Suggestion: Additional evaluations using Comparisons Tab might be informative.
See Help for more information.
24
Steps for Action D:
At window: Do:
Comparisons tab Click on Graphical Comparisons.
Graphical-Comparisons
Options Select Density-Histogram Plot.
Click on Apply.
Density-Histogram
Plot Click on Done.
using the Comparison tab, as suggested by the latter part of the Absolute Evaluation
message. The Density-Histogram Plot based on the final histogram is shown in
Figure 2.3. The closeness of the density function to the histogram visually confirms the
quality of the Weibull representation. Note that we could have simultaneously
plotted the density functions of several distributions in Figure 2.3.
25
Density-Histogram Plot
0.17
0.14
0.10
Density/Proportion
0.07
0.03
0.00
0.13 0.63 1.13 1.63 2.13 2.63 3.13
Interval Midpoint
13 intervals of w idth 0.25 1 - Weibull
Figure 2.3. Density-Histogram Plot for the fitted Weibull distribution and the
service-time data.
26
Steps for Action E:
At window: Do:
Graphical-Comparisons
Options Select Distribution-Function-Differences Plot.
Click on Apply.
Distribution-Function-
Differences Plot Click on Done.
Graphical-Comparisons
Options Click on Done.
in Figure 2.4. The plot shows the differences between the Weibull distribution function
and the sample distribution function over the range of the data. [The sample
distribution function, which is an estimate of the true underlying distribution function of
the data, is defined at a particular value of x as (approximately) the proportion of
observations in the sample that is less than or equal to x.] Since the vertical differences
in the plot are close to 0, this is further indication that the Weibull distribution is a good
model for the data.
27
Distribution-Function-Differences Plot
0.20
0.13
0.07
Difference (Proportion)
0.00
-0.07
-0.13
-0.20
0.06 0.50 0.93 1.37 1.81 2.24 2.68 3.11
x
Use caution if plot crosses line 1 - Weibull (mean diff. = 0.00501)
28
Steps for Action F:
At window: Do:
Comparisons tab Click on Goodness-of-Fit Tests.
Options for
Goodness-of-Fit Tests Select Anderson-Darling Test.
Click on Apply.
Options for
Goodness-of-Fit Tests Click on Done.
(the most-powerful test available in ExpertFit) to see formally whether our data could
have been generated from the specified Weibull distribution. (You may want to read
the discussion of goodness-of-fit tests in the Goodness-of-Fit Tests and Their
Interpretation tutorial in the software before proceeding.) We will perform the test at a
level (alpha) of 0.05. Since the Anderson-Darling statistic, 0.205, is less than critical
value, 0.750, we do not reject the Weibull distribution. You should keep in mind that
failure to reject by this test does not necessarily mean that the Weibull distribution is
exactly the distribution that produced the data; this test tends to have low power for
small to moderate sample sizes. (We also performed the Kolmogorov-Smirnov Test
and Chi-Square Test and they also did not reject the Weibull distribution.)
In summary, there is no reason to believe based on the above heuristics and
tests that the Weibull distribution does not provide a good model for the service-time
data.
29
Steps for Action G:
At window: Do:
Comparisons tab Click on Applications tab.
Applications tab Click on Simulation Representation in the
Use a Specified Distribution (Model) section.
Simulation-Representation
Options Select the simulation software of your choice.
Click on Apply.
Simulation-Software
Representation Click on Done.
Simulation-Representation Click on Done.
Options
Applications tab In the File menu, select Close Data Analysis.
G: If you are using ExpertFit in the context of simulation modeling, we now see how to
put the selected Weibull distribution into the proper format for several different
simulation-software products using the Applications tab. In particular, the above
actions show how to represent the Weibull distribution in the software product of your
choice; the actual representations for selected products are shown in Table 2.8.
30
Table 2.8. Simulation-software representations for the Weibull distribution.
31
Example 2.2: Ship-Loading Times
This sample of n = 856 observations corresponds to loading times (in hours) for
an oil tanker. The Data-Summary Table is given in Table 2.9; the positive skewness
and the fact that the mean is larger than the median both suggest the underlying
distribution of the data has a longer right tail than left tail.
A histogram of the data with a lower endpoint of 0, an interval width of 0.1, and
23 intervals is given in Figure 2.5 – the latter two choices were obtained by trial and
error. Note that the histogram is reasonably smooth, skewed to the right, and is
definitely shifted away from the origin.
The process of finding a distribution that is a good representation for the data
once again begins by selecting Automated Fitting. The log-logistic distribution was
found by ExpertFit to provide the best representation (see Table 2.10), with a Relative
Score of 100.00. The Absolute Evaluation for the log-logistic distribution is “Good,”
which indicates that there is no reason for concern. Also, the model mean and sample
mean differ by only 0.34 percent.
32
Histogram
0.25
0.20
0.15
Proportion
0.10
0.05
0.00
0.05 0.35 0.65 0.95 1.25 1.55 1.85 2.15
Interval Midpoint
23 intervals of w idth 0.1
33
Table 2.10. Evaluation of the candidate models.
Evaluation: Good
Suggestion: Additional evaluations using Comparisons Tab might be informative.
See Help for more information.
34
We continue our evaluation of the log-logistic distribution by displaying the
Density-Histogram Plot in Figure 2.6. This plot seems to indicate that the log-logistic
distribution provides a good fit for the ship-loading times.
Density-Histogram Plot
0.27
0.22
0.16
Density/Proportion
0.11
0.05
0.00
0.05 0.35 0.65 0.95 1.25 1.55 1.85 2.15
Interval Midpoint
23 intervals of w idth 0.1 1 - Log-Logistic
Figure 2.6. Density-Histogram Plot for the fitted log-logistic distribution and
the ship-loading times.
35
The Distribution-Function-Differences Plot for the log-logistic distribution is
displayed in Figure 2.7. The small vertical differences (errors) suggest that this model
provides a good fit. The P-P Plot in Figure 2.8 also indicates a good fit, since the plot
is close to the straight line with a slope of 1 and a y-intercept of 0.
Distribution-Function-Differences Plot
0.20
0.13
0.07
Difference (Proportion)
0.00
-0.07
-0.13
-0.20
0.37 0.63 0.89 1.14 1.40 1.66 1.92 2.18
x
Use caution if plot crosses line 1 - Log-Logistic (mean diff. = 0.00517)
36
P-P Plot
1.00
0.80
0.60
Model Value
0.40
0.20
0.00
0.00 0.20 0.40 0.60 0.80 1.00
Sample Value
Range of sample 1 - Log-Logistic (discrepancy = 0.01424)
Figure 2.8. P-P Plot for the fitted log-logistic distribution and the ship-loading
times.
37
Example 2.3: Repair Times for a Machine
This example discusses a data set where no model provides a good fit and
where the use of an empirical distribution is recommended.
This sample of n = 288 observations corresponds to repair times for a machine
used for manufacturing household products. Since the repair times were generally
rounded to the nearest 5 minutes, we converted the observations to real numbers to
allow a greater number of distribution choices. The Data-Summary Table is given in
Table 2.11, from which it appears that the underlying distribution of the data is skewed
to the right.
A histogram of the data with a lower endpoint for the first interval of 0, an interval
width of 11, and 17 intervals is given in Figure 2.9. It is clear that the histogram is
positively skewed and has a long right tail.
Using Automated Fitting, we find that the “best” fitting model is a Pearson type
V distribution, which has a Relative Score of 96.00. However, the Absolute
Evaluation is “Bad.” Finally, there is a significant error in the model mean relative to
the sample mean of 6.74 percent.
38
Histogram
0.40
0.32
0.24
Proportion
0.16
0.08
0.00
5.50 38.50 71.50 104.50 137.50 170.50
Interval Midpoint
17 intervals of w idth 11
39
The poor quality of the Pearson type V representation is confirmed by the
Density-Histogram Plot in Figure 2.10. (You might also try the Frequency-
Comparison Plot.)
Density-Histogram Plot
0.40
0.32
0.24
Density/Proportion
0.16
0.08
0.00
5.50 38.50 71.50 104.50 137.50 170.50
Interval Midpoint
17 intervals of w idth 11 1 - Pearson Type V(E)
Figure 2.10. Density-Histogram Plot for the fitted Pearson type V distribution and
the repair-time data.
40
The Distribution-Function-Differences Plot for the Pearson type V distribution
in Figure 2.11 crosses the blue-dashed rectangle, which strongly indicates that this
distribution is not a good representation for the data. Goodness-of-fit tests are not
applicable to the inverse Gaussian distribution since, for example, the location
parameter was not estimated by the method of maximum likelihood.
Distribution-Function-Differences Plot
0.20
0.13
0.07
Difference (Proportion)
0.00
-0.07
-0.13
-0.20
5.00 30.71 56.43 82.14 107.86 133.57 159.28 185.00
x
Use caution if plot crosses line 1 - Pearson Type V(E) (mean diff. = 0.01272)
41
In summary, none of the fitted continuous distributions appears to provide a
good representation for the repair-time data. (This is not surprising since there are only
32 distinct values in a sample with a range of [5, 185] – no discrete distribution works
either.) Therefore, if we are doing simulation modeling, we must resort to the use of an
empirical distribution. The empirical distribution function based on the distinct sample
values is shown in Figure 2.12. You can employ the (Use an Empirical Distribution)
Simulation Representation option at the Applications tab to put this empirical
distribution into the proper format for your simulation software. Furthermore, you can
use the Copy button to place the simulation-software representation into the Windows
Clipboard.
0.80
0.60
F(x)
0.40
0.20
0.00
5.00 30.71 56.43 82.14 107.86 133.57 159.28 185.00
x
Empirical using 288 sample values
Figure 2.12. Empirical distribution function for the repair-time data based on the
unique data values.
42
Example 2.4: Weekly Product Sales
In this example we illustrate how ExpertFit can be used to analyze an integer data set.
The Data-Summary Table for n = 156 weekly sales of a product over a 3-year period
[see Law (2007, p. 325)] is given in Table 2.12. (Note that values range from 0 to 11.)
A histogram of the sales data, starting at 0 and using 12 intervals that each contain one
value, is given in Figure 2.13. Its shape is similar to the probability mass function of a
geometric distribution.
Using Automated Fitting, ExpertFit found that the geometric distribution with
pˆ (0) = 0.346 provides the best representation, receiving a Relative Score of 83.33 (see
Table 2.13). This score is low because the second-best model is the negative binomial
distribution with sˆ = 1 and pˆ (0) = 0.346 , which is the same as the above geometric
distribution. If the negative binomial distribution is deleted (see View/Delete Models at
the Models tab), then the Relative Score of the geometric distribution is 100.00. The
Absolute Evaluation for the geometric distribution is “Indeterminate,” which indicates
that additional evaluations using the Comparisons Tab are strongly recommended.
43
Histogram
0.38
0.30
0.23
Proportion
0.15
0.08
0.00
0 2 4 6 8 10
Value
12 intervals of w idth 1
44
Table 2.13. Evaluation of the candidate models.
Evaluation: Indeterminate
45
To determine whether the geometric distribution is a good representation for the
sales data, we present a Frequency-Comparison Plot for the geometric distribution in
Figure 2.14. The agreement between the histogram and the expected proportions for
the geometric distribution is good except possibly for the interval corresponding to x = 1.
Frequency-Comparison Plot
0.38
0.30
0.23
Proportion
0.15
0.08
0.00
0 2 4 6 8 10
Value
12 intervals of w idth 1 1 - Geometric
Figure 2.14. Frequency-Comparison Plot for the fitted geometric distribution and
the sales data.
46
A Distribution-Function-Differences Plot and P-P Plot for the geometric
distribution are given in Figures 2.15 and 2.16, respectively. Neither of these plots
gives us any particular reason to think that the geometric distribution is not a good
representation for the sales data.
Distribution-Function-Differences Plot
0.20
0.13
0.07
Difference (Proportion)
0.00
-0.07
-0.13
-0.20
0 2 4 6 8 10
x
Use caution if plot crosses line 1 - Geometric (mean diff. = 0.00884)
47
P-P Plot
1.00
0.80
0.60
Model Value
0.40
0.20
0.00
0.00 0.20 0.40 0.60 0.80 1.00
Sample Value
Range of sample 1 - Geometric (discrepancy = 0.04460)
Figure 2.16. P-P Plot for the fitted geometric distribution and the sales data.
48
We conclude this example by performing an Equal-Width Chi-Square Test for
the geometric distribution, since equal-probable intervals are not available for discrete
distributions. It is recommended for discrete distributions that the intervals be chosen
so that the probabilities (expected numbers or counts) under the hypothesized model
are approximately equal for all intervals. One way to do this is to note that the mode
(most-likely value) of the geometric distribution is 0; furthermore, pˆ (0) = 0.346 . The
large value for the mode limits our choice of intervals and we end up with the three
intervals given in Table 2.14, where most of the calculations for the chi-square test are
also presented. (These intervals were obtained using the View/Group Cells button.)
Note that the expected count for each interval is at least 5, as is recommended. Since
the chi-square statistic value of 1.930 is less than the critical value of 5.991
corresponding to a level of 0.05 and 2 degrees of freedom, we do not reject the
geometric distribution at level 0.05.
In summary we have no reason to believe that the geometric distribution is not a
good model for our data.
49
3. Data Analysis Module – Advanced Mode
In this chapter we discuss Advanced Mode for the Data Analysis module,
which is accessed from the Mode pull-down menu at the top of the screen. Advanced
Mode contains a large number of features that are not in Standard Mode – these
features will be of interest to the sophisticated user. However, a user can switch from
one mode to another at any time.
The use of Advanced Mode for the Data Analysis module to determine what
probability distribution best represents a data set is based on the sequential application
of the four tabs shown in Table 3.1. (These tabs are similar to those in Table 2.1.)
Data Used to read in a data set from a file, enter a data set at the
keyboard, or paste in a sample from the Clipboard
The options available in these four tabs are shown in Tables 3.2 through 3.5,
respectively.
50
Table 3.2. Options for the Data tab.
51
Table 3.3. Options for the Models tab.
Fit a Class of Models Automatic fitting of all models in a particular class (i.e.,
non-negative continuous, bounded continuous, and
unbounded continuous) – parameters are typically
estimated from the data
52
Table 3.4. Options for the Comparisons tab.
53
Table 3.5. Options for the Applications tab.
54
Although there are different ways that these four tabs could be used to
determine the best distribution for a data set, the following are the explicit steps that we
recommend (same as for Standard Mode):
1. Obtain a data set using the Data tab – see Section 1.2 for a discussion of the type
and amount of data required.
2. View the resulting Data-Summary Table (at the Data tab) – provides information on
the center, shape, and range of the true density (mass) function.
3. Make a histogram of your data (used in Step 5) using the Data tab – see the
Constructing a Histogram from Your Data tutorial in the online help.
4. Select the distribution that is the best representation for your data using the
Automated Fitting option at the Models tab.
5. Confirm using the Comparisons tab that the best distribution as determined by
ExpertFit is, in fact, satisfactory in an absolute sense – see Section 2.1 for
recommendations.
6. If you are doing simulation modeling, then either represent the best-fitting
distribution (if good in an absolute sense) or an empirical distribution based on your
data (if the best distribution is not satisfactory) in your simulation software using the
Applications tab.
An example of the use of Advanced Mode for the Data Analysis module is
given in Section 3.1.
55
3.1. Example
We now present an example of the use of Advanced Mode for the Data
Analysis module, following the six-step approach outlined at the beginning of this
chapter. We use Normal Precision in fitting distributions to the data.
56
Example 3.1: Testing the Homogeneity of Two or More Data Sets
57
Frequency-Comparison Plot
0.22
0.18
0.13
Proportion
0.09
0.04
0.00
12.38 21.88 31.38 40.88 50.38 59.88 69.38
Interval Midpoint
Example 3.1-1: Processing Times 1 Example 3.1-2: Processing Times 2
58
Distribution Function Plot
1.00
0.80
0.60
F(x)
0.40
0.20
0.00
17.00 24.83 32.66 40.49 48.31 56.14 63.97 71.80
x
Example 3.1-1: Processing Times 1 Example 3.1-2: Processing Times 2
Figure 3.2. Distribution Function Plot for the two data sets.
59
Box-Plot Comparisons
60
We now use Automated Fitting (with Normal Precision) to determine what
distribution best represents the Merged data set, which consists of 1748 observations.
It turns out that the Pearson type V distribution provides the best fit with a Relative
Score of 97.92. The Absolute Evaluation is “Indeterminate,” which means that more
evaluations need to be performed at the Comparisons tab before the quality of the
representation provided by the Pearson type V distribution can be determined. The
Anderson-Darling test says to reject the Pearson type V distribution at level 0.05;
however, for this large sample size, the test is very powerful and may reject a
hypothesized distribution whose error may be practically insignificant. In fact, the
Density-Histogram Plot, the Distribution Function Plot, the Distribution-Function-
Differences Plot, and the P-P Plot all indicate that the Pearson type V distribution
provides a reasonably good representation of the data.
61
4. Task-Time Models Module
In some simulation studies it is not possible to obtain good data on the random
variables of interest, so the usual statistical techniques are not applicable to the
problem of selecting corresponding probability distributions. For example, if the system
being studied does not currently exist in some form, collecting data from the system is
obviously not possible. This difficulty can also occur for existing systems, if the number
of required probability distributions is large and the time available for the simulation
study prohibits the necessary data collection and analysis. In addition, sometimes there
are data available from an existing system, but the data are not in a format suitable for
use in a simulation model (e.g., the data were collected by an automated system). For
such situations ExpertFit provides guidance on modeling a task time in the Task-Time
Models module. Note that the use of no-data models is not a substitute for a careful
analysis of data collected from your system, if this is possible.
Consider the continuous random variable corresponding to the time to complete
some task (e.g., a machine repair time or a customer service time in a bank). ExpertFit
allows you to model such a random variable by a Weibull, lognormal, or triangular
distribution. For a particular distribution, you must give subjective estimates of the
minimum task time, the most-likely task time (the mode), and the 100pth percentile of
the task time. Allowable percentiles for the Weibull and lognormal distributions are the
90th (the default), 95th, and 99th; the 100th percentile (the maximum value) is also
available for the triangular distribution. More information on the use of the Task-Time
Models module can be found in the Modeling Task Times in the Absence of Data
tutorial, which is accessed from the Help pull-down menu in ExpertFit.
After completely specifying a task-time model, you can display characteristics of
the model such as its density function or percentiles. You can also represent the task-
time model in the simulation-software product of your choice.
62
4.1. Organization and Options
A Task-Time Model is based on the sequential use of the two tabs given in
Table 4.1.
The options for the Models and Applications tabs are given in Tables 4.2 and
4.3, respectively.
Specify a Model Used to create new (or to modify existing) models for a
task time
63
Table 4.3. Options for the Applications tab.
64
4.2. Example
65
Example 4.1: Modeling a Task Time
Click on Done.
66
A: Suppose it is thought that the minimum and maximum times to perform some task
are 1 and 11 minutes. Furthermore, suppose that the most-likely time to perform the
task is believed to be 4 minutes. Then the density function of the ExpertFit-specified
triangular distribution is given in Figure 4.1. Suppose that we want to know the 95th
percentile of our model. Using the Percentile Table button in the Characteristics
option, we get the percentile table shown in Table 4.4. From this table we get that 95
percent of the time the task-time random variable will take on values less than or equal
to 9.129 minutes.
The use of a triangular distribution is a simple approach for modeling a task time
in the absence of data, and it is usually possible to get estimates of the three
parameters. However, the triangular distribution does not have a very flexible shape
[see Law (2007 p. 370)]. Therefore, one could use a Weibull distribution or a
lognormal distribution instead, which are also supported by ExpertFit.
67
Density Function Plot
0.20
0.16
0.12
f(x)
0.08
0.04
0.00
1.00 2.43 3.86 5.29 6.71 8.14 9.57 11.00
x
1 - Triangular
68
Steps for Action A (continued):
At window: Do:
Characteristics Options Select Percentile Table.
Click on Apply.
69
Table 4.4. Percentiles of the specified triangular distribution.
Percent 1 - Triangular
0.0 1.00000
0.1 1.17321
0.5 1.38730
1.0 1.54772
2.5 1.86603
5.0 2.22474
10.0 2.73205
25.0 3.73861
50.0 5.08392
75.0 6.81670
90.0 8.35425
95.0 9.12917
97.5 9.67712
99.0 10.16334
99.5 10.40389
99.9 10.73542
100.0 11.00000
70
5. Machine-Breakdown Models Module
U1 D1 U2 D2
Time
0
End of cycle End of cycle
71
The busy time before failure of the machine, B, is assumed to have a gamma
distribution with a shape parameter α equal to 0.7 and a scale parameter βB to be
specified, as shown in Figure 5.2. We chose the gamma distribution because of its
flexibility (i.e., its density can assume a wide variety of shapes) and because it has the
general shape of many busy-time histograms when α is less than or equal to 1. The
particular shape parameter of 0.7 for the gamma distribution was determined by fitting a
gamma distribution to a number of different sets of busy-time data, with 0.7 being the
average shape parameter obtained.
Note that busy time for a machine is only accumulated when the machine is
doing productive work, not when it is blocked or starved. For example, suppose that
the first busy time generated from the gamma distribution is 60.7 minutes and that each
part takes exactly 1 minute to be processed. Then the machine fails while processing
its 61st part. However, the simulation clock might be somewhat larger than 60.7 when
the machine fails, due to starving or blocking for the machine.
At the instant the machine fails, we assume that the downtime of the machine, D,
begins. The downtime of the machine is assumed to have a gamma distribution with a
shape parameter α equal to 1.3 and a scale parameter βD to be specified, as shown in
Figure 5.3. This particular shape parameter was determined by fitting a gamma
distribution to a number of different sets of downtime data, with 1.3 being the average
shape parameter obtained.
In order to determine the values of the scale parameters βB and βD, ExpertFit
asks you to give subjective estimates for two of the following three basic machine
characteristics:
• Machine efficiency
• Mean downtime for the machine
• Mean number of downs in some time period (e.g., in an 8-hour shift)
The efficiency of a machine is defined to be the long-run proportion of potential
processing time (i.e., parts present and machine not blocked) during which the machine
is actually processing parts. If the machine is never starved or blocked, then the
efficiency is the long-run proportion of time that the machine is processing parts.
72
Density Function Plot
15.50
12.91
10.33
7.75
f(x)
5.17
2.58
0.00
0.00 0.87 1.73 2.60 3.46 4.33 5.20 6.06
x
73
Density Function Plot
0.58
0.48
0.38
0.29
f(x)
0.19
0.10
0.00
0.00 1.09 2.19 3.28 4.38 5.47 6.57 7.66
x
74
The mean downtime of a machine is the mean amount of time that elapses from
the instant the machine breaks down until the instant that it is repaired; it includes both
the time spent waiting for a repairman (if any) and the repair time itself.
The mean number of downs in some time period (called the Time Frame in
ExpertFit) such as a shift (possibly, non integral) is, more specifically, the mean number
of busy-time/downtime cycles in a time period. For example, if the mean number of
downs is exactly 2, then the machine fails and is subsequently repaired an average of 2
times in a time period.
If the machine is subject to significant starving or blocking, then you must also
give the mean number of parts produced per time period and the mean part-processing
time in order for ExpertFit to compute βB and βD.
The default gamma distributions used for busy time and downtime have location
parameters of zero; thus, they can take on arbitrarily small positive values. However, in
practice one or both of these distributions might have a minimum possible value (i.e., its
location parameter) that is a positive number. For example, it might be known that the
minimum possible downtime is 10 minutes. Thus, ExpertFit allows you to specify a
positive value for the minimum possible downtime or for the minimum possible busy
time.
75
5.1. Organization and Options
Models Used to construct models for the busy-time and downtime distributions
The options for the Models and Applications tabs are given in Tables 5.2 and
5.3, respectively.
Specify a Model Used to create models for the busy-time and downtime
distributions
76
Table 5.3. Options for the Applications tab.
Time-Frame Report Displays the expected total time during a time frame that
the machine will be busy, down, and either blocked or
starved
77
5.2. Examples
78
Example 5.1: Modeling Machine Downtimes in the Absence of Data
At window: Do:
Project 1 Click on New.
Project-Element Editing Select Construct distributions in the absence of
data.
Select Machine-Breakdown Models.
In Project-Element Name edit box, enter Example
5.1.
Click on OK.
Project 1 Click on Analyze.
Models tab Click on Specify a Model.
Specify/Modify
Machine-Breakdown Model Click on Create a New Model.
Assumptions about
the Machine Specify the Machine efficiency to be 0.9.
Specify the Mean downtime to be 60.0.
Click on Additional Machine Characteristics tab.
Assumptions about
the Machine Specify the Minimum downtime to be 10.0.
Click on OK.
Specify/Modify
Machine-Breakdown Model Click on Done.
79
A: Consider a machine that is never starved or blocked. Suppose that the machine
has an efficiency of 0.9; that is, it is actually producing parts 90 percent of the time.
When the machine goes down, the mean downtime is 60 minutes. However, the
minimum possible downtime is 10 minutes. These characteristics are entered using
the commands on the previous page. Note that the default values for the Blocking
and/or Starving are Significant checkbox, Characteristics to be Entered, and
Time Unit are correct. Time Frame is not used in this example.
The machine busy-time and downtime distributions have now been completely
specified, and all of the specified and calculated (shown in blue) machine
characteristics are shown on the Specify/Modify Machine-Breakdown Model screen.
In particular, note that the mean number of downs (actually the mean number of busy-
time/downtime cycles) per 8-hour shift has been calculated to be 0.8. This makes
sense since the mean length of a busy-time/downtime cycle is 10 hours.
80
Steps for Action B:
At window: Do:
Models tab Click on Applications tab.
Applications tab Click on Time-Frame Report.
Specify Units and Model for
Time-Frame Report Click on Apply.
Click on Done.
Specify Units and Model for
Time-Frame Report Click on Done.
Applications tab Click on Models tab.
(This step anticipates proceeding to Example 5.2.)
distributions. For example, a machine Time-Frame Report is given in Table 5.4. Note
from this report that the machine is expected to be busy 90 percent of the time, which is
another way of saying that its efficiency is 0.9 (also see Table 5.5).
81
Table 5.4. Machine Time-Frame Report for the specified busy-time and downtime
models.
82
Example 5.2: Continuation of Previous Example
At window: Do:
Models tab Click on Specify a Model.
Specify/Modify
Machine-Breakdown Model Click on Create a New Model.
Assumptions about Click on the Blocking and/or Starving are
the Machine Significant checkbox.
Specify the Machine efficiency to be 0.9.
Specify the Mean downtime to be 60.0.
Click on Part Production Characteristics tab.
Assumptions about
the Machine Specify the Mean number of parts produced
to be 100.
Specify the Mean part-processing time to be 4.0.
Click on Additional Machine Characteristics tab.
Assumptions about
the Machine Specify the Minimum downtime to be 10.0.
Click on OK.
Specify/Modify
Machine-Breakdown Model Click on Done.
83
A: Suppose for the machine of Example 5.1 that blocking/starving is now significant.
For example raw materials might arrive to the machine on an intermittent basis.
Suppose also that the mean number of parts produced per 8-hour shift (the default
value of the Time Frame) is 100 and the mean part-processing time is 4 minutes.
We set the Blocking and/or Starving are Significant checkbox to “on” at the Basic
Machine Characteristics tab, in addition to specifying the Machine Efficiency and
Mean Downtime as in Example 5.1. Because blocking/starving is now significant, we
must specify values for Mean number of parts produced and Mean part-processing
time at the Part Production Characteristics tab. Finally, the Minimum downtime is
specified at the Additional Machine Characteristics tab.
84
Steps for Action B:
At window: Do:
Models tab Click on Applications tab.
Applications tab Click on Time-Frame Report.
Specify Units and Model for
Time-Frame Report Select model 2 in the Model for Time-Frame Report
scroll list.
Click on Apply.
Click on Done.
Specify Units and Model for
Time-Frame Report Click on Done.
Application tab In the File menu, select Close Machine-Breakdown
Models.
B: In Table 5.5 we display a machine Time-Frame Report, from which we can see
that the machine is expected to be busy 83.33 percent of time. Recall in Example 5.1
(see Table 5.4) that the machine was expected to be busy 90 percent of the time. Also
the Mean number of downs per 8-hour shift is different for the two examples. Thus, the
simulation results for the two examples will be different provided that the machine does,
indeed, experience blocking/starving in the second example.
85
Table 5.5. Machine Time-Frame Report for the specified busy-time and downtime
models.
86
6. Distribution Viewer
87
7. Batch Mode
88
References
Evans, M., N. Hastings, and B. Peacock, Statistical Distributions, Third Edition, John
Wiley, New York (2000).
Johnson, N. L., S. Kotz, and N. Balakrishnan, Continuous Univariate Distributions,
Volume 1, Second Edition, Houghton Mifflin, Boston (1994).
Johnson, N. L., S. Kotz, and N. Balakrishnan, Continuous Univariate Distributions,
Volume 2, Second Edition, Houghton Mifflin, Boston (1995).
Johnson, N. L., S. Kotz, and A.W. Kemp, Univariate Discrete Distributions, Second
Edition, Houghton Mifflin, Boston (1992).
Law, A. M., Simulation Modeling and Analysis, Fourth Edition, McGraw-Hill, New York
(2007).
89
Appendix A. Distributions Included in ExpertFit
In this appendix we present important information on the thirty-two continuous and eight
discrete standard distributions available in ExpertFit. (ExpertFit also supports three types of
empirical distributions.) These distributions are organized according to the following categories:
90
Category Distribution Notation Page
Bounded Beta beta(a, b, α1,α2) 94
Continuous Johnson SB JSB(a, b, α1,α2) 111
Power Function power(a, b, α) 125
Triangular triang(a, b, m) 129
Uniform U(a, b) 130
Discrete Bernoulli Bernoulli(p) 93
Binomial bin(t, p) 95
Discrete Uniform DU(i, j) 98
Geometric geom(p) 107
Hypergeometric hyper(N, n, s) 108
Logarithmic Series logseries(α) 114
Negative Binomial negbin(s, p) 119
Poisson Poisson(λ) 124
A continuous random variable can take on any value in some interval of the real line [e.g.,
(0, ∞)]. A non-negative continuous distribution restricts the random variable to be strictly larger
than a specified lower-bound value. A bounded continuous distribution restricts the random
variable to be strictly larger than a specified lower-bound value and strictly smaller than a
specified upper-bound value. An unbounded continuous distribution places no restrictions on the
values of the random variable. A discrete random variable can take on some subset of the non-
negative integers; the subset depends upon the specific distribution.
91
The following table describes the special symbols used in this Appendix:
∫t
z −1
Γ( z ) = e −tdt fo r z > 0
0
Γ(u ) Γ( v )
N o te th a t B (u , v ) = B (v , u ) = .
Γ(u + v )
IID Independent,
identically distributed
∼ Is distributed as
∈ Is contained in
92
Bernoulli - Bernoulli(p) Discrete
1 − p if x = 0
Mass p (x) = p if x = 1
0
otherwise
Parameter p ∈ ( 0 ,1)
Range {0 , 1}
Mean p
Variance p (1 − p )
0 if p < 0.5
Mode 0 and 1 if p = 0.5
1 if p > 0.5
Comment:
1. The Bernoulli(p) and bin(1, p) distributions are the same.
93
Beta – beta(a, b,α1, α2) Bounded Continuous
x − a α1 −1 b − x α2 −1
Density f (x) = b − a b − a if a < x < b
(b − a ) B ( α1 ,α 2 )
0
otherwise
Range (a , b)
α1
Mean a + (b − a)
α1 + α2
α1α2
Variance (b − a ) 2
(α1 + α2 ) 2 (α1 + α2 + 1)
α1 − 1
a + (b − a ) α + α − 2 if α1 > 1,α 2 > 1
1 2
Comments:
1. The beta(a, b,1,1) and U(a, b) distributions are the same.
2. If X 1 and X 2 are independent random variables with X i ∼ gamma(0, β , α i ) , then
Y = X 1 /( X 1 + X 2 ) ∼ beta(0,1, α1 , α 2 ) .
94
Binomial - bin(t, p) Discrete
t x t−x
p (1 − p ) if x ∈ {0, 1,...,t}
Mass p(x) = x
0
otherwise
Range {0,1, … , t )
Mean tp
Variance tp (1 − p )
p ( t + 1) − 1 and p ( t + 1) if p ( t + 1) is an integer
Mode
p ( t + 1) otherwise
Comments:
1. If Y1 , Y2 , …, Yt are independent Bernoulli(p) random variables, then
Z = Y1 + Y2 + + Yt ∼ bin(t , p ) .
2. The bin(1, p) and Bernoulli(p) distributions are the same.
3. If X 1 , X 2 , … , X m are independent random variables with X i ∼ bin(ti , p ) , then
Y = X 1 + X 2 + + X m ∼ bin(t1 + t2 + + tm , p ) .
4. The bin(t , p ) mass function is symmetric if and only if p = 1/ 2 .
95
Cauchy - Cauchy(γ, β) Unbounded Continuous
−1
x − γ
2
96
Chi-Square - chisq(γ, ν ) Non-negative Continuous
( x − γ )(ν − 2) / 2 − ( x −γ ) / 2
ν /2 e if x > γ
Density f (x) = 2 Γ(ν / 2)
0
otherwise
Range (γ , ∞ )
Mean γ +ν
Variance 2ν
γ + ν − 2 if ν > 2
Mode
γ otherwise
Comment:
1. The chisq(0,ν ) and gamma(0, 2,ν / 2) distributions are the same.
97
Discrete Uniform - DU(i, j) Discrete
1
if x ∈ {i,i + 1,..., j}
Mass p (x) = j − i + 1
0 otherwise
Parameters i and j integers with i ≤ j ; i is a location parameter
and j − i is a scale parameter
Range {i, i + 1, …, j}
i+ j
Mean
2
(j − i + 1) 2 − 1
Variance
12
Comment:
1. The DU(0, 1) and the Bernoulli(1/2) distributions are the same.
98
Erlang - m-Erlang(β)
or Erlang(γ, β, m) Non-negative Continuous
(x − γ ) m −1 −(x − γ )
m
f (x) = β (m − 1)!
exp if x > γ
Density β
0
otherwise
Range (γ , ∞ )
Mean γ + mβ
Variance mβ 2
Mode γ + β ( m − 1)
Comments:
1. When γ = 0, the notation m-Erlang(β) is typically used.
2. The expo(γ, β) and Erlang(γ, β, 1) distributions are the same.
3. The Erlang(γ, β, m) and gamma(γ, β, m) distributions are the same.
99
Error - error(h) Unbounded continuous
h 2
Density f (x) = e − (xh ) for all real numbers x
π
Range ( − ∞ ,∞)
Mean 0
1
Variance
2h 2
Mode 0
Comment:
−1
1. The error(h) and N(0, h 2 ) distributions are the same.
100
Exponential - expo(γ, β) Non-negative Continuous
1 −(x − γ )
exp if x > γ
Density f (x) = β β
0
otherwise
Range (γ , ∞ )
Mean γ +β
Variance β2
Mode γ
Comments:
1. The expo(γ, β) distribution is a special case of both the gamma and Weibull
distributions (for shape parameter α = 1, scale parameter β, and location
parameter γ in both cases).
2. If X 1 , X 2 , …, X m are independent expo(0, β) random variables, then
Y = X 1 + X 2 + + X m ∼ gamma(0, β , m) , also called the m-Erlang(β)
distribution.
101
Exponential Power - expower(γ, β, α) Unbounded Continuous
x − γ 2 /α
exp −
2β
Density f (x) = for all real numbers x
α /2 α
2(2 β ) Γ 1 +
2
Range ( − ∞, ∞ )
Mean γ
3α
2α β 2 Γ
Variance 2
α
Γ
2
Mode γ
Comments:
1. The expower(γ, β, 1) and N(γ , β ) distributions are the same.
102
Extreme Value Type A - EVTA(γ , β ) Unbounded Continuous
Density f ( x ) = (1/ β ) exp[( x − γ ) / β ]exp{− exp[( x − γ ) / β ]} for all real numbers x
Range ( − ∞, ∞ )
Mean γ − 0.57722 β
Variance 1.64493β 2
Mode γ
Comments:
1. X ∼ Weibull(0, β , α ) if and only if Y = ln X ∼ EVTA(ln β ,1/ α )
2. X ∼ EVTA(γ , β ) if and only if Y = β exp[( X − γ ) / β ] ∼ expo(0, β )
3. X ∼ EVTA(γ , β ) if and only if Y = − X has an extreme value type B distribution with
parameters −γ and β , denoted EVTB( −γ , β )
103
Extreme Value Type B - EVTB(γ , β ) Unbounded Continuous
Density f ( x ) = (1/ β ) exp[ −( x − γ ) / β ]exp{− exp[ −( x − γ ) / β ]} for all real numbers x
Range ( − ∞, ∞ )
Mean γ + 0.57722 β
Variance 1.64493β 2
Mode γ
Comments:
1. X ∼ Weibull(0, β , α ) if and only if Y = − ln X ∼ EVTB( − ln β ,1/ α )
2. X ∼ EVTB(γ , β ) if and only if Y = β exp[− ( X − γ ) / β ] ∼ expo(0, β )
3. X ∼ EVTB(γ , β ) if and only if Y = − X ∼ EVTA(−γ , β )
104
F - F(ν 1 ,ν 2 ) Non-negative Continuous
ν 1ν1 / 2ν ν2 2 / 2 x (ν1 − 2) / 2
if x > 0
Density f ( x) = B (ν 1 / 2,ν 2 / 2)(ν 2 +ν 1 x)(ν1 +ν 2 ) / 2
0
otherwise
Range (0, ∞ )
ν2
Mean for ν 2 > 2
ν2 − 2
2ν 22 (ν 1 + ν 2 − 2)
Variance for ν 2 > 4
ν 1 (ν 2 − 2) 2 (ν 2 − 4)
ν 2 (ν 1 − 2)
for ν 1 > 2
Mode ν 1 (ν 2 + 2)
does not exist on (0, ∞) otherwise
Comment:
1. The F (ν 1 ,ν 2 ) and PT6(0,ν 2 /ν 1 ,ν 1 / 2,ν 2 / 2) distributions are the same.
105
Gamma - gamma(γ, β, α) Non-negative Continuous
(x − γ )α −1 − (x − γ )
α exp if x > γ
Density f (x) = β Γ(α ) β
0
otherwise
Range (γ , ∞ )
Mean γ + αβ
Variance αβ 2
γ + β (α − 1) if α ≥ 1
Mode
γ otherwise
Comments:
1. The expo(γ, β) and gamma(γ, β, 1) distributions are the same.
2. For a positive integer m, the gamma(0, β, m) distribution is called the m-Erlang(β)
distribution, and the gamma(γ, β, m) and the Erlang(γ, β, m) distributions are the same.
3. If X 1 and X 2 are independent random variables with X i ∼ gamma(0, β , α i ) , then
Y = X 1 /( X 1 + X 2 ) ∼ beta(0,1, α1 , α 2 ) .
4. X ∼ gamma(γ , β , α ) if and only if Y = 1/( X − γ ) has a Pearson type V distribution with
location parameter 0, scale parameter 1/β , and shape parameter α, denoted PT5(0,1/ β , α )
5. The chisq(0,ν ) and gamma(0, 2,ν / 2) distributions are the same for ν ∈ {1, 2, …} .
6. If X 1 , X 2 , …, X m are independent random variables with X i ∼ gamma(0, β , α i ) , then
Y = X 1 + X 2 + + X m ∼ gamma(0, β , α1 + α 2 + + α m ) .
106
Geometric - geom(p) Discrete
p(1 − p) x if x ∈ {0,1,...}
Mass p(x) =
0 otherwise
Parameter p ∈ (0,1)
Range {0,1, …}
1− p
Mean
p
1− p
Variance
p2
Mode 0
Comments:
1. If Y1 , Y2 , … is a sequence of independent Bernoulli(p) random variables and
X = min{i : Yi = 1} − 1 , then X ∼ geom( p ) .
107
Hypergeometric - hyper( N , n, s ) Discrete
s N − s
x n−x
Mass p ( x) =
N
n
ns
Mean
N
( ns / N )(1 − s / N )( N − n)
Variance
N −1
Comment:
1. Suppose that there is a population of N items of which s are successes (i.e., have a certain
property). If we select n of these items without replacement, then the number of successes
obtained has a hypergeometric distribution with parameters N, n, and s.
108
Inverse Gaussian - IG(γ, β, α) Non-negative Continuous
1
α 2 −α (x − γ − β ) 2
Density f (x) = 2π (x − γ )3 exp 2 if x > γ
2β (x − γ )
0 otherwise
Range (γ , ∞)
Mean γ +β
β3
Variance
α
Mode γ + β ( 1 + θ 2 − θ ) , where θ = 3β / 2α
Comments:
1. The parameter β has elements of a shape parameter since it affects the skewness
and kurtosis.
2. The IG(γ ,1, α ) and Wald(γ , α ) distributions are the same.
109
Inverted Weibull - IW(γ, β, α) Non-negative Continuous
α −α −1 x − γ
−α
Range (γ , ∞)
γ + β Γ 1 −
1
Mean for α > 1
α
2 1 2
2
Variance β Γ 1 − − Γ 1 − for α > 2
α α
γ + β
α α
Mode
α +1
Comments:
1. X ∼ IW(0, β , α ) if and only if Y = X −1 ∼ Weibull(0,1/β , α )
110
Johnson SB - JSB(a, b, α1, α2) Bounded Continuous
α 2 (b − a ) 1
2
exp − α1 + α 2 ln x − a if a < x < b
Density f (x ) = (x − a)(b − x ) 2π 2 b − x
0 otherwise
Range (a, b)
1 1 − 2α 22
Mode Bimodal when α 2 <
2
and α1 <
α2
− 2α 2 tanh -1 ( )
1 − 2α 22 ;
otherwise unimodal.
Comments:
1. X ∼ JSB( a, b, α1 , α 2 ) if and only if Z = α1 + α 2 ln
X −a
∼ N(0 ,1)
b− X
2. The density function is (skewed left, symmetric, skewed right) as the shape parameter
α1 is ( > 0, = 0, < 0) .
3. lim f ( x) = lim f ( x ) = 0 for all values of α1 and α 2
x→ a x →b
111
Johnson SU - JSU(γ, β, α1, α2) Unbounded Continuous
2
α2 1 x −γ x − γ
2
f (x ) = exp − α1 + α 2 ln + + 1
Density
2π (x − γ ) 2 + β 2 2
β β
for all real numbers x
Range ( −∞, ∞)
1 α1
Mean γ − β exp 2 sinh
2α 2 α2
Mode (
γ + β y , where y satisfies y + α1α 2 y 2 + 1 + α 22 y 2 + 1 ln y + y 2 + 1 = 0 )
Comments:
X −γ X −γ
2
1. X ∼ JSU(γ , β , α1 , α 2 ) if and only if Z = α1 + α 2 ln + + 1 ∼ N(0 ,1)
β β
2. The density function is (skewed left, symmetric, skewed right) as the shape parameter
α1 is ( > 0, = 0, < 0) .
112
Laplace - Laplace (γ, β) Unbounded Continuous
1 x −γ
Density f ( x) = exp − for all real numbers x
2β β
Range ( − ∞, ∞ )
Mean γ
Variance 2β 2
Mode γ
Comments:
1. X ∼ Laplace(γ , β ) if and only if Y = e X has the log-Laplace distribution with location
parameter 0, scale parameter eγ , and shape parameter 1/ β , denoted LP(0, eγ ,1/ β )
2. The Laplace distribution is also called the double exponential distribution.
3. The Laplace(γ , β ) and expower(γ , 2 β , 2) distributions are the same.
113
Logarithmic series - logseries(α) Discrete
θα x
if x ∈ {1, 2, …}, where θ = −1/ log(1 − α )
Mass p (x) = x
0
otherwise
Range {1, 2, …}
θα
Mean
1− α
θα (1 − θα )
Variance
(1 − α ) 2
Mode 1
114
Logistic - logistic(γ, β) Unbounded Continuous
exp[ − (x − γ ) / β ]
Density f (x ) = for all real numbers x
β {1 + exp[− ( x − γ ) / β ]}2
Range ( − ∞ ,∞ )
Mean γ
β 2π 2
Variance
3
Mode γ
Comments:
1. X ∼ logistic(γ , β ) if and only if Y = e X is distributed as a log-logistic distribution with
location parameter 0, scale parameter eγ , and shape parameter 1/ β , denoted LL(0, eγ ,1/ β )
2. The logistic distribution has greater “tail weight” than the normal distribution.
115
Log-Laplace - LP(γ, β, α) Non-negative Continuous
α α −1
X −γ
β if β > x > γ
2β
−α −1
Density f (x) = α X −γ
β if x ≥ β
2β
0 otherwise
Range (γ , ∞)
βα 2
Mean γ+ 2 for α > 1
α −1
β 2α 2 (2α 2 + 1 )
Variance for α > 2
(α 2 − 1)2 (α 2 − 4)
Comment:
1. X ∼ LP(γ , β , α ) if and only if Y = ln( X − γ ) has the Laplace distribution with location
parameter ln β and scale parameter 1/ α , denoted Laplace(ln β ,1/ α )
116
Log-logistic - LL(γ, β, α) Non-negative Continuous
X −γ
α −1
α
β if x > γ
2
Density f (x) = X − γ α
β 1 +
β
0 otherwise
Range (γ , ∞)
π
Mean γ + βθ cosecant(θ ) for α > 1, where θ =
α
π
Variance β 2θ {2cosecant(2θ ) − θ [cosecant(θ )]2 } for α > 2, where θ =
α
1
γ + β α − 1 α
if α > 1
Mode α +1
0 otherwise
Comment:
1. X ∼ LL(0, β , α ) if and only if Y = ln X has the logistic distribution with location
parameter ln β and scale parameter 1/ α , denoted logistic(ln β ,1/ α )
117
Lognormal - LN(γ, β, α) Non-negative Continuous
1 − [ln(x − γ ) − β ]2
exp if x > γ
Density f (x) = (x − γ ) 2πα 2 2α 2
0
otherwise
Range (γ , ∞ )
α2
Mean γ + exp β +
2
Mode γ + exp ( β − α 2 )
Comments:
1. X ∼ LN(γ , β ,α ) if and only if Y = ln(X − γ ) ∼ N(lnβ ,α )
2. lim f (x) = 0, regardless of the parameter values
x →γ
118
Negative Binomial - negbin(s, p) Discrete
s + x − 1 s
p (1 − p) x if x ∈ {0,1,…}
Mass p(x) = x
0
otherwise
Range {0,1, …}
s (1 − p )
Mean
p
s (1 − p )
Variance
p2
s (1 − p ) − 1
Mode Let y = .
p
Then the mode is y and y + 1 if y is an integer; it is y + 1 otherwise.
Comments:
1. If Y1 , Y2 , …, Ys are independent geom(p) random variables, then
Z = Y1 + Y2 + + Ys ∼ negbin( s, p ) .
119
Normal - N(γ, β) Unbounded Continuous
1 − (x − γ ) 2
Density f (x) = exp 2 for all real numbers x
2πβ 2 2β
Range ( − ∞ ,∞ )
Mean γ
Variance β2
Mode γ
Comments:
1. The N(0,1) distribution is called the standard normal distribution.
2. If X 1 , X 2 , …, X k are independent standard normal random variables, then
Y = X 12 + X 22 + + X k2 has a chi-square distribution with k degrees of freedom, denoted
chisq(k).
3. X ∼ N(γ , β ) if and only if Y = e X ∼ LN(0 ,eγ , β )
120
Pareto - Pareto(γ, β) Non-negative Continuous
β γ β +1
if x > γ
Density f (x) = γ x
0
otherwise
Range (γ ,∞ )
βγ
Mean for β > 1
β −1
βγ 2
Variance for β > 2
( β − 1)2 ( β − 2 )
Mode γ
Comments:
1. The location parameter γ must be strictly positive and, thus, all data values must be as
well.
2. The parameter γ has attributes of a scale parameter since it affects the variance.
3. The parameter β has attributes of a shape parameter since it affects higher moments
such as the skewness and the kurtosis.
121
Pearson Type V - PT5(γ, β, α) Non-negative Continuous
(x − γ )− (α +1) −β
−α
f (x) = β Γ(α )
exp if x > γ
Density x −γ
0
otherwise
Range (γ , ∞)
β
Mean γ+ for α > 1
α −1
β2
Variance for α > 2
(α − 1)2 (α − 2 )
β
Mode γ+
α +1
Comments:
1. X ∼ PT5(γ ,β ,α ) if and only if Y = 1 / (X − γ ) ∼ gamma(0 ,1 / β ,α ). Thus, the Pearson type V
distribution is sometimes called the inverted gamma distribution.
2. Note that the mean and variance only exist for certain values of the shape parameter α.
122
Pearson Type VI - PT6(γ, β, α1, α2) Non-negative Continuous
[(x − γ )/β ] 1
α −1
if x > γ
Density f (x) = β B (α1 , α 2 ){1 + [( x − γ ) / β ]}α1 +α 2
0 otherwise
Range (γ , ∞)
β α1
Mean γ+ for α 2 > 1
α2 −1
β 2α1 (α1 + α 2 − 1)
Variance 2
for α 2 > 2
(α 2 − 1) (α 2 − 2 )
β (α1 − 1)
γ + if α1 > 1
Mode α2 +1
does not exist on ( γ , ∞ ) otherwise
Comments:
1. X ∼ PT6(γ ,β ,α1 ,α 2 ) if and only if Y = ( X − γ ) / (X − γ + β ) ∼ beta(0 ,1,α1 ,α 2 )
3. The F(ν 1 ,ν 2 ) and PT6(0,ν 2 /ν 1 ,ν 1 / 2,ν 2 / 2) distributions are the same for ν 1 ∈ {1, 2, …} and
ν 2 ∈ {1, 2, …} .
4. Note that the mean and variance only exist for certain values of the shape parameter α 2 .
123
λ)
Poisson - Poisson(λ Discrete
e−λ λ x
if x ∈ {0,1, …}
Mass p (x) = x!
0
otherwise
Parameters λ>0
Range {0,1, …}
Mean λ
Variance λ
λ − 1 and λ if λ is an integer
Mode
λ otherwise
Comment:
1. Let Y1 , Y2 , … be a sequence of non-negative IID random variables, and let
i
X = max i: ∑ Y j ≤ 1 . Then the distribution of the Yi’s is expo(0, 1/λ) if and only if
j =1
X ∼ Poisson(λ ) .
2. If X 1 , X 2 , …, X m are independent random variables with X i ∼ Poisson(λi ) , then
Y = X 1 + X 2 + + X m ∼ Poisson(λ1 + λ2 + + λm ) .
124
Power Function - power(a, b, α) Bounded Continuous
α x − a α −1
if a < x > b
Density f (x) = b − a b − a
0
otherwise
Range (a,b)
a + bα
Mean
α +1
(b − a) 2α
Variance
(α + 1) 2 (α + 2)
a if α < 1
Mode does not uniquely exist if α = 1
b if α > 1
Comment:
1. The power(a,b,α ) and beta(a, b, α ,1) distributions are the same.
125
Random Walk - RW(γ, β, α) Non-negative Continuous
1
−α 1 − β ( x − γ ) 2
α 2
if x > γ
exp
Density f (x) = 2π ( x − γ ) 2β ( x − γ )
2
0 otherwise
Range (γ , ∞)
α +β
Mean γ+
αβ
α + 2β
Variance
α 2β
Mode γ+ ( )
1 + θ 2 − θ / β , where θ = β / 2α
Comments:
1. The parameter β has attributes of a shape parameter since it affects the skewness and kurtosis.
2. X ∼ RW(0, β , γ ) if and only if Y = X −1∼ IG(0, β , α )
126
Rayleigh - Rayleigh(γ, β) Non-negative Continuous
2(x − γ ) x − γ 2
exp − if x > γ
Density f (x) = β 2 β
0 otherwise
Range (γ , ∞)
β π
Mean γ+
2
β 2 (4 − π )
Variance
4
β
Mode γ+
2
Comment:
1. The Rayleigh(γ , β ) and Weibull(γ , β , 2) distributions are the same.
127
Student’s t - t(ν) Unbounded Continuous
ν +1
Γ
Density f (x) = 2 for all real numbers x
2 (ν +1) / 2
ν x
πν Γ 1 +
2 2
Range ( − ∞ ,∞)
ν
Variance for ν > 2
ν −2
Mode 0
Comment:
1. The t(1) and Cauchy(0, 1) distributions are the same.
128
Triangular - triang(a, b, m) Bounded Continuous
2( x − a)
(b − a ) ( if a < x ≤ m
m − a )
Density f (x) = 2 ( b − x )
( b − a )( b − m ) if m < x < b
0 otherwise
Parameters a, b, and m are real numbers with a < m < b; a is a location parameter, b - a is a
scale parameter, m is a shape parameter.
Range (a, b)
a+b+m
Mean
3
a 2 + b 2 + m 2 − ab − am − bm
Variance
18
Mode m
Comment:
1. The triangular distribution is used as a rough model for a continuous random variable in the
absence of data.
129
Uniform - U(a, b) Bounded Continuous
1
if a < x < b
Density f (x) = (b − a )
0 otherwise
Range ( a, b)
a+b
Mean
2
( b − a )2
Variance
12
Comments:
1. The U(a, b) and beta(a, b, 1, 1) distributions are the same.
130
Wald - Wald(γ, α) Non-negative Continuous
1/ 2
−α ( x − γ − 1) 2
α
exp if x > γ
Density f (x) = 2π ( x − γ )3 2(x − γ )
0 otherwise
Range (γ , ∞ )
Mean γ +1
1
Variance
α
3
Mode γ + 1 + θ 2 − θ , where θ =
2α
Comment:
1. The Wald(γ , α ) and IG(γ ,1, α ) distributions are the same.
131
Weibull - Weibull(γ, β, α) Non-negative Continuous
α −1
x − γ α
α β ( x − γ ) exp −
−α
if x > γ
Density f (x) = β
0 otherwise
Range (γ , ∞)
β 1
Mean γ+ Γ
α α
β2 2 1 1
2
Variance 2Γ − Γ α
α α α
1
γ + β α − 1 α
if α ≥ 1
Mode α
γ otherwise
Comments:
1. The Weibull(γ , β ,1) and exp o(γ , β ) distributions are the same.
2. The Weibull(γ , β , 2) and Rayleigh(γ , β ) distributions are the same.
132
Index
A
absolute-error plot, 53
absolute evaluation, 23-25, 32, 34, 38, 43, 61
Advanced Mode, 1, 9, 11, 50, 55, 56
AnyLogic, 4, 31
Anderson-Darling test, 13, 15, 29, 41, 53, 61
Applications tab, 11, 13, 14, 50, 54, 63, 64, 76, 77
Arena, 4, 31
ASCII format, 6, 88
Automated Fitting, 12, 14, 23, 32, 38, 43, 52, 55, 61
AutoMod, 4, 31
AutoSched AP, 4
B
Batch Mode, 88
Bernoulli distribution, 4, 93
beta distribution, 4, 94
binomial distribution, 4, 95
blocked machine, 71, 72, 75, 77, 80, 82-86
box plot, 51
box-plot comparisons for homogeneity, 51, 57, 60
box-plot comparisons plot, 53
box-plot comparisons percentile table, 53
busy-time distribution, 3, 71-77, 80-82, 86
C
Cauchy distribution, 4, 96
chi-square distribution, 4, 97
chi-square test, 13, 15, 29, 37, 49, 53
133
Clipboard, 2, 6, 11, 12, 42, 50, 51, 88
coefficient of variation, 18, 32, 38
Comparisons tab, 11, 13, 14, 25, 50, 53, 55
confirmation options, 15
correlation plots, 12, 51
continuous distributions, 4, 6, 90, 91
Crystal Ball, 4, 5
customer support, 9
D
Data Analysis Module, 3, 11-61
data requirements, 6, 7
data summary, 12, 14, 18, 32, 38, 43
Data tab, 11, 12, 14, 50, 51, 55
delete a model, 12, 43, 52, 63, 76
delete data, 12, 51
density function, 15, 25, 68, 76
density function plot, 13, 54, 64, 68, 73, 74, 77
density-histogram plot, 13, 15, 25, 26, 35, 40, 53, 61
discrete distributions, 4, 6, 42, 43, 49, 91
discrete uniform distribution, 4, 98
distribution function comparisons, 51, 59
distribution function, 27, 51, 53, 61
distribution-function-differences plot, 13, 15, 27, 28, 36, 41, 47, 53, 57, 61
distribution function plot, 13, 54, 57, 59, 64, 77
Distribution Viewer, 87
dongle (see security key)
downtime distribution, 3, 71-77, 80-82, 86
E
empirical distribution, 13, 14, 38, 42, 54, 55
Erlang distribution, 4, 99
134
error distribution, 4, 100
evaluation report, 13, 53
Excel, 2, 6, 88
ExpertFit software architecture, 10
exponential distribution, 4, 15, 101
exponential power distribution, 4, 102
ExtendSim, 4, 31
extreme value distribution, 4, 103, 104
F
F distribution, 4, 105
Flexsim, 4, 31
Frequency-comparison plot, 13, 15, 46, 53, 57, 58
frequency-comparison table, 53
frequency table, 12, 51
G
gamma distribution, 4, 24, 37, 72-75, 106
generate random values, 54
geometric distribution, 4, 43-49, 107
glossary, 1
goodness-of-fit tests, 13, 29, 37, 41, 49, 53, 61
Anderson-Darling (see Anderson-Darling test)
chi-square (see chi-square test)
Kolmogorov-Smirnov (see Kolmogorov-Smirnov test)
H
help, online, 1, 9, 14, 29, 55
High Precision, 1, 9, 16
histogram, 12, 19-22, 32, 33, 38, 39, 43, 44, 51
histogram comparisons, 53, 57, 58
homogeneity tests, 51, 57-60
135
hypergeometric distribution, 4, 108
I
identically distributed data (see moving-average plot)
importing data from Excel, 6, 88
independent data (see correlation plots)
index, feature, 1
installation instructions, 8
integer data, 6, 14, 43-49
inverse Gaussian distribution, 4, 109
inverted Weibull distribution, 4, 110
J
Johnson SB distribution, 4, 111
Johnson SU distribution, 4, 112
K
Kolmogorov-Smirnov test, 13, 15, 29, 37
Kruskal-Wallis test, 51, 57
L
Laplace distribution, 4, 113
lexis ratio, 43
likelihood-function table, 53
location parameter, 23, 24, 34, 75
logarithmic series distribution, 4, 114
logistic distribution, 4, 115
log-Laplace distribution, 4, 116
log-logistic distribution, 4, 32-37, 117
lognormal distribution, 4, 3, 15, 37, 62, 67, 118
136
M
Machine-Breakdown Models, 3, 71-86
machine efficiency, 72, 79-84, 86
mass function, 43
mean, 17, 18, 32, 38, 43
mean downtime of a machine, 72, 75, 79, 80, 82, 83, 86
mean number of downs, 72, 75, 80, 82, 85, 86
mean number of parts produced, 75, 83, 84, 86
mean part-processing time, 75, 83, 84, 86
median, 17, 18, 32, 38, 43
MedModel, 4
merge data sets, 51, 57
Micro Saint, 4
mode, 15, 49, 62
Models tab, 11, 12, 14, 23, 50, 52, 63, 76
moment table, 13, 54, 64, 77
moment-comparison table, 53
moving-average plot, 51
N
negative binomial distribution, 4, 43, 45, 119
normal distribution, 4, 23, 120
Normal Precision, 1, 9, 56
O
OPNET Modeler, 4
outliers, 6
P
P-P plot, 13, 15, 36, 37, 47, 48, 53, 61
Pareto distribution, 4, 121
137
Pearson type 5 distribution, 4, 38-41, 61, 122
Pearson type 6 distribution, 4, 123
percentile for a probability p, 13, 54, 62, 64, 66, 67, 77
percentile table, 13, 54, 64, 67, 69, 70
Poisson distribution, 4, 45, 124
positive skewness, 17, 18, 32, 38, 43
power function distribution, 4, 125
probability for an x, 37, 64, 77
probability plots, 13, 15, 36, 37, 47, 48, 53, 61
ProcessModel, 4
Project, 1, 2, 10, 57
ProModel, 4, 31
Q
Q-Q plot, 53
R
random variates (see generate random values)
random walk distribution, 4, 126
Rayleigh distribution, 4, 127
raw-error plot, 53
real data, 6, 14, 17-42, 57-61
relative-discrepancies table, 53
relative score, 23, 24, 32, 34, 38, 43, 45, 61
repair-time data, 38-42
@Risk, 4, 5
S
sales data, 43-49
sample distribution function, 27
scale parameter, 23, 24, 31, 34, 72-75
scatter plot, 12, 51
138
security key, 8
ServiceModel, 4
service-time data, 17-31
shape parameter, 2-13, 72-74
ship-loading data, 32-37
Simio, 4, 31
SIMPROCESS, 4
simulation-software representation 11, 13, 14, 30, 31, 42, 63, 76, 77
skewed to the right, 17, 18, 32, 38, 43
skewness, 17, 18, 32, 38, 43
SLX, 4
Standard Mode, 1, 9, 11, 50
starved machine, 71, 72, 75, 77, 80, 82-86
Student’s t distribution, 4, 128
subset selection, 12, 51
survivor function plot, 53
T
table of errors, 53
Task-Time Models, 3, 62-70
test-statistics comparison, 53
trend analysis (see moving-average plot)
transformations of data, 12, 51
triangular distribution, 3, 4, 62, 66-70, 129
tutorials, 1
U
uniform distribution, 4, 130
uniform distribution, discrete, 4, 98
V
variance, 18, 32, 38, 43
139
W
Wald distribution, 4, 131
Weibull distribution, 3, 4, 23-31, 37, 62, 67, 132
Windows, 8
WITNESS, 4, 31
140