Development of a Human Resources Transition Simulation
Model in Slovenian Armed Forces
Andrej Skraba, Miroljub Kljajié
Andrej Knafli¢, Davorin Kofjaé, Iztok Podbregar
University of Maribor, Faculty of Organizational Sciences
Cybernetics & DSS Laboratory
Kidriéeva cesta 55a, SI-4000 Kranj, Slovenia
e-mail:{andrej.skraba, miroljub.kljajic, andrej.knaflic, davorin.kofjac}
@fov.uni-mb.si
iztok.podbregar@fvv.uni-mb.si
Abstract
The paper describes development of continuous and discrete model of human resources
transitions in large organization. The model considers eight different ranks. The calibration
of the model was performed where the historical data was used to determine time constants
of transitions and fluctuations. Basic simulation runs were performed in order to complete
predictive validation of the model. Optimization of the model was performed by application
of pattern search algorithm considering the key parameters. By this means the target strategy
of the system could be determined. Development and comparison of continuous and discrete
event model was performed. Discrete event model was applied in the validation phase. Hybrid
approach to the problem provided higher level of confidence. System dynamics methodology
proved to be appropriate as the tool for initial development of the model and structural vali-
dation reference.
Keywords: simulation model, system dynamics, human resources, transitions, optimization, target func-
tion, strategy
1 Introduction
In a large and complex organizational structure, human resour
manding and on going process. From an abstract
is represented as a delay chain, for which we need to determine
s management (HRM) is a de-
system dynamics (SD) point of view, this problem
parameters, that the target:
functions will be achieved. In such case, prediction or anticipation is a very important point of
view [1, 2, 3, 8, 9, 10, 11], and serves as a prerequisite for efficient management of complex systems.
Problem of dynamics can be addressed by continuous or event based simulation. Applying both
methodologies, additional validation of developed models can be achieved using their compari-
son. It is important to take advantages each of the approaches has, while considering differences
amongst them, especially while preparing input data for simulation. While solving this kind of
problems, event based simulation is to be more accurate. However, higher accuracy is bonded to
time demanding data preparation. Comparison between discrete and continuous model provides
means of structural and quantitative validation.
Main objective of present study is development of a dynamical model, which allows structural
pursuit and foresecing of fluctuation amongst individual ranks in a large and complex organizational
structure; in our case Slovenian Armed Forces. Developed
sm should provide optimal strategy
to achieve desired rank structure. The goal structure should be achieved by the consideration of
parameter limits and longer time constants which determine system response. By the means of
the optimization the best solution for armed forces development should be provided. Due to the
restructuring the addressed problem is of significant importance. The study is based on a system
simulation method, where SD and Discrete Event Simulation (DES) are used. SD procedure i
more aggregated and that is why more understandable to users, while DES assures higher accuracy
of gained results. In this phase we developed two differently aggregated SD models, based on
ions amongst ranks. The SD model is built from the common entities of system
7 js and rates. Individual level represents number of rank members (a number of
persons positioned in a rank) and are labeled as A, B, C, ... , while rate represents the number
of persons who enter or leave each rank in one time unit. Dynamical model is developed in a
numerical computing environment and programming language MATLAB. Time unit of simulation
run in a continuous model is chosen as one month. Calibration of this model was performed using
a historical data, where simulation results perfectly get along with anticipations. While time period
for such simulation is free to choose in our case simulation was performed for a period of 10 and 15
years. Besides accuracy, discrete event model also includes more detailed data, as it also considers
duration of training and education, which individual spends in a rank before the transition to the
next level. Model is in a stage of evaluation where complexity - benefit ratio is consider
2 Metodology
Long-term HR planning is a strategic and very important part in a process of preparation and
realization of such a complex organizational system also for the sake of cost reduction. All these
entities are a part of a complex social system, where we can expect results only on a longer
time periods because of a large time constants and delays across feedback loops [16]. Behaviour
of such systems is very sensitive to slightest initiatives from environment and different decision
policy proceedings. This means that the developed model must enable a simulation (prediction) of
personnel structure dynamics; considering transitions amongst different ranks and their fluctuation
from inside and outside of the system.
= Demand = Demand = Demand
forrankA forankB forrankN
oO ~N Oo
“raining ney aing
HR capactly A vamcapactty 8 tam capacity N
- anon ‘ td. Instructors
i ar Rank 8 =e N\ training “> Ronkn ™
yr technology 7 ‘technology aa Training
capaciyA capaciyA eee
Discharge Retention Discharge Retention ai, Retention Sapacty N
rateA rates vaio rate rato ‘ate
Figure 1: Causal Loop Diagram for rank transitions.
The model is based on the system dynamics methodology [13]. There are many examples
of system dynamics methodology implementation in the field of workforce management [7, 5, 6].
The main purpose of the system developed within our research framework is to perform and
support a process of complex decision making concerning human resources management by man-
machine interaction. Different models designed simulate the flows of personnel through various
ranks, answering to a question what would be the system behaviour like considering different
input resources. This means, that user can identify potential manpower shortfall and calculate
financial impacts for a chosen policy and examine possible curative policies or rationalizations only
by mimicking operations of real-world system. But this can only be achieved, if we develop a
validated simulation model, which will take into consideration model feedback loop dependencies
scially time constants which define human resourc
agram of the system [4]. Transitions in the system are dependant on demand for
particular rank and training capacity which ig basieally hierarchical. In ternis of
this would mean causal dependance of each particular rank. Number of membe
depenc
influences number of instructors as well as number of rank A members
es dynamics. Figure 1
"
tem dynamics
sin each rank
on the demand which determines the training human resource capacity. This positively
On the other hand, the
instructors diminish the number of members in particular rank since these instructors work with
higher rank. The instructor effort is supported by new technologies in military training. While
the structure is repetitive representing a chain of negative feedback loops, the task to achieve the
target function, meaning a particular dynamic of particular rank members is main issue addressed
in the research. Particular rank is also determined by discharge rate and retention rate.
a b n
ft ft ft
> LAE ee ts
Rv Ra Re Ry
A cannonical a
form
r aq,
c . v
Fa Fe Fy
Figure 2: Structure of system dynamics model of rank transitions in the canonical form. Shown
structure represents delay chain of first order delay elements.
Such SD model, can be easily
of next activities
adopted in a group decision-making process [12] by realization
1. Definition of structure, development of model, definition of goals and criteria.
2. Gathering of data and calibration of model.
3. Validation of model.
4. Preparation of simulation scenarios and what-if analysis.
5. Simulation scenario analysis
system development.
6. Simulation test and initiation of a model.
Figure 2 represents the stock and flow diagram of an SD model, which represents transitions
of manpower amongst individual ranks. The structure shows that a cl system
is examined in a way where individual workforce type is produced within the system rather than
imported from the outside or in a stric system, where i.e. higher rank manpower
bust be from next rank below. This also means that the change at one rank can create chain
demand in other ranks. Stock variables (levels) describe the state of the sys
number of manpower inzank A and rank B, while flow variables (rates) illustrate the rates of
change of stocks, such as fluctuation and transition rates, as stated below:
Ry is an clement of change of stock, denoted by value r, which tells the ratio of transitions from
a prior rank (in this case out of s es) to a rank A,
R,4 is an element of change of stock, denoted by value a, which tells the ratio of transitions from
arank A to a rank B,
Rp i
d labour fore
ystem bounda
an element of change of stock, denoted by value b, which tells the ratio of transitions from
a rank B to a next rank (in this case out of s
em boundaries),
F4 is an element of change of stock, denoted by value c, which tells the fluctuation out of the
system from a rank A,
Fg is an clement of change of stock, denoted by value d, which tells the fluctuation out of the
system from a rank B.
ole xb
it as
len “@iour pw @l gr — Pf ial . Soules Fe Poe %
sis lve] vou |
AOE, cpealcaan Lael Sapa aT * .
for “@ieur ob _ |
Figure 3: Model of rank transition developed by Discrete Event Methodology in MAT-
LAB/SimEvents.
sful transitions into rank B.
of evaluations of monthly transitions from rank A to rank B and can
Now lets try to define variable a, which represents ratio of a suc
This rate depict on a basis
be defined as:
m
Ao
()
where m is the expected number of members moving from rank A to rank B in an interval
of unit length of time, Ao is the start size of a rank A (k = 0) and 7 represents a unit length of
time (it tells if transitions are done yearly, monthly or even in shorter period of time). The other
equation in this case could be:
n
= (2)
;
where 7) being the transition probability and 7 a time needed for rank A recruitment.
Equation 3 depicts the following fictious ratio d, which in our model represents fluctuation ratio
from rank B, where m = 15, So = 375 and r = 12:
15 1 1
- 35 ~ 30 (8)
All other ratios (a, b, c and r) can now be calculated in a same manner. Results gathered in
such way depend on the accuracy and quality of data, used for calculations. This means that the
best way is to combine data about yearly transitions amongst individual ranks with expectations
of how long individu: in each rank.
Model on a figure 2 can also be represented as a
stem of difference equations [15, 14]:
A(k+1) = A(k) + At(R,(k) — Ra(k) — Fa(k)) (4)
Ra(k) = aA(k)
Fa(k) = cA(k)
Blk+1) = B(k) + At(Ra(k) — Ra(k) — Fe(k))
Rp(k) = bB(k)
Fa(k) = dB(k)
N(k+ 1) = N(k)+At(Re(k) — Rv(k) — Fw(k))
Ry(k) = nN(k)
Fy(k) = vN(k)
Figure 3 represents a model of transitions amongst ranks, developed by principles of DES
constructed by MATLAB/SimEvents. Element Rank A characterize a server entity in a discrete
event simulation, initialized by the start size of a rank A. Here we use event based generator of
s. The level of Rank A element is controlled by a step function, which represents an input of
tem. Variation of transitions of entities from Rank A to the next level is presented as frequency
s of historical data about
dynamics and transitions amongst individual ranks. Fluctuation ratios are also presented as user-
defined functions, gained by analysis of historical data about fluctuations. This is then followed as
an output to the next level, which in our case is Rank B.
Structure of discrete model is identical to the structure of continuous model. In a discrete model
the transitions amongst ranks are represented as stochastic distribution functions. This means that
duration which entity spends in individual rank is represented as cumulative distribution function
gathered from a data base. Model enables the variation of discrete transitions by principles of
Monte Carlo simulation and is by that more accurate than the continuous one.
distribution of entity stay time in Rank A, which was acquired by analys
2
1.5
1
Generated entities
0.5
0)
—————
100 150
Time [month]
Figure 4: Generated entities as rank fluctuation
Figure 4 shows the dynamic of entity generatio:
from clement Rank A. All individual entities are generated according to given fluctuation process
distribution. Each and every one of these points represent fluctuation of one manpower as entity
of Rank A. While developing simulation models, where modelling of complex systems in addressed,
implementation of hybrid simulation in simulation tool plays an important role [18]. In our case,
combination of continuous and discrete simulation was performed using MATLAB. This functionality
was build in a position to participate in the process of predictive, replicative, structural and
pragmatical validation.
Examples of simulation results from continuous and di
which in our case represents fluctuations
rete model are shown in a figure 5.
mulation runs in a 15 years simulation
Group of curves from a discrete model results represent 50
on
350
300
Average with
confidence
250 bounds
Results of 50
o simulation runs
@ 200 of discrete evenit
a. model
ra
5
© 150
Results-of
100 f continuous
‘simulation
model
50
0 20 40 60 80 100 120 140 160 180
Figure 5: Comparison of continuous and discrete simulation model results. (confidence level a =
0.05 with average of fifty simulation runs.
duration (180 months). Because of the scale of time constants, represented in models, a longer
period of simulation time is required. As we can see on this figure, considering given parameter
limitations, in 5 year duration of simulation target conditions could not be met. In a proc
of model validation, results from continuous and discrete model should match. This form of
procedure is very important, especially regarding parameters hypersensitivity. Stati t-test
can be performed, to confirm if results correspond, by confirmation or rejection of parameter
adequacy.
Model of transitions, developed by principles of continuous simulation using MATLAB/Simulink
is presented on a figure 6. Step and discrete time delay function Z~! is realized as input to a
Rank A. Parameter values and simulation results are rendered through MATLAB Workspace. Level
of element textitRank A depends on initialization value init A and the difference of 4 integrator
inputs, presented by parameters fluctuation A ratio and transition A ratio. According to discrete
model, main difference is, that here transition ratio is defined only by parameter transition A ratio,
which is the critical element in definition of model.
3 Definition of strategy according to target functions
Figure 7 shows the structure of the optimization problem. The core of the system is Rank Structure,
where different ranks hierarchically interact with each other. The system is driven by the Desired
Rank Structure Dynamics. In our case target functions were determined. Depending on the Rank
Structure Adequacy the Optimization Algorithm determines the Parameter Sct Values according
to the Parameter Boundaries. Response of the Rank structure is in the next cycle compared to
the Desired Rank Structure Dynamics where the procedure repeats until the satisfactory Rank
Structure is reached.
Definition of strategy was defined as optimization problem. Optimization problem is a proc
of finding the best solution from all feasible solutions. Figure 8 describes the dynamics of criteria
function value variation trough 600 iterations of simulation. Iteration is presented as calculation
of the value of criteria function, minimisation of which is required:
rth
mip T= 33> (Cn X(n) 6)
n=1i=0
-
(== }$ 5]
‘Adp >|
oute
Ea, EbE) ih:
(pt. p2]
nputAT
input AZ
outaaly
Figure 6: Rank transition model developed by the principles of continuous
in MATLAB/Simulink.
imulation developed
where u is a set of input parameters with some value limitations, n is the number of optimized
ranks and by that a number of known criteria functions C,,. We need to specify target function for
Iculate the mean square deviation from the value of observed parameter
Xn. ty is a dis Sum of mean square deviations from the values of criteria functions are
then summed again according to the set of target functions [17]. If the init values of ranks are
known as 1(0),r(0),73(0),...,2n(0), we are able to manage the control functions in a following
manner:
each observed rank and c:
ete tim
= filai,r2,2%3,- 1 U1, U2, US, +) Un)
Fy Ut U2, US, «4 Un)
= fo(r1,22,73,-
= fa(w1, 22,03 1 U1, U2, U3, +; Un)
ral 01 22,203 oy Bny Ur, UD, Ug, «4, Un)
where u» are time dependant control variables. In our scenario, the primary objective is to
define the values of u, variables in a such way, that values of rates are brought from their init
values:
21(0), 72(0), 3(0),..., 7n(0) (6)
to their target value
x1 (te), v2(tk),
Strategy
Determination
Rank
Parameter Structure
Boundaries
a
Optimization Desired
Algorithm Rank
‘Structure
as Dynamics
+
Rank
‘Structure
Adequacy
Figure 7: Optimization problem key components.
essed by equation 5, a numerical method was used.
that probability for deter-
mination of analytical or mathemati iderably small. In our
in search of optimum or definition of best strategy, according to target functions, application of
pattern search algorithm was performed.
For minimisation of criteria function expr
Assumption of empirical functions in such optimization problems c
al accession of optimum is co
ie,
x10° Minimum value of criteria function: 3046.4437
25
Value of criteria function
i) 100 200 300 400 500 600
Iteration
Figure 8: Value of criteria function with respect to algorithm iteration. Application of pattern
search algorithm for determination of optimal strategy (600 iterations).
Figure 9 represents attainability of target functions for observed ranks A, B, C and D. Target
function should be expressed for each and every rank, which is to be observed. According to given
target and criteria functions expre
considering the limitations of key parameters. The dashed curves on a figure 9 denote the target
functions of individual ranks, while solid lines denote system response or in our case the actual
number of manpower in a rank.
As a result, the system calculates an optimal strategy for achieving target functions within a
set of parameter limitations. The example of parameter variation according to the stated targed
d by equation 5 our goal is to aim for minimal deviations
400
350
300
8g ‘rank C
@
~ 250 target function
S /
S
a
200
target D
ZL
150 response
0) 40 20 30 40 50 60 70
Time [month]
Figure 9: Target function approximation (dashed lines) for ranks A, B, C in D (full lines)
Change of parameters according to the target function
0.04
0.035
Parameter value
°
So
26
G8
1 2 3 4 bd 6 7
Time[month x 10*]
Figure 10: Example of parameter variation according to the target function.
function is shown in Figure 10.
The dynamics of the parameter variation is not convenient for a di
conclusion about the computed strategy. Therefore the tabular results
representation. Table 1 (*values are altered due the confidence reasons) shows the strategy deter-
mination for particular rank. The number of officers in rank is shown, the inputs to particular
rank, transitions to the next rank and Fluctuation. In the last two rows the parameter values
Transitions to the next rank and Fluctuation parameters are shown. Table 1 is one of the possi
stem and provided to the decision make
pending on the parameter constraints, feasibility and government policy the possible acceptable
strategy could be selec
ision maker to draw a
are more convenient for
solutions automatically computed by developed
ed and implemented.
4 Conclusion
The main conclusion after performed development phase is, that system dynamics approach con-
tributes to the main development cycle with its transparency. The structures developed by system
dynam: One of the impor-
approach were referential during the entire development pro
Table 1: Strategy determination for particular rank
month 0 10 20 30 40 50 60 70
No. inrank 4620 4633 4627 4598 4590 4604 4593 4607
Rank inputs 86 58 23 34 56 28 56 x
Trans. next 59 49 44 38 38 39 30 x
Fluctuation 14 15 8 4 4 0 12 x
Trans. par. 0.0328 0.025 0.0101 0.0152 0.025 0.0128 0.025 0
Fluct. par. 0.0098 0.0098 0.0061 0.0032 0.0037 0 0.0098 0
tant advantage of the continuous models is the possibility to perform the optimization and target
function strategy search. In this regard the Markov chain approach is less suitable [4]. Hybrid
approach was almost mandatory in order to provide the appropriate level of validation confidence.
Due to the importance of the problem addressed, the validation was the crucial methodological
topic. Presented approach incorporating system d s methodology is structured in a way
that the prediction of what will happen in a system if current policies are kept is enabled. Op-
timization problem solving technique on the other hand provides answers to the question what
kind of policies should be implemented to fulfil given goals. Developed system based on MATLAB
provides a computational engine which provides the dynamical strategy according to the stated
target function. Stated optimization problem represents significant computational burden which
could be effectively addressed by the means of parallel computing. On the example of Slovenian
this is of primary importance on account of ongoing restructuring proc
Presented approach provides: effective algorithm to provide the strategy for management: of
amit
Armed Forces
large scale HRM problem with high importance for state. Results of the system are applicable to
the real world system almost instantancously.
Future research process will include model testing actions 2 . Another
aspect of future research will include the development of sophisticated simulation graphical user
interface (GUI) and introducing feedback information in human resources management process.
Final tool should generate a list of possible strategies how to deploy human resources management
policies in a large and complex organizational system according to given target functions. Such
tool could be applicable in different type of complex workforce planning processes like in Slovenian
army and other large organizations.
Acknowledgement
This research is sponsored by ARRS ~ Slovenian Research Agency ~ (Target research programmes -
and Security 2006-2010", project code: M5-0175).
‘Science for Peace
References
[1] Dubois D., Resconi G. (1992) Hyperincurisvity — a new mathematical theory, 1°ted. Presses Universitaires de Lidge
[2] Rosen R. (1985) Anticipatory Sy
ems. Pergamon Press
[3] Kljajié M. (1998) Modeling and Understanding the Complex System Within Cybernetics. Ramaekers M. J. (Ed.), 15th
International Congress on Cybernetics. Association International de Cybernetique. Namur. 864 — 869
[4] Wang J. A review of Operations Research Applications in Workforce Planning and Potential Modelling of Military
Training. DSTO Systems Sciences Laboratory. Edinburgh Australia. 2005.
[5] Trost CS, “A dynam
(John Wiley & §
model of work quality in a government oversight organization” , System Dynamics Review 18(4),
sster, 2002) 473-495
[6] Mayo DD, Callaghan MJ and Dalton WJ “Aiming for restructuring success at London Underground”, System Dy-
namics Review 17(3), (John Wiley & Sons, Chichester, 2001) 261-289.
[7] Bajracharya A, Ogunlana SO, Bach NL. “Effective organizational infrastructure for training activities: a case study
of the Nepalese construction sector”, System Dynamics Review 16(2), (John Wiley & Sons, Chichester, 2000) 91-112.
[8] Kljajié M. (2001) Contribution to the meaning and understanding of anticipatory systems. Dubois D. M. (Ed.).
Computing anticipatory systems, (AIP conference proceedings, 573). Melville (New York): American Institute of
Physics, 400 — 411
10
{9]
[10]
fu
(22)
(13)
(ua
(15)
[16]
(7)
[1s]
Kljajié, M., Skraba, A., Kofjaé, D. Bren, M., (2005) Discrete Cobweb Model Dynamics with Anticipative Structure.
WSEAS Tranactions on Systems. Vol. 4, Issue 5.
Skraba, A., Kljaji¢, M., Kofjaé, D. Bren, M., Mrkaié M. (2005). Periodic Cycles in Discrete Cobweb Model. WSEAS
‘Transactions on Mathematics. Issue 3, Vol. 4, July 2005. pp. 196-203.
Skraba A, Kljajié M, Kofjaé D, Bren M, Mrkaié M (2006) Anticipative cobweb oscillatory agents and determination
of stability regions by lyapunov exponents. WSEAS Transactions on Mathematics 12(5):1282-1289
Skraba A., Kljajié M, Leskovar R., “Group Exploration of SD Models - Is there a Place for a Feedback Loop in the
Decision Process?” , System Dynamics Review, (John Wiley & Sons, Chichester, 2003) 243-263.
Forrester JW. 1973. Industrial Dynamics. MIT Press, Cambridge, MA.
Kreyszig, E. (1993) Advanced Engineering Mathematics. 7" ed. John & Wiley Sons Inc.
Luenberger, D. G. (1979) Introduction to Dynamics Systems \ Theory, Models & Applications. John & Wiley Sons
Inc.
Wiener N. (1961). Cybernetics or Control and Communication in the Animal. MIT Press: Cambridge, MA.
Watkins T (2007) “Pontryagin’s Maximum Principle”. http: //waw.sjsu.edu/faculty/watkins/pontryag. htm
(15/1/2007)
Kljajié M, Bernik I., Skraba A., “Simulation Approach to Decision Assessment in Enterprises” , Simulation, (Simulation
Councils Inc., 2000) 199-210.
ll
