5

Conceptual Modelling (Descriptive

Model)

Introduction
In chapter 4, we covered the use of diagramming techniques to represent the explanatory
model that provides a structural representation of the system. In general, pure explanatory
models are seldom found, and in a DES model, we usually depend on descriptive sub-
models. For conceptual modelling, when employing a descriptive model, the emphasis is on
the collection and analysis of quantitative data from observations of system behaviour. There
are a number of descriptive modelling (also termed input modelling in simulation studies)
methods, but we may be constrained in our choice by the availability and time required
for data collection. Greasley and Owen (2016) show examples from the literature of where
descriptive approaches to modelling people’s behaviour have been applied to model differ-
ent aspects of model content such as resource availability, process durations and decisions.
The following eight input modelling methods are presented:

1 Estimation
2 Theoretical distribution
3 Empirical distribution
4 Bootstrapping
5 Mathematical equation
6 Cognitive architectures
7 Machine learning
8 Trace

1. Estimation
When little or no data exists because the system does not currently exist or there is no time
for data collection, then an estimate must be made. This option may also be used in the
simulation model development phase before further data collection is undertaken when time
permits. One approach is to simply use a fixed value representing an estimate of the mean
time of the process, but this does not represent the stochastic variability in the process dura-
tion. However, this treatment of what appears to be probabilistic behaviour as deterministic
may be acceptable if it attains the level of detail required by the conceptual model.

DOI: 10.4324/9781003124092-5
72 Conceptual Modelling (Descriptive Model)
2. Theoretical Distribution
Deriving a theoretical distribution is the most common descriptive modelling method in
simulation and may be feasible when over around 25 sample data points are available. This
method can ‘smooth out’ certain irregularities in the data to show the underlying distribution
and is not restricted to generating values within the data sample used. Another important
reason to employ a theoretical distribution is that it allows us to easily change the parameters
of our distribution for scenario analysis. For example, if we are using the exponential distri-
bution for the interarrival rate, we can simply change the mean parameter of the exponential
distribution to represent either slower or faster arrival rate scenarios.
There are two options for choosing an appropriate theoretical distribution. The choice
can be made based on known properties of the variability of the modelling area being rep-
resented. For example, times between arrivals (interarrival times) are often modelled using
an exponential distribution. The overview of the main theoretical distributions given in this
chapter includes common applications of these distributions to different modelling areas.
Some examples of common usage are shown in table 5.1
The second approach and standard procedure is to match a sample distribution to a theo-
retical distribution by constructing a histogram of the data and comparing the shape of the
histogram with a range of theoretical distributions. A histogram is a bar chart that shows the
frequency distribution of the sample data. It should be noted that the histogram summarises
the frequency distribution and does not show how it varies over time.
Further examples of choosing a theoretical distribution are provided here:

• A sample size of below 25 to 30 is probably too small to fit a theoretical distribution

with any statistical confidence, although it may be appropriate to construct a histogram
to assist in finding a representative distribution.
• Statistical theory suggests that if we assume arrivals are one at a time, the number of
arrivals in a time interval is independent of the number of arrivals in the earlier time
interval, and the arrival rate is independent of time (over short time periods this may be
true), then arrival times can be represented by a Poisson process, and thus, interarrival
times can be represented using the exponential distribution.
• The triangular distribution is often used when there is a lack of data. Here, the param-
eters of the distribution can be estimated by consulting personnel with knowledge of
the process or planned process. For the triangular distribution, it will be necessary to
estimate the range parameters (minimum value and maximum value) and the most likely
parameter.
• Process durations can be represented using a uniform or triangular distribution with the
minimum and maximum values at a percentage variability from the mean. For example,
a mean of 100 with a variability of ±20% would give values for a triangular distribution
of 80 for minimum, 100 for most likely and 120 for maximum.

Table 5.1 Example Distribution Type by Modelling Area

Modelling Area Distribution

Interarrival times Exponential

Process/activity duration Gamma, Weibull, log-normal, triangle
Other elements: failure time Exponential, Weibull
 Conceptual Modelling (Descriptive Model) 73
• The normal distribution may be used when an unbounded (i.e. the lower and upper
levels are not specified) shape is required. The normal distribution requires mean and
standard deviation parameters.
• When only the minimum and maximum values are known and behaviour between those
values is not known, a uniform distribution generates all values with an equal likelihood.
• The probability values at a decision point can be derived from sample data by simply
calculating the number of occurrences within the sample for each decision choice. So a
choice for a decision made 24 times out of a sample of 100 observation provides a prob-
ability of 0.24.

Once a potential distribution candidate is found, it is necessary to estimate the parameters

of the distribution which provides the closest fit to the data. The relative goodness of fit can
be determined by using an appropriate statistical method. Examples of ‘goodness of fit’ tests
include the chi-square test, the Kolmogorov-Smirnov test and the Anderson-Darling test. See
Law (2015) for more details on these tests and the procedure. Figure 5.1 shows an example
of a histogram of sample data – in this case, machine breakdown events in a factory. A Pois-
son distribution has been fitted to the data using a chi-square test.
Although in this case a spreadsheet has been used for the analysis (see Greasley, 2004),
most simulation practitioners use a statistical software package to undertake the matching

Figure 5.1 Histogram of breakdowns and fitted Poisson distribution using the chi-square test
74 Conceptual Modelling (Descriptive Model)
process, such as the Input Analyzer for Arena (see chapter 5A) and Stat:Fit for Simio (see
chapter 5B) and Simul8 (see chapter 5C). It should be noted that if a statistical software
package is employed, then the software will simply find the best mathematical fit of the data
to a distribution. Because the software can check against a wide range of theoretical distribu-
tions, then in some instances, it may make sense to choose a distribution which offers a very
close fit rather than the closest fit. This is in order to employ a distribution that is suggested
by statistical theory. For example, this may mean choosing an exponential distribution for
interarrivals even if it is not the closest fit. This decision will need to be made based on the
experience and intuition of the simulation practitioner.
The following provides details of some of the potential continuous and discrete theoreti-
cal distributions that could be used to represent variability. These represent a subset of the
distributions available in simulation software packages, such as Arena, Simio and Simul8.
A wider selection of distributions is covered in Law (2015).

Continuous Distributions
Continuous distributions can return any real value quantity and are often used to model
interarrival times, timing and duration of breakdown events, and process durations. For a
continuous random variable, the mapping from real numbers to probability values is gov-
erned by a probability density function F(x). A number of continuous distributions are now
shown. Please note that these are examples and the distribution shape will change depending
on the parameter values chosen. The shape of the distribution is described by its skewness
(the asymmetry of the distribution about its mean) and its kurtosis (the degree of peakedness
of the distribution).

Beta
The beta distribution is used in project management networks for process duration. It is most
often used when there is limited data available from which to derive a distribution. It can also
be used to model proportions such as defective items as the distribution can only go between
0 and 1. The parameters beta and alpha provide a wide range of possible distribution shapes
(figure 5.2).

Figure 5.2 Beta (beta, alpha) distribution

 Conceptual Modelling (Descriptive Model) 75
Exponential
The exponential distribution (which may also be referred to as the negative exponential) is
used for independent interarrival events or time between process failures. As such, it is often
used to represent the interarrival time between customers or orders, where the majority of
arrivals occur quite close together with occasional long gaps between arrivals. When used
to represent process failures, the exponential should only be used when the history of the
process is not relevant to future failure times (so if past tool wear is a factor in equipment
breakdown, the exponential should not be used). This is because the exponential distribution
has ‘no memory’ of previous events (figure 5.3).

Figure 5.3 Exponential (mean) distribution

Gamma
The gamma distribution has parameters alpha (shape) and beta (scale), which determine a
wide range of distribution shapes. The gamma distribution is used to measure process dura-
tion. When the shape parameter is close to 1, exponential-like distributions may be a close
fit, and when the shape parameter is close to 2, Weibull-like distributions may be a close fit
(figure 5.4).

Figure 5.4 Gamma (beta, alpha) distribution

76 Conceptual Modelling (Descriptive Model)
Erlang
The distribution is Erlang when the shape (skew) parameter k is an integer (typically, values
between 2 and 5 are used). The Erlang distribution is used to model several sequential and
independent service phases within one distribution when the parameter k is used to represent
the number of service phases (figure 5.5).

Figure 5.5 Erlang (mean, k) distribution

Log-Normal
The log-normal distribution is used to model the product of a large number of random quanti-
ties. Its shape is similar to the gamma distribution, and it can also be used to represent process
durations, in particular those that have a relatively low amount of variability (figure 5.6).

Figure 5.6 Log-normal (log mean, log standard deviation) distribution

Normal
The normal distribution has a symmetrical bell-shaped curve. It is used to represent quanti-
ties that are sums of other quantities using the rules of the central limit theorem. The nor-
mal distribution is used to model manufacturing and service process times and travel times.
Because the theoretical range covers negative values, the distribution should be used with
 Conceptual Modelling (Descriptive Model) 77
care for positive quantities such as process durations. A truncated normal distribution may
be used to eliminate negative values, or the log-normal distribution may be an appropriate
alternative (figure 5.7).

Figure 5.7 Normal (mean, standard deviation) distribution

Triangular
The triangular distribution is difficult to match with any physical process but is useful for
an approximate match when few data points are available, and the minimum, mode (most
likely) and maximum values can be estimated. Thus, the triangular distribution might be
used when requesting an estimate of minimum, maximum and most likely times from the
process owner. What can be a useful property of the triangular distribution is that the values
it generates are bounded by the minimum and maximum value parameters. The mean value
is calculated by (minimum + mode + maximum)/3 (figure 5.8).

Figure 5.8 Triangular (minimum, mode, maximum) distribution

78 Conceptual Modelling (Descriptive Model)
Uniform
This distribution has a rectangular shape and specifies that every value between a minimum
and maximum value is equally likely. It is sometimes used when only the range (minimum
and maximum) values are known and no further information on the distribution shape is
available, but a triangular distribution would usually provide a better alternative (figure 5.9).

Figure 5.9 Uniform (minimum, maximum) distribution

Weibull
The Weibull distribution can be used to measure reliability in a system made up of a number
of parts. The assumptions are that the parts fail independently and a single part failure will
cause a system failure. If a failure is more likely to occur as the activity ages, then an alpha
(shape) value of more than one should be used. The distribution can also be used to model
process duration (figure 5.10).

Figure 5.10 Weibull (beta, alpha) distribution

 Conceptual Modelling (Descriptive Model) 79
Discrete Distributions
Discrete distributions return only whole-number or integer values and are used to model
whole-number values, such as decision choices, batch sizes and customer numbers. For a
discrete random variable, the mapping from discrete numbers to probability values is gov-
erned by the probability mass function P(x). The binomial and Poisson discrete distributions
are now shown.

Binomial
The binomial distribution is used to model repeated independent trials such as the number
of defective items in a batch, the number of people in a group or the probability of error
(figure 5.11).

Figure 5.11 Binomial (probability, trials) distribution

Poisson
The Poisson distribution can model-independent events separated by an interval of time.
If the time interval is exponentially distributed, then the number of events that occur in the
interval has a Poisson distribution. It can also be used to model random variation in batch
sizes (figure 5.12)

Figure 5.12 Poisson (mean) distribution

80 Conceptual Modelling (Descriptive Model)
Dependent and Non-stationary Processes
When using the previously mentioned distribution fitting procedures, it is important to rec-
ognise that they are based on the assumption of independent and identically distributed data.
However, this may not always be the case; for example, weekly demand amounts may be
partly based on inventory held, which is based on previous order amounts. This example
requires a time-series model in which the correlations between the observations in the series,
termed autocorrelation, are taken into account.
Another issue is when input processes change over time, they are termed non-stationary
processes. For example, many arrival patterns, such as customer arrivals in service sys-
tems, vary greatly with time due to outside factors such as lunch breaks and customer
behaviour and are thus modelled as non-stationary Poisson processes. In order to model
this behaviour, periods of time when the arrival rate remains constant must be identified
(e.g. for each hour of the day) and an arrival rate calculated for each time period per unit
of time. It should be noted that simply changing the parameter (e.g. mean) value of the
distribution for each time period will generally cause incorrect results, and the following
method should be used.

1 Determine the maximum arrival rate for all time periods.

2 Generate arrivals using a stationary Poisson process (i.e. exponential distribution
between arrivals).
3 ‘Thin out’ the arrivals by only allowing them to enter the system with a probability of the
arrival rate for the current time period divided by the maximum arrival rate determined
in the first step.

Details are provided on how to model non-stationary processes in Arena (chapter 7A) and
Simio (chapter 7B).

3. Empirical Distribution
An empirical or user-defined distribution is a distribution that has been obtained directly
from raw data. This could take the form of a summary of event log data as used by process
mining software. An empirical distribution is usually chosen if a reasonable fit cannot be
made with the data and a theoretical distribution. A disadvantage is that it lacks the capability
of a theoretical distribution to be easily scaled (by adjusting its parameters) for simulation
scenario analysis. Most simulation practitioners use a statistical software package to define
an empirical distribution, such as the Input Analyzer for Arena (see chapter 7A) and Stat:Fit
for Simio (see chapter 7B) and Simul8 (see chapter 7C).

4. Bootstrapping
This approach involves sampling randomly from raw data, which may be held in a table
or spreadsheet. This creates a discrete distribution with values drawn from a data set and
with their probabilities being generated by the number of times they appear in the data set.
This method does not involve the traditional methods of fitting the data to a theoretical
or empirical distribution and thus may be relevant when these traditional methods do not
produce a distribution that appears to fit the data acceptably. This approach will only gen-
erate data values that occur within the sampled data set, and so the method benefits from
 Conceptual Modelling (Descriptive Model) 81
large data sets to ensure that the full range of values that might be generated are present in
the data set. For example, occasionally, large service times may have an important effect
on waiting time metrics, and so the data set should be large enough to ensure these are
represented.

5. Mathematical Equation
Examples of this approach in DES include the following use of mathematical equations to
compute:

• The choice of customers using self-service or face-to-face hotel check-in facilities.

• The shortest path to determine customer movements in a retail store.
• Worker performance in terms of job accuracy and job speed.
• The selection of tasks from a queue dependent on workload and task importance.
• The choice of service provider depending on variables, including queue length for ser-
vice, distance to the service provider and service process duration.
• The effect of operator fatigue on task duration.

The most common use of mathematical equations is the use of learning curve equations
which predict the improvement in productivity that can occur as experience of a process is
gained. If an estimate can be made of the rate at which a process time will decrease due to
increased competence over time, then more accurate model predictions can be made. Math-
ematically, the learning curve is represented by the function

Y = axb

where
y represents the process time of the xth process activation,
a is the initial process time,
x is the number of process activations and
b equals ln p/ln 2.
Here, ln = log10 and p is the learning rate (e.g. 80% = 0.8).
Thus, for an 80% learning curve,
b = ln 0.8 / ln 2 = −0.322.

To implement the learning curve effect in the simulation, a matrix can be used, which
holds the current process time for the combination of each process and each person undertak-
ing this process. When a process is activated in the simulation, the process time is updated,
taking into account the learning undertaken by the individual operator on that particular
process. The use of the learning curve equation addresses the assumption that all operators
within a system are equally competent at each process at a point in time. The log-linear learn-
ing curve is the most popular and simplest learning curve equation and takes into account
only the initial process time and the workers learning rate. Other learning curve equations
can be used, which take into account other factors such as the fraction of machine time that is
used to undertake the process. Examples of constructing a mathematical equation for a DES,
such as a learning curve equation, are contained in Nembhard (2014), Dode et al. (2016) and
Malachowski and Korytkowski (2016).
82 Conceptual Modelling (Descriptive Model)
6. Cognitive Architectures
When attempting to model human behaviour in a simulation model, the use of cognitive
models can be employed. An example is the use of the Psi theory, which is a theory about
how cognition, motivation and emotion control human behaviour. Here, task performance
can be modelled over time, showing how competence oscillates based on experienced suc-
cess and failure at a new task. This approach provides an alternative to the traditional learn-
ing curve effect, which implies a continuous improvement over time without any setbacks
in performance due to failures. Another cognitive model that has been used is the theory of
planned behaviour, which takes empirical data on demographic variables and personality
traits and transforms these into attitudes toward behaviour, subject norms and perceived
behavioural control. Cognitive architectures used to represent the cognitive process include
the physical, emotional, cognitive and social (PECS) architecture (Schmidt, 2000) and the
theory of planned behaviour (TPB) (Ajzen, 1991). Examples of their use in the context of
DES are in the works of Riedel et al. (2009) and Brailsford et al. (2012).

7. Machine Learning
Machine learning (ML) techniques are increasingly being used in simulation projects. While
simulation is based on a model of a system, ML generates a model of behaviour from the
relationship between input data and output data. The combined use of simulation and ML
falls under a hybrid modelling approach (see chapter 13).
A decision tree is an example of a representation of an ML algorithm that is not ‘black
box’, and thus it is possible to trace the path by which an outcome is reached. This makes
them ideal for modelling as they can be translated into if-then-else statements for use in the
simulation model. Figure 5.13 provides a simple example of a credit application decision tree.

Figure 5.13 Decision tree for credit application

 Conceptual Modelling (Descriptive Model) 83
The decision tree can then be directly converted to program code, as shown in the follow-
ing section. The rules derived are indicated by the comments to the right of the code. This
program code can then be used in the simulation model, for example, in a conditional deci-
sion at a decision point.

IF Held Account 3+ Years THEN

IF Employed THEN
Offer Credit {Rule 4}
ELSE
IF Has Security THEN
Offer Credit {Rule 3}
ELSE
Refuse Credit {Rule 2}
ENDIF
ENDIF
ELSE
Refuse Credit {Rule 1}
ENDIF

8. Trace
Trace-driven simulations use historical process data or real-time information directly. An
example would be a simulation of a historical demand pattern (for example, actual customer
or order arrival times) in order to compare the current process design with a proposed process
design. The advantage of a trace or data-driven simulation is that validation can be achieved by
a direct comparison of model performance metrics to real system performance over a historical
time period. In addition, model credibility is enhanced as the client can observe model behav-
iour replicating real-life events. The disadvantages are the need for suitable data availability
and the possible restriction in scope derived from a single trace that may not reflect the full
variability of the process behaviour. Trace simulation is normally used for understanding pro-
cess behaviour (descriptive analytics) and for checking conformance to the ‘official’ process
design and cannot be used for predictive analytics such as scenario analysis as it only contains
historical data. Trace simulation is similar to process mining which uses historical event logs,
although process mining can check conformance of the process map in addition to operational
process performance. A summary of the trace can be used to form an empirical distribution.

Summary of Modelling Methods

Table 5.2 provides guidance on the various descriptive modelling methods:

Table 5.2 Descriptive Modelling Methods

Disadvantages Advantages Comments

Estimation Lack of accuracy. May be the only option. May be used in the
model development
phase before further
data collection is
undertaken.

(Continued)
84 Conceptual Modelling (Descriptive Model)
Table 5.2 (Continued)

Disadvantages Advantages Comments

Theoretical No available theoretical Can ‘smooth’ data to the Best choice if a

distribution distributions may fit underlying distribution. reasonable data fit can
data. Generates values Generates values be made.
outside of the data outside of data sampled.
range which may not For example, rare
be appropriate. events at the tail of the
distribution which may
be missing from a small
data sample collected.
Compact method of
representing data values.
Distribution is easy to
scale for scenario
analysis.
Empirical Cannot usually generate Provides distribution An option if no
distribution values outside when no theoretical theoretical distribution
the range of data distribution provides an can be fit.
(therefore may miss adequate fit to data.
‘extreme’ values).
Difficult to scale for
scenario analysis.
Bootstrapping Distributions are derived Avoids an issue of poorly Likely to require large
only from values held fitted distributions. data set to ensure
in the data set. coverage of a full
range of data values.
Equation May require extensive Shows relationships that Can be useful
data collection. are a consequence of for expressing
simulated time, such as relationships between
learning effects. variables.
Cognitive Difficult to validate Attempts to model factors Limited use due to
cognitive models. leading to individual demanding data
human behaviour. collection and
validation needs.
Machine Using some ML Can employ complex logic Need for modeller to
learning techniques, the logic processes. have analytic skillset.
implemented may not May require interface
be inspected. between simulation
and ML software.
Trace Only uses values held in Replicates reality to check Not suitable for
the data set. for conformance. prediction.

What Variability Are We Modelling?

When modelling variability, we need to consider both the granularity of our modelling
and the need to distinguish the ‘normal’ variability we wish to model and variability
caused by a change in the process.
In terms of granularity, we need to ensure that we model variability at a sufficient
level of detail to ensure that it can be represented correctly. For example, if data is col-
lected of the checkout process at a supermarket, this will entail placing goods onto a
 Conceptual Modelling (Descriptive Model) 85

conveyor, scanning the items by the checkout operative and then paying for the items.
To collect data on the overall process ignores that the conveyor and scanning process
times are dependent on the number of items being bought while the payment process is
not. Thus, if we are using the method of fitting to a theoretical distribution, the overall
process should not be fitted to a single distribution but broken down for analysis. As
well as sequential processes, care should also be taken with parallel processes in which
different options may require fitting to separate distributions. For example, mutually
exclusive payment types (cash, card, touch payment) will have different distributions.
In addition, we need to note that the theoretical distribution represents the normal
variability, which is inevitable and what we would expect of a process design operat-
ing in a given environment. We can only remove this variability by improving the
design of the process. If the variability is normal (i.e. the variability is due to random
causes only), then the process is in a stable state of statistical equilibrium, and for
example, if we are modelling using a probability distribution, then the parameters of
its probability distribution are unchanging. Conversely, if the performance variability
is abnormal (i.e. due to a change in process procedures or environment, for example),
then the abnormal variability disturbs the state of the statistical equilibrium of the pro-
cess leading to a change in the parameters of the distribution.

Summary
Eight input modelling methods are presented in this chapter: estimation, theoretical distribu-
tion, empirical distribution, bootstrapping, mathematical equation, cognitive architectures,
machine learning and trace. The most common approach employed in DES is the theoretical
distribution method, which provides a compact method of representing data values, with the
distribution easy to scale for scenario analysis.

Exercises
• Compare and contrast theoretical and empirical distribution input modelling methods.
• Evaluate the trace input modelling method for a simulation study.
• Evaluate the use of machine learning as an input modelling method.
• Compare the process of deriving a theoretical and empirical distribution.
• A number of observations have been made of arrivals to a supermarket (see table 5.3).
The frequency of arrivals in a ten-minute period is as follows. Using a chi-square ‘good-
ness of fit’ test, find if the arrivals can be described by a Poisson distribution with a mean
of 1 at a significance level of 0.1.

Table 5.3 Supermarket Arrivals

Number of Arrivals Frequency

0 70
1 80
2 34
3 12
4 4
Total 200
86 Conceptual Modelling (Descriptive Model)
Table 5.4 Customer Serve Time

Serve Time Observed Relative Cumulative Frequency

(minutes)
Less than 1 0.02
Less than 2 0.29
Less than 3 0.54
Less than 4 0.59
Less than 5 0.90
Less than 6 0.95
Less than 7 1.00

Table 5.5 Arrival Times

4.9 4.4 3.8 5.9 4.6

2.9 3.9 3.8 6.8 5.3
3.5 1.3 3.1 4.3 3.5
6.5 5.3 4.6 5.1 3.5
5.4 4.3 4.7 3.7 3.8
5.5 2.6 6.2 4.5 8.8
3.3 6.5 7.2 2.8 6.2
5.0 6.9 3.9 7.6 5.3
2.7 1.9 2.0 5.2 4.6
5.1 4.3 2.4 4.6 7.9

• A number of observations have been taken of customers being served at a cafeteria (see
table 5.4). Analysis has revealed a mean of 2 minutes and a standard deviation of 0.5
minutes for the data. Investigate using a Kolmogorov-Smirnov test that the observations
are normally distributed at a significance level of 0.1.
• The following data has been collected on arrival times (see table 5.5). Find the param-
eters of an empirical distribution that fits the data.

References
Ajzen, A. (1991) The theory of planned behaviour, Organizational Behaviour and Decision Processes,
50(2), 179–211.
Brailsford, S.C., Harper, P.R. and Sykes, J. (2012) Incorporating human behaviour in simulation mod-
els of screening for breast cancer, European Journal of Operational Research, 219(3), 491–507.
Dode, P., Greig, M., Zolfaghari, S. and Neumann, W.P. (2016) Integrating human factors into dis-
crete event simulation: A proactive approach to simultaneously design for system performance and
employees’ well-being, International Journal of Production Research, 54(10), 3105–3117.
Greasley, A. (2004) Simulation Modelling for Business, Ashgate Limited.
Greasley, A. and Owen, C. (2016) Behavior in models: A framework for representing human behavior,
in M. Kunc, J. Malpass, and L. White (eds.) Behavioral Operational Research: Theory, Methodology
and Practice, Palgrave Macmillan.
Law, A.M. (2015) Simulation Modeling and Analysis, 5th edition, McGraw-Hill Education.
Malachowski, B. and Korytkowski, P. (2016) Competence-based performance model of multi-skilled
workers, Computers and Industrial Engineering, 91, 165–177.
Nembhard, D.A. (2014) Cross training efficiency and flexibility with process change, International
Journal of Operations and Production Management, 34(11), 1417–1439.
Riedel, R., Mueller, E., Von Der Weth, R. and Pflugradt, N. (2009) Integrating human behaviour into
factory simulation – A feasibility study, IEEM 2009 – IEEE International Conference on Industrial
Engineering and Engineering Management, 2089–2093.
Schmidt, B. (2000) The Modelling of Human Behaviour, SCS Publications.