Expert Systems With Applications 55 (2016) 172183

Contents lists available at ScienceDirect

Expert Systems With Applications

journal homepage: www.elsevier.com/locate/eswa

Using fuzzy logic to implement decision policies in system

dynamics models
Daniel Cardoso de Salles1, Armando Celestino Gonalves Neto1, Lino Guimares Marujo
Department of Industrial Engineering, Polytechnic School, Universidade Federal do Rio de Janeiro, Av. Athos da Silveira Ramos 149, PO Box 64548, Cidade
Universitria, Rio de Janeiro 21941-909, Brazil

a r t i c l e

i n f o

Article history:
Received 11 December 2014
Revised 26 January 2016
Accepted 27 January 2016
Available online 13 February 2016
Keywords:
Fuzzy logic
System dynamics
Balance dynamic scorecard
Managerial dashboard

a b s t r a c t
This paper aims to show how the use of fuzzy systems can enhance the application of system dynamics
(SD), with the construction of virtual worlds, for organizational learning. By doing so, the main contribution is to propose a fuzzy-SD integrated methodology that allows a natural language modeling of decision
policies. At rst it will be revised the recent literature with practical and theoretical fuzzy-SD integration
showing that the motivation and purpose of this article have not been explored yet. Following, it will
be scrutinized the origin of system dynamics to theoretically justify the proposed fuzzy-SD integration
for organizational learning. Finally, it will be presented a hypothetical case study based on the balance
scorecard model to illustrate the methodology in action.
2016 Elsevier Ltd. All rights reserved.

1. Introduction
Almost simultaneously, two revolutionary knowledge elds
were developed: Systems Dynamics, created by Jay Forrester at
Massachusetts Institute of Technology in the 50s (Forrester, 1961),
and Fuzzy Logic created by Lot Zadeh in the 60s (Zadeh,
1975a). Both arose from the perception that the tools of their times
were unable to deal appropriately with real problems (Sterman,
20 0 0; Ross, 2005).
System Dynamics (SD) can be dened as the branch of control
theory that deals with socioeconomic systems and the branch of
science that deals with management issues controllability (Coyle,
1996). Forrester perceived the deciency in our decision-making
process when dealing with the complexity of dynamic systems. To
overcome this deciency, he proposed modeling tools that allowed
discussion, rening and simulation of mental models exposed to
different policies. His greatest contribution was the Stock and Flow
Language that allowed the creation of virtual worlds, controlled experimental laboratories to enhance learning.
Fuzzy Logic (FL) and related Fuzzy Sets Theory have been created with the purpose of adapting the mathematical tools of logic
to types of uncertainty such as vagueness and approximation, characteristics of natural language and of human mental models. By
doing so, FL enables a representation of human knowledge by lin

Corresponding author.
E-mail addresses: daniel_cardoso@poli.ufrj.br (D.C. de Salles),
armandoneto@poli.ufrj.br (A.C. Gonalves Neto), lgmarujo@poli.ufrj.br (L.G. Marujo).
1
Tel.: +55 21 3938 8065.
http://dx.doi.org/10.1016/j.eswa.2016.01.048
0957-4174/ 2016 Elsevier Ltd. All rights reserved.

guistic IF-THEN expressions, characteristic of approximate reasoning (Zadeh, 1975a, b).

Besides the Introduction, this article will be structured as it follows: on Section 2 it will be revised all FL-SD integrations proposed in the literature; on Section 3 it will be introduced a theoretical discussion on learning in SD environment showing how the
use of FL could enhance its potential; on Section 4 it will be properly presented the FL-SD integrated methodology; on Section 5 it
will be presented a case study exemplifying the methodology application; and nally Section 6 contains the discussion and conclusion remarks.

2. Review of literature
Traditionally, literature uses the FL approach especially for natural language processing and inaccurate knowledge in expert systems, process control and pattern recognition (Karavezyris, Timpe,
& Marzi, 2002). At the other extreme, the literature of dynamic
systems permeates many areas of study, especially socioeconomic
and administrative, basically applicable to any system we wish to
study.
The rst authors to integrate these two approaches were Pankaj
and Sushil (1994), who proposed a method for qualitative analysis
of causal loops using fuzzy linguistic uncertainties to incorporate
the perceptions and beliefs of the modeler. As in most of later integration proposals their motivation was the perception that variables relationships in humans mental models are best expressed
in natural language.

D.C. de Salles et al. / Expert Systems With Applications 55 (2016) 172183

Most of the available researches integrate FL to SD models in

order to consider fuzzy variables (their relations and arithmetic, or
soft descriptions) when data was unavailable, with low credibility
or because specic variables presented fuzziness uncertainty. Examples of such motivation and practical integrations can be seen
by Campuzano, Mula, and Peidro (2010); Kunsch and Springael
(2008) and Karavezyris et al. (2002). In these applications, FL was
not used to dene policies to control the model but to deal with
uncertainty in some variables of the model. The control of their
models was presented in a classical way (mixed in the model and
with crisp parameters).
Others like Song, Song, and Zhang (2015), and Orji and Wei,
2015a, 2015b) used SD models to simulate alternative scenarios
that were subsequently ranked by multi-objective programming
(MOP), which embedded FL. In a similar application, Xu, Deng, and
Yao (2014) used fuzzy to model uncertain parameters and a MOP
to select alternative policies that were individually evaluated in a
SD model.
In a more elaborated extension of this application, Chang and
Ko (2014); Xu and Li (2011) and Wu and Xu (2013) used the MOP
to control each step of the SD model. This proposal is in a sense
similar to the proposed methodology because the controller is totally dissociated from the model to be controlled. However, the focus was not on the learning throughout development and testing
of politics, since the control was automated by a MOP controller
that searched for optimization.
Another different FL-SD integration can be found by Sabounchi,
Triantis, Sarangi, and Liu (2014), who used FL to model a decision
rule inside their SD model (users preference on transportation).
As discussed in the article, the specic decision is naturally fuzzy
which demands the introduction of a Fuzzy Inference System (FIS)
(Mamdani & Assilian 1975). In a sense this article presented a practical application of the currently proposed FIS controller; however,
there is a main difference that should be highlighted: the model
decision policies were not transferred to the FIS controller since
the stakeholders policy decision was still mixed into the model
(classically). The FIS was introduced only to model a specic variables relation that was, by nature, fuzzy. However, although not
very different in a practical way, there is a considerable theoretical
difference since the complete dissociation of control policies proposed in this present article is the main aspect to reinforce virtual
worlds learning.
Baradaran and Keshavarz (2015) also used a FIS (specically a
Sugeno type) in part of their model to express specic relations
between variables that are naturally fuzzy. As in the previous article, the application was conceptually different from the proposed
in this article since the decision policies are still provided classically inside the SD model.
Finally, another recent integration proposal is found by Arasteh
and Aliahmadi (2014); Nasirzadeh, Khanzadi, and Razaie (2014)
and Khanzadi, Nasirzadeh, and Alipour (2012). These articles mix
in their SD models crisp variables, whose relations were obtained
from historical data, and fuzzy variables (imbedding expert knowledge), whose relations were obtained by fuzzy systems (simulating mental models of experts). Since not all the control decisions
were disaggregated in a FIS (that could provide crisp commands to
govern the model), these applications had to extend the models
arithmetic to deal with fuzzy numbers. In all these articles, their
goals were to use the SD model to simulate a specic output of
the models (a fuzzy output considering the extension principle).
Because of that, there was no focus on the modeling and testing of
different decision policies, which in all these cases was stated once
in the construction of the model.
After the revision of the theoretical and practical FL-SD integrations in the literature, it can be seen that there are no applications focused on virtual worlds learning. Consequently, the main

173

contribution of this article will be to propose a FL-SD integration

methodology specically to enhance learning with the use of virtual worlds. Since the proposal and testing of policies are the basis
of learning throughout experimentation in virtual worlds models,
it is essential to built white-box models, which means according
to Wieland and Gutzler (2014): simple enough to be understood
by stakeholders; constructed with mutual collaboration of experts
and stakeholders; and fast-prototyping to support fast interactive
modeling. Before presenting and exemplifying the methodology, it
will be introduced a theoretical discussion on system thinking and
learning to justify the integration proposal.
3. Theoretical discussion on system thinking and learning
Systems Dynamics (SD) is a methodological approach derived
from systems thinking, aiming to facilitate learning in the complex,
feedback, multi-loop, multi-states and non-linear systems in which
we live (Forrester, 1961).
According to Sterman (20 0 0), SD was originally developed to
promote and enhance the double loop learning. He stated that
the key to circumvent the diculties is the so-called virtual
worlds, thus encouraging the explicit and ongoing discussion of
mental models, which are not limited by analytical tractability.
They are based on realistic assumptions about human behavior and
use all types of data, available for specifying and estimating relationships (Forrester, 1987).
However, some of the practical barriers to the double
loop learning cannot be overcome simply by virtual worlds
applications. In fact, Sterman (20 0 0) pointed out that the use of
virtual worlds has some pitfalls, e.g., not being able to develop
the ability to scientic reasoning and to overcome deciencies in
group learning processes and the commonly veried video game
syndrome (closely related to the lack of scientic method).
Another barrier to learning in virtual worlds, mentioned by
Sterman (20 0 0), is that an individuals mental model is fuzzy, incomplete, loosely dened and variable in terms of time. Therefore,
each individual interprets certain content differently and adapts his
or her model particularly.
Scholars agree that, for effective group learning with SD, it is
necessary to declare and eciently test the assumptions, and to
communicate their results in a standardized manner (that way
reaching a group learning consensus). Otherwise, efforts would be
diffuse, and learning could be lost through some practices, such as
video game syndrome.
A primary cause of these problems is the fact that in most SD
models, decisions are either parameterized in a classical logic or
provided directly (manually) by the operator of the experiment.
The rst option is inconsistent with our real decision-making process (as mentioned, fuzzy by nature), and therefore, has little
practical value. In this sense, nearly three decades ago, Morecroft
(1988) predicted that the biggest and most challenging possible
collaboration in SD, would focus on using policy-maker knowledge
at a symbolic/conceptual level.
The second option, manual parameterization, is the main origin
of the lack of scientic method and video game syndrome problems, because the user (even if he or she does not change) by not
declaring explicitly his decision-making process (i.e., the policy hypothesis) can easily lose the learning focus. Another cause for these
problems stems from the inability to communicate (to the group)
in ordinary language the hypothesis of each experiment and its results (Oliva, 2003).
We believe that the use of fuzzy logic, with clearly-stated
linguistic rules, is the perfect solution to externalize all the
hypotheses and decision-making under development. As highlighted by Ross (2005), natural language is one of the most
powerful ways of conveying knowledge and information that

174

D.C. de Salles et al. / Expert Systems With Applications 55 (2016) 172183

Control
Indicators

Policys
hypotheses

System
dynamics
Model

Fuzzy policy
system

Dynamics
hypotheses

Control
Decisions
Fig. 1. Scheme of the integration of fuzzy system and SD model.

1. Problem Articulation
(Boundaries Selection)
1.a Theme Selecon
1.b Key Variables
1.c Time Horizon
1.d Dynamic problem denion

2. Dynamic
Hypotheses
2.a Inial hypothesis generaon
2.b Endogenous focus
2.c Mapping the structure

5. Formulation and
Evaluation of Fuzzy
Policies 5.a Set controlled variables
5.b Dene indicators and frequency
5.c Set Mamdani system

3. Formulation

4. Testing
4.a Comparison with reference modes
4.b Robustness under extreme
condions
4.c Sensivity

3.a Specify the structure and relaonships

3.b Esmate the parameters, behaviors,
relaonships and inial condions
3.c Test consistency with purpose and boundaries
Fig. 2. Methodology of cyclic FL-SD Troubleshooting.

humans have in relation to problems and situations involving

reasoning and decisions. Despite its vagueness and ambiguity, the
language makes individuals able to understand each other. Therefore, to model the process of human reasoning it is necessary to
emulate our natural language (and its linguistic variables).
The great advantage of FL over other approaches to support
decision-making is that its language is very simple, exible and
easily enough to be understood and communicated. Which are, as
mentioned, requirements to avoid black-box tools, and make the
model useful for the decision-maker and his team.

lier, and Schwaninger (1998) and by Kunsch, Springael, and Brans

(2001), consisting of technical support decision-making using SD
with multi-criteria decision tools (Kunsch & Springael, 2008).
The proposed separation does not change the traditional sequence of activities and thus allows us to preserve the structure
of the problem-solving methodology by Sterman (20 0 0), adapted
on Fig. 2.
We believe that this is the major contribution of this work and
therefore we will use a hypothetical case study to exemplify the
methodology step by step.

4. Proposed methodology
5. The case study: a hypothetical organization
In the proposed methodology, the system to be governed (modeled on SD stock and ow language) is separated from the human
decisions system (policies), which can now be modeled in a natural language using the fuzzy Mamdani system (Mamdani & Assilian, 1975) and formally declared, facilitating the application of the
scientic method and disciplining policies testing. Fig. 1 illustrates
this interaction.
This methodology can be included in a more general framework
called Adaptive Control Methodology (ACM) developed by Brans,
Kunsch, and Mareschal (2002); Brans, Macharis, Kunsch, Cheva-

According to Ghemawat and Cassiman (2007), and Gary, Kunc,

Morecroft, and Rockart (2008) to incorporate dynamics into our
strategy thinking is a great, perhaps the greatest, challenge that
the strategy eld faces to advance.
Another inuential author in the eld (Porter, 1991) also
pointed out that several frames can explain differences between
the undertakings at given moments in time, but our understanding of the dynamic processes that leads to a superior position in
the market is still very limited.

D.C. de Salles et al. / Expert Systems With Applications 55 (2016) 172183

175

Investment
+

Financial

Profit
Revenue

+ Price
Demand

Customers
-

Delivery Time

Internal
Processes

+ Propensity to
consumption

Production
Capacity

Costs

+
+

+
Productivity

Staff

Learning
And Growth

+
Processes
Improvement

Innovations

Fig. 3. Four layers and dimensions of system structure.

These quotations from highly respected authors in the eld

demonstrate the continued relevance of studying the dynamics of
strategic systems. For this reason and the great prevalence of studies in this area using SD, we decided to build a strategy model (and
a problem) to exemplify the methodology.
It should be noted, before beginning, that the objective of this
case study is to implement the methodology, keeping the model
simple but avoiding triviality.
Problem articulation (Boundaries selection):
(a) Dene the theme/problem: Company XYZ presently cannot
meet its demand and seeks to increase its prot by making
decisions on price and staff level to attract customers.
(b) Dene key variables: The key variables will be productivity,
innovation, process improvement, team size and the product
price.
(c) Set time horizon: Two and a half years (30 months)
(d) Dene the dynamic problem: Make decisions on pricing and
staff level to obtain greater prot for two and a half years.
5.1. Formulation of dynamic hypotheses:
(a) Generate the initial hypotheses:
The consumer decision is affected negatively by the delivery
time and the product price and positively by innovative products.
Investments in process improvements and innovations
have decreasing scale returns.
The company has xed labor costs, and the only variable
cost is material.
(b) Guarantee endogenous focus:
This aspect should never be underestimated; as Richardson
(2011) pointed out, this is a sine qua non condition of the SD
approach. In this case study that means to ensure that all
the elements explaining the behavior of demand and production problem were considered in the model. To do so, it

is assumed that the initial problem arose from previous innovation investments that made
the product attractive for
the market (considering its low price) and made the demand
to exceed the production capacity.
(c) Mapping system structure:
Fig. 3 shows the causal loop diagram
balance scorecard structure.

aligned with

the

5.2. Formulation of simulation models:

Specify the structure and relationship (non controlled) between
variables;
It should be noticed that in the original model of Sterman
(20 0 0), the relationships expressing decision rules were also dened at this stage, and now they are fully transferred to the fuzzy
controller. Completion of these steps means that only the natural
modus operandi of the system should be modeled, so the two steps
are:
(a) Estimate the parameters, behaviors, relationships and initial
conditions;
(b) Test consistency with purpose and system borders.
We modeled a exible and consistent behavior with the problem stated previously. Fig. 4 exhibits the stocks and ows model
built to simulate the problem.
Table 1 contains the formula and some explanations for all the
variables in the model.
Test phase:
(a) Comparison with reference behavior.
Also known as model calibration phase, it is an essential step
in modeling activity, especially in SD as a problem solving
approach. Validation is essential and represents an iterative

176

D.C. de Salles et al. / Expert Systems With Applications 55 (2016) 172183

Fig. 4. The stocks and ows structure of the proposed SD model.

process of gradual condence gain in the likelihood of the

model and its assumptions (Van Horn, 1971, cited by Saleh,
Oliva, Kampmann, & Davidsen, 2010). Depending on the specic problem, and the availability of studies and data, it can
be an extremely cumbersome step. Clearly the fuzzy integration has no effect on this step (it can even simplify it by
separating controls from the natural behavior of the system, but it depends on each case).
(b) Assess robustness against extreme conditions.
This step has the role of adjusting the model to a particular
referral behavior. Since the case is ctional we will simply present and explain the behavior of the out of control model. The Figs 57 show the behavior of its variables
grouped according to their balance scorecard position.
In Fig. 5, the nancial dimension, we note that, after initial oscillation, all indicators (earnings, sales and total cost) stabilize at some same level, for the obvious reason that all depend on the volume of shipment.
In Fig. 6, it can be seen that the propensity to consumption starts at very high levels, but quickly drops as delivery time increases. Production cannot meet 17 days to deliver the product, therefore wasting the potential of the market (keeping the propensity to consume permanently below
100). With this behavior, it is clear that the innovation is not
able to compensate the delivery delay in costumers opinion, despite its increase (as shown in Fig. 5). It should be
noted that the increase in process improvements (also seen

in Fig. 5) and the resulting productivity increase, are not able

to reduce the delivery time considering the demand they
face.
In Fig. 7, the internal processes dimension, we note that capacity utilization uctuates around its limit (100%), unable to
lower the high in-process inventories and consequently improve their service level (which requires to process the intermediate inventories and the new coming requests).
As we mentioned in the opening statement of the problem,
without changing the parameters, the production is saturated and unable to reduce its process inventory and delivery time, failing to meet the potential demand and resulting
in prot stagnation.
(c) Sensibility analysis:
To exemplify this item we varied the percentage split on investment, which remains xed in the original model. We tested four
proles of the evolution of this variable in time, and compared
their Net Present Value (NPV). The result is shown in Fig. 8.
As mentioned, it is evident that, in the model conguration, capacity is limited. Therefore, the increases in productivity are more
effective because they lead to a delivery time improvement and
consequently higher demand.
Innovation, on the other hand, stimulates demand directly, further deteriorating the delivery time, which reduces demand backwards, representing a balanced causal loop interaction. This explains why models that prioritized investments in process improvements achieved better results, compared on NPV terms.

D.C. de Salles et al. / Expert Systems With Applications 55 (2016) 172183

177

Table 1
Equations explanation for each variable in the stock and ow model.
FORMULAE
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24

OBSERVATIONS

Dt = 0.5
Innovation_Investment_percentual = 50 0 0/150 0 0
Innovation(t) = Innovation (tdt) + (Innovation_Investment) dt
INIT Innovation = 50 0 0
Innovation_Investment = (Investment / Innovation)
Innovation_Investment_percentual
Processes_Improvement (t) = Processes_Improvement
(tdt) + (Processes_Improvement_Investment) dt
INIT Processes_Improvement = 10 0 0 0
Processes_Improvement_Investment = (Investiment/Processes_Improvement)
(1Innovation_Investment_percentual)
Price(t) = Price (tdt) + (Delta_Price) dt
INIT Price = 50
Delta_Price = 0
Production(t) = Production (tdt) + (Orders - Expedition) dt
INIT Production = 0
Orders = 10 0 0 0 Propensity_to_ consumption / 100
Expedition = Productivity Staff
Staff(t) = Staff(tdt) + (Delta_Staff) dt
INIT Staff = 20
Delta_Staff = 0
Labor_Cost = Staff 2500
Material_Cost = 5
Total_Cost = (Expedition Material_Cost) + Labor_Cost
Investment = 0.5 Prot
Prot = RevenuesTotal_Cost
Productivity = (450 Processes_Improvement)/10 0 0 0

25

Propensity_to_ consumption = IF TIME > 2 THEN 100 (50/Price)

Delivery_Time) (Innovation/50 0 0) ELSE 100

26
27

Delivery_Time = CTMEAN(Expedition) 30
Capacity_Utilization = (Orders/Capacity) 100

28
29

Revenue_Variation = DERIVN(Vendas,1)
Delta_Revenue = IF Revenue = 0 THEN 0 ELSE (Revenue_Variation/Revenue)
100
Revenue = Expedition Price
NPV = NPV(Prot,0.005)

30
31

(17/

5.3. Formulation and evaluation of fuzzy policies

The fuzzy system must receive from the SD model only the indicators that will be used to make decisions, returning command
orders to the SD model. This communication interaction is held at
every delimited control interval, which can be equal to the simulation interval or not.
In systems that reproduce some reality, policies can arise naturally from normal practices. In the present case, since the system
is hypothetical, its response to different parameters is dicult to
predict.
The policy selection, whether dened traditionally (in the
model itself or manually) or in a separate fuzzy system, generally requires evaluation of multiple possibilities, involving many
parameters, and at several levels. According to Saleh et al. (2010),
the approaches of exploratory scenarios (what-if scenarios) and exhaustive testing by experts, are the most common ways to generate and evaluate policies in models of dynamic systems. Others,
like Kleijnen (1995), used automated optimization software to analyze the parameters in search for a better solution given an objective function. Also in the same work, the authors presented a technique to connect the behavior of the system structure to its parameters, suggesting a systematic approach to evaluating policies. The
diculty to implement this proposal is beyond the scope of this
work, and we will use the traditional testing methods.
It is noteworthy that, within this methodology, experiments
and impressions-based decisions are not only possible but

An example of a classic parameterized policy (the investment split)

As mentioned, with decreasing scale returns

As mentioned, with decreasing scale returns

Without control is null

The initial full propensity demand is 10 0 0

Rule determining that half of the prot is always directed to investments

Initial productivity of 450 being proportionally increased by the process
improvement stocks growth
This formula was built based on the previously declared proportionality, as can
be seen inverse to price (with initial price 50 for 100% satisfaction) and
delivery time (with 17 days as a 100% satisfaction target) and positively with
product innovation (with the initial innovation stock value as 100%
satisfaction)
Native STELLA formula transformed into daily basis (ISEE Systems, 2009)
A non-traditional manner to calculate utilization to measure the capacity of
the production to process in good timing new orders
The normalized revenue variation indicator

A native STELLA functions with a nominal 0.01 monthly interest rate

recommended, since the goal is not to obtain the single optimal

policy, but to learn more about the system and policies possibilities to govern it.
Despite the need for testing, it should be emphasized that the
greater the understanding of the model and the inuence of each
parameter on its behavior, the more skilled is the modeler to set a
policy. For this reason, the previous model testing step is crucial to
the quality of the policies to be proposed at this stage.
Following, we will present the setup and results of only one
controller.
(a) Dene the controlled variables:
The controlled variables will be the price changes and the size
of the staff (hiring/dismissal).
(b) Dene indicators for monitoring and its frequency:
The indicators will be the Capacity Utilization and the Delta
Revenue (or sales percentage variation). The intention was
to maintain increasing sales (depending on demanded quantity and price), and the utilization below its saturation
(to ensure satisfactory delivery time). The measurement frequency was one per month (two ranges of simulation).
We chose these two indicators because the utilization is a good
indicator to access the quality of delivery and also idleness
(or excessive size of staff). And the sales variation is a good
indicator to motivate changes in price and/or staff selection,

178

D.C. de Salles et al. / Expert Systems With Applications 55 (2016) 172183

Fig. 5. The nancial dimension behavior.

Fig. 6. The propensity to consumption and delivery time.

assuming that at some point there would be an ideal balance between price level and team size that would not allow
more signicant sales growth.
(c) Dene fuzzy controller or fuzzy policy system:
The fuzzy system best suited to implement rules in natural language is the Mamdani system (Mamdani & Assilian 1975). To
build such system it must be considered the following items:

Membership functions of inputs

To start, it should be dened the domain of each of the
indicators. This can be obtained by normalizing the calculation of the indicators or experimentally study the model to
determine the value of variation bands. In real applications,
if the model is adjusted to reference behaviors, this denition would be less problematic. Fig. 9 exhibits the chosen

D.C. de Salles et al. / Expert Systems With Applications 55 (2016) 172183

179

Fig. 7. Internal processes behavior.

Fig. 8. Proles of sensitivity analysis and its results.

Fig. 9. Staff utilization member function.

membership functions associated with their linguistic terms.

As can be seen, they are triangular or trapezoidal for simplicity and because there is no specic motivation for other
prole.

Membership functions of outputs

As in the previous step, it is necessary to dene the possible

values that outputs could take. Unlike the inputs that are results of
the model, these parameters are exogenous to the model, and in
most cases, can assume values at the policy-makers will.
Even in these situations, the realism of these values should be
taken into account. You cannot release certain values that do not

make sense or viability in real life, because you can lose all practical value of learning.
We dened, for the price control, a $10 range of variation. The
changes of staff, for reasons merely subjective, was set to layoff not
exceeding ve employees and hire not above ten employees (in an
interval of 30 days).
As can be noted in Fig. 9 the Capacity Utilization was considered medium close to 70% (steeper for larger values) and this will
be the control target.
Fig. 10 highlights the range mentioned and price maintenance
between $2. In Fig. 11, it can be seen that the Delta Revenue was
dened very sensitive to the variations, considering null between

180

D.C. de Salles et al. / Expert Systems With Applications 55 (2016) 172183

Fig. 10. Membership function for price range variation.

Fig. 11. Membership function for revenue variations.

3% and positive/negative triggered from 2.5%. As can be noticed

from Fig. 12, we built one more membership function for staff variation granting more commands options.
The rules are self-explanatory (Table 2) and it can be seen two
well-dened reactions. The rst, when utilization is high, there
should be hiring; and when it is low, there would be dismissals.
The second, the prices are used to control demand, while seeking
for better prot levels.
5.4. Dashboard of results:
The same graphs for the non-controlled model are shown in
Fig. 13, additionally with a graph showing the control decisions
made by the fuzzy system.
As can be noticed in Fig. 13, the propensity to consume has an
initial drop driven by price increases and then it goes back to an
upward trajectory, drawn by a delivery time close to its minimum
(15 days, or a range of simulation) and the growth of product innovation.
It can be noted that the capacity utilization is maintained at
around 70%, as desired; and also, as a result, process inventories
remain at a lower level.
The dimension of learning and growth shows the same upward trajectory, since they are modeled as stocks, and, in this

case, growing at a faster pace than in the baseline behavior, although the graph scale does not allow the comparison
to Fig. 5.
At the nancial dimension, it should be noted that the total
cost now behaves desynchronized with prot and sales, because its
xed portion changes with the workforces hiring and ring. Also,
note that the prot presented an increasing pattern with the
increase in the propensity to consume and the price level. For
comparison, the NPV of this model was $1.132 107 , a result considerably higher than the model without control (shown in the
sensitivity analysis, Fig. 8).
Interestingly, after achieving half of the simulation, the control
was stabilized keeping the price around $66, and a workforce of 27
employees.
This hypothetical case can give some managerial insights on
companies releasing new innovative products. As usual the product is initially offered at a relative low price (with low prot margin) to foment demand, however as the demand grows at fast rate
the enterprise should increase their capacity (in this case related
uniquely to staff size) to avoid deterioration in delivery service.
Parallel to the capacity increase, this study suggest that the price
could be raised gradually (as the innovation and good delivery service guarantee an eager demand), until a stabilization level where
prot margins are optimized without reducing the demand for the

D.C. de Salles et al. / Expert Systems With Applications 55 (2016) 172183

181

Fig. 12. Membership function for staff variation.

Table 2
Logical rules for fuzzy modeling membership functions.
IF

Capacity utilization IS

1
2
3
4
5
6
7
8
9

Low
Low
Low
Medium
Medium
Medium
High
High
High

AND

Delta revenueIS
Positive
Null
Negative
Positive
Null
Negative
Positive
Null
Negative

THEN

Delta price IS

Delta staffIS

Keep
Reduce
Reduce
Keep
Keep
Reduce
Raise
Raise
Raise

Reduce
Reduce
Reduce
Reduce
Keep
Keep
Raise
Raise
High Raise

Fig. 13. Dashboard of the results of the Fuzzy-DS model of scorecard.

182

D.C. de Salles et al. / Expert Systems With Applications 55 (2016) 172183

innovative product (that also evolutes with R&D investments that

adds functionalities to the original product).
6. Discussion and conclusions
The structure of the proposed methodology and the objectives of integration could be satisfactorily exemplied by the case
study. As sustained in the theoretical discussion section, we believe
that Fuzzy Logic (FL) is a more appropriate tool to express business decisions (policies), accurate with the decision-makers way
of thinking, in natural language, and perfectly complementary to
the System Dynamics (SD) modeling approach.
As discussed, the proposed methodology is different from all
the other FL-SD integrations because it focuses on the learning
process with virtual worlds. This focus reects on some specicities of the proposal like the complete segregation of the control
policies from the to-be controlled model and the urgency to keep
it a white-box model.
As shown, the proposed separation of the fuzzy control system
does not bring signicant additional diculties to the traditional
method proposed by Sterman (20 0 0) for solving problems with
system dynamics tools, which can be easily adapted to a FL-SD
integrated methodology.
The methodology, by forcing the modeler to explicitly state their
assumptions and mental models of policies, strikes away regular
problems of deploying virtual worlds, such as, the video game syndrome, the lack of scientic method, and the diculties to discuss
and communicate policies, as well as showing their results at team
level.
In practical terms, the separation does not bring more complexity to the implementation phase, because the tools provided by
MatLab (The MathWorks Inc., 2013) and STELLA (ISEE Systems,
2009) and their interfaces are extremely user friendly, keeping the
modeled problem accessible and avoiding to intimidate the user
with black box tools.
The main limitation of this methodology is that FL is not always the most suitable way to express decision policies, it is
only appropriate when decision policies could be declared in
natural language involving fuzzy variables. This is normally the
case of business strategic rules but not the case for engineering systems. For example, Polat and Bozdag (2002) compared
the performance of a fuzzy controller with other classical controllers and concluded that there were no indications that a FL
was a better a priori alternative. Reinforcing the argument that
FL should be chosen for its theoretical/conceptual suitability, if
this is not the case, one should consider which control technique has the best performance for the particular system being
evaluated.
Secondly, as previously discussed, learning on virtual worlds
is dependent on policy design and testing. Although the FIS controller disciplines and favors these activities, when experimentation is cumbersome it could be necessary to use a numerical
method to search for better solution policies (as in the MOP proposals cited in the literature review). It is noteworthy that there
is a risk that these additions wipe out the simplicity needed
for stakeholders involvement, an interesting analysis for further
studies.
Finally, this proposal opens several opportunities to test the
methodology in other complex and real problems, especially strategic related, involving larger teams discussions. Possibly, within
more complex models, it wont be viable to carry out exhaustive
testing and scenario analysis practices. However, real life expertise
tends to reduce the range of alternative policies to be tested. In
these situations we believe that the FL user-friendly implementations will play a decisive role to facilitate discussion and share
knowledge about SD control policies.

References
Arasteh, A., & Aliahmadi, A. (2014). A proposed real options method for assessing
investments. The International Journal of Advanced Manufacturing Technology, 70,
13771393.
Baradaran, V., & Keshavarz, M. (2015). An integrated approach of system dynamics simulation and fuzzy inference system for retailers credit scoring. Economic
Research-Ekonomska Istraivanja, 28, 959980.
Brans, J. P., Macharis, C., Kunsch, P. L., Chevalier, A., & Schwaninger, M. (1998). Combining multi-criteria decision aid and system dynamics for the control of socio-economic processes: an iterative real-time procedure. European Journal of
Operational Research, 109, 428441.
Brans, J. P., Kunsch, P. L., & Mareschal, B. (2002). Management of the future: a system dynamics and MCDA approach. In D. Bouyssou, E. Jacquet-Lagrze, P. Perny,
R. Slowinski, D. Vanderpooten, & Ph. Vincke (Eds.), Aiding Decisions with Multiple Criteria: Essays in Honour of Bernard Roy (pp. 483502). Kluwer Academic
Publishers, Dordrecht, the Netherlands.
Campuzano, F., Mula, J., & Peidro, D. (2010). Fuzzy estimations and system dynamics
for improving supply chains. Fuzzy Sets and Systems, 161, 15301542.
Chang, Y., & Ko, T. (2014). An interactive dynamic multi-objective programming
model to support better land use planning. Land Use Policy, 36, 1322.
Coyle, R. G. (1996). System dynamics modelling: a practical Approch. CRC Press.
Forrester, J. W. (1961). Industrial dynamics. Cambridge, MA: MIT Press.
Forrester, J. W. (1987). 14 obvious Truths. System Dynamics Review, 3, 156159.
Gary, M. S., Kunc, M., Morecroft, J. D. W., & Rockart, S. F. (2008). System dynamics
and strategy. System Dynamics Review, 24, 407429.
Ghemawat, P., & Cassiman, B. (2007). Introduction to the special issue on strategic
dynamics. Management Science, 53, 529536.
Inc, TheMathWorks (2013). MATLAB. Natick, Massachusetts, U.S.A.
ISEE systems. (2009). STELLA. Version 9.1.
Karavezyris, V., Timpe, K. ,, & Marzi, R. (2002). Application of system dynamics and
fuzzy logic to forecasting of municipal solid waste. Mathematics and Computers
in Simulation, 60, 149158.
Khanzadi, M., Nasirzadeh, F. ,, & Alipour, M. (2012). Integrating system dynamics and
fuzzy logic modeling to determine concession period in bot projects. Automation
in Construction, 22, 368376.
Kleijnen, J. P. C. (1995). Sensitivity analysis and optimization of system dynamics
models: regression analysis and statistical design of experiments. System Dynamics Review, 11, 275288.
Kunsch, P. L., Springael, J., & Brans, J. P. (2001). An adaptive control methodology
based on system dynamics and MCDA case study: the CO2 energy tax in the
residential sector. In A. Colorni, M. Paruccini, & B. Roy (Eds.), Aide Multi Critre
la Dcision/Multiple Criteria Decision Aiding, Report of the Commission of the
European Communities, Joint Research Centre JRC-ISPRA, EUR - report 19808 EN
(pp. 145162).
Kunsch, P. ,, & Springael, J. (2008). Simulation with system dynamics and fuzzy reasoning of a tax policy to reduce CO2 emissions in the residential sector. European Journal of Operational Research, 185, 12851299.
Mamdani, E. H., & Assilian, S. (1975). An Experiment in linguistic synthesis with a
fuzzy logic controller. International Journal of Man-Machine Studies, 7, 113.
Morecroft, J. D. W. (1988). System dynamics and microworlds for policymakers. European Journal of Operational Research, 35, 301320.
Nasirzadeh, F., Khanzadi, M. ,, & Razaie, M. (2014). Dynamic modeling of the quantitative risk allocation in construction projects. International Journal of Project
Management, 32, 442451.
Oliva, R. (2003). Model calibration as a testing strategy for system dynamics models.
European Journal of Operational Research, 151, 552568.
Orji, I. J., & Wei, S. (2015a). An innovative integration of fuzzy-logic and systems
dynamics in sustainable supplier selection: a case on manufacturing industry.
Computers & Industrial Engineering, 88, 112.
Orji, I. J., & Wei, S. (2015b). Dynamic modeling of sustainable operation in green
manufacturing environment. Journal of Manufacturing Technology Management,
26, 12011217.
Pankaj, S., & Sushil, K. (1994). A fuzzy set theoretic approach to qualitative analysis
of causal loops system dynamics. European Journal of Operational Research, 78,
380393.
Polat, S., & Bozdag , C. F. (2002). Comparison of fuzzy and crisp systems via system
dynamics simulation. European Journal of Operational Research, 138, 178190.
Porter, M. E. (1991). Towards a dynamic theory of strategy. Strategic Management
Journal, 12, 95117.
Richardson, G. P. (2011). Reections on the Foundations of System Dynamics. System
Dynamics Review, 27, 219243.
Ross, T. J. (2005). Fuzzy logic with engineering applications. Hoboken, NJ, USA: John
Wiley & Sons.
Sabounchi, N. S., Triantis, K. P., Sarangi, S., & Liu, S. (2014). Dynamic simulation modeling and policy analysis of an area-based congestion pricing scheme
for a transportation socioeconomic system. Transportation Research Part A, 59,
357383.
Saleh, M., Oliva, R., Kampmann, C. E., & Davidsen, P. I. (2010). A comprehensive analytical approach for policy analysis of system dynamics models. European Journal
of Operational Research, 203, 673683.
Song, J., Song, D., & Zhang, D. (2015). Modeling the concession period and subsidy
for bot waste-to-energy incineration projects. Journal of Construction Engineering
and Management, 141, 04015033.
Sterman, J. D. (20 0 0). Business Dynamics: Systems Thinking and Modeling for a Complex World. Irwin/McGraw-Hill.

D.C. de Salles et al. / Expert Systems With Applications 55 (2016) 172183

Van Horn, R. L. (1971). Validation of simulation results. Management Science, 17,
247258.
Wieland, R., & Gutzler, C. (2014). Environmental impact assessment based on dynamic fuzzy simulation. Environmental Modelling & Software, 55, 235241.
Wu, Z., & Xu, J. (2013). Predicting and optimization of energy consumption using
system dynamics-fuzzy multiple objective programming in world heritage areas.
Energy, 49, 1931.
Xu, J., & Li, X. (2011). Using system dynamics for simulation and optimization of one
coal industry system under fuzzy environment. Expert Systems with Applications,
38, 1155211559.

183

Xu, J., Deng, Y., & Yao, L. (2014). Sustainable development-oriented industrial restructuring modeling and analysis: a case study in Leshan. Clean Technologies
and Environmental Policy, 16, 267279.
Zadeh, L. A. (1975a). The concept of a linguistic variable and its application to approximate reasoning-I. Journal of Information Science, 8, 199245.
Zadeh, L. A. (1975b). The concept of a linguistic variable and its application to approximate reasoning-II. Information Sciences, 8, 301357.