To Radicalize or to Deradicalize? Is that the Question?
Erik Pruyt* (TU Delft) & Jan Kwakkel (TU Delft)
August 17, 2011
Abstract.
Radicalization and deradic
lization are deeply uncertain dynamic proc Exploring
and analyzing many plausible futures and assessing the robustness of policies to reinforce
desirable evolutions seem more useful for such processes than trying to predict their precise
development over time and optimize the associated policy response. This paper illustrates
how the combination of System Dynamics Modeling and Exploratory Modeling and Analysis
could be helpful for exploration and decision making in case of deeply uncertain dynamic
issues such as de/radicalization processes. In this paper, different System Dynamics models
about radical and non-radical activism are presented and analyzed, both separately and jointly,
but always under deep uncertainty. The different models are treated as alternative dynamic
assumptions about how/why activism may become more or less extremist/harmful. These
dynamic assumptions are included as structural/model uncertainties and are combined with
many other uncertainties in order to generate a large ensemble of plausible scenarios. Finally,
this ensemble of plausible scenarios is used to test policies for fighting radicalization under deep
uncertainty. The ‘hot teaching and testing cases’ related to the models used for this analysis
are provided in the appendiz.
Keywords: Exploratory System Dynamics, Deep Uncertainty, EMA, ESDMA, Radi-
calization, Deradicalization
1 Introduction
1.1 Radicalization & Deradicalization
Radical activism and (further) radicalization are complex and uncertain risks: these shambering
phenomena could be self dissolving, unsolvable or could be the potential breeding ground for acute
and/or chronic cris
fall of radical versus non-radical actions against dictatorial regimes in Tune:
es (Pruyt 2010b). De/radicalization of animal rights activism and the rise and
, Egypt, Lybia, etc.
are interesting examples of such unpredictable processes. It is precisely the combination of dynamic
complexity and deep uncertainty that makes them unpredicable but nevertheless ‘explorable’.
Dynamic complezity is found in situations where cause and effect are subtle, and where the
effects over time of interventions are not obvious (Senge 1990). It arises, among else, from the
complex interactions of systems, markets, institutions, products, regulators, (groups of) actors,
and policies/regulations. And deep uncertainty is found in situations where analysts do not know,
or the parties to a decision cannot agree on (i) the appropriate conceptual models that describe
the relationships among the key driving forces that will shape the long-term future [e.g. dif:
ferent drivers and underlying structures than today], (ii) the probability d
represent uncertainty about key variables and parameters in the mathematical representations
of these conceptual models, and/or (iii) how to value the desirability of alternative outcomes
(Lempert, Popper, and Bankes 2003).
‘ibutions used to
“Corresponding author: Erik Pruyt, Delft University of Technology, Faculty of Technology, Policy and Manage-
ment, Policy Analysis Section; P.O. Box 5015, 2600 GA Delft, The Netherlands — E-mail: e.pruyt@tudelft .n1
Pruyt & Kwakkel, 2011. To Radicalize or to Deradicalize? In: Proc. Int. Conf. SD Society — 2
1.2 Methodology: ESD x EMA = ESDMA
Conventional forecasting, planning, and analysis
complexity (Senge 1990) and even less so for dealing with deep uncertainty: prediction of dynamic
behaviors and certainty about probabilities, validity, and optimality cannot be obtained for (fu-
ture) multi-dimensional systems characterized by high degrees of dynamic complexity and deep
uncertainty. Moreover, improving models by increasing the level of detail or their size does mostly
not help much: after all, ‘all models are wrong [but] some are just more useful’ (Sterman 2002).
It may even be harmful if there is little time to act or uncertainties cannot be reduced, since it is
very time-consuming and may generate the illusion that all unce d.
Instead of focussing on predictability, optimality, and attempting to develop ever more detailed
models validated upon past conditions, it may be more useful to develop alternative models, explore
different model formulations and a plethora of uncertainties, and test effectiveness and robustness
of policies in the face of these parametric and structural uncertainti
methods are not suited for dealing with dynamic
ainties can be and are redu
s (SD) is used for modeling and simulating dynamically com-
plex issues and analyzing their resulting non-linear behaviors over time in order to develop and
the effectiveness of structural policies. Mainstream System Dynamicists have assumed for
ies are omnipresent, and hence, that trajectories generated with SD simu-
lation models should not be interpreted quantitatively as point or trajectory predictions, but that
they should be interpreted qualitatively as general ‘modes of behavior’. And although univariate
sensitivity analysis and multivariate sensitivity analysis are mostly performed, they are mainly
aimed at validation close to the base case behavior — not exploration over the entire uncertainty
space. It scems therefore that in traditional SD the omnipresence of uncertainties is accepted but
that they are not really explored or explicitly taken into account.
However, SD mod so be built specifically for the purpose of exploring the potential
s on dynamically comp nes. Such Exploratory System Dynamics
(ESD) models are preferably fast-to-build and casily-manageable models, and consequently, rather
simple and highly aggregated. ESD is an interesting approach for exploring uncertainties, and
testing the effectiveness of policies in the face of these uncertainties. However, ESD in isolation
may be insufficiently broad and systematic to firmly base policymaking under deep uncertainty
on.
But the combination of ESD with Exploratory Modeling and Analysis (EMA — a methodology
for exploring deep uncertainty and testing the robustness of polici subsection 2.2) may be
useful and sufficient for broadly and y exploring and analyzing plausible dynamics
under deep uncertainty, and for testing the effectiveness and robustness of policies without ne-
glecting deep uncertainty and dynamic complexity. EMA consists of using exploratory models
(not necessarily SD models) for generating tens of thousands of scenarios (called an ‘ensemble of
future worlds’) in order to analyze and test the robustness of policy options across this ensemble
of future worlds ~ in other words whether the outcomes are acceptable for all transient scenarios
generated by sweeping the entire multi-dimensional uncertainty space. As such, it can be used
to generate insights and understanding about the functioning of systems and the robustness of
policies, by taking deep uncertainty seriously into account (Lempert, Popper, and Bankes 2003)
(Agusdinata 2008). In EMA, the question is not ‘when to measure more’ nor ‘when to model
better’, but how to explore and analyze dynamically complex systems under deep uncertainty’
and ‘which policies do effectively and robustly improve system behavior under deep uncertainty’.
Since EMA requires handy models for generating (thousands of) plausible scenarios, and
ESD requires methods for exploring deep uncertainty, they are actually natural complemen-
tary allies (Pruyt 2007), and could be combined as Exploratory System Dynamics modeling and
Analysis (ESDMA). Examples of ESDMA can be found in (Lempert, Popper, and Bankes 2003),
(Pruyt 2010b; Pruyt and Hamarat 2010a; Pruyt and Hamarat 2010b), and in these proceedings
in (Pruyt, Kwakkel, Yucel, and Hamarat 2011; Pruyt, Logtens, and Gijsbers 2011; Pruyt 201a;
Kwakkel and Slinger 2011).
influence of uncertainti
Pruyt & Kwakkel, 2011. To Radicalize or to Deradicalize? In: Proc. Int. Conf. SD Society — 3
1.3 De/Radicalization and ESD & ESDMA
As stated above: De/radicalization processes are mostly too complex and uncertain to be pre-
dicted whether or not models are used. Models may nevertheless be useful to explore the extent
of plausible evolutions. One way to generate different plausible evolutions is to focus on uncer-
tainties from start to finish and sweep the entire uncertainty space. The enormous amount of
scenarios generated that way can often be clustered. Similar patterns may then be represented
by one representative scenario. Both the full set of scenarios and the reduced set of scenarios
may then be used to gain a better understanding of plausible evolutions. A better understanding
of those plausible evolutions may lead to better plausibly counter-intuitive!— policies. And the
appropriateness of these policies may as well be tested over the entire ensemble of scenarios. This
is what will be illustrated in this paper.
1.4 Organization
First, the methodology is further explained in section 2. Second, a first de/radicalization model
is introduced and used for ESDMA in section 3. Then, a second de/radicalization is introduced
and used for ESDMA in section 4. An ESDMA on both models and various structural variants
is presented in section 5. Finally, some concluding remarks are formulated in section 6. ‘Hot’
teaching and testing cases related to the SD models are included and referred to in appendix A.
2 Methodology: Exploratory System Dynamics Modeling
and Analysis
2.1 Exploratory System Dynamics (ESD)
Exploratory System Dynamics (ESD) refers to the development and use of fast-to-build and ee
to-use SD models for ‘quick and dirty’ exploration of (a plethora of) possible behaviors and plau-
sible scenarios, and for developing a rough idea about the effectiveness and to a lesser extent the
robustness~ of potential policies
uch, it may be used to quickly explore plausible behaviors with a relatively simple model
-or relatively simple models~ without explicitly aiming to uncover the structures underlying the
real issue and forecasting the behavior or the probabilities of the behaviors.
The focus of ESD lies then on testing whether behaviors of interest (c.g. plausible trajectories
that require attention) can be generated (at all), on exploring plausible types of behaviors, and on
identifying policies that may be effective. Traditional tools such as scenario anal:
and what-if analysis can be used in ESD too, but in view of a different goal than in traditional
SD (exploration instead of model validation).
ESD could be used for generating interesting behaviors, plausible trajectories that require
attention, test whether specific behaviors can be generated at all, etc. It cannot be used to make
predictions or derive firm conclusions about specific values or sets of conditions that may lead to
this or that behavior, or to interpret outcomes in a probabilistic sense. Issues for which ESD is
useful are just too uncertain and complex, and that the ESD models developed are just too wrong
for any form of prediction (but also more useful for the purpose of exploration). Hence, ESD may
be good as an introduction to dealing with complex and uncertain issues, especially for imagining
plausible modes of behavior; it is not useful for detailed analysis in view of detailed implementation
or to (totally) reduce uncertainties. ESD modeling is thus most appropriate for quick and dirty
modeling if time is at a premium and for assumption based modeling if uncertainties are too deep
for traditional approaches.
isk analysis
1 Counter-intuitive policies in the case of de/radicalization may be (i) less control instead of mor
changing the (entire) population instead of forcing radical groups to change, (iii) to spend resoure
and expectation management of radical groups,
ii) gradually
on perception
Pruyt & Kwakkel, 2011. To Radicalize or to Deradicalize? In: Proc. Int. Conf. SD Society — 4
Although ESD in isolation may be appropriate for some issues @ , it may also
be insufficiently systematic for well-founded decision support for iss sed by deep
ESD may still be useful for that purpose, but then as a scenario generator for EMA.
uncertainty.
2.2 Exploratory modeling and Analysis (EMA)
Exploratory modeling and Analysis (EMA) can be used (i) to explore the influence of uncertain-
ties”, and (ii) to test the effectivene:
EMA con:
issue of interest, (ii) generating an ensemble of future worlds (thousands to millions of
) by sweeping uncertainty ranges and varying uncertain structures and boundaries, (iii)
simulating and analyzing the dynamic behaviors, bifurcations, et cetera, (iii) and/or specifying a
variety of policy options (preferably adaptive ones), and simulating, calculating, and comparing
the performance of the various options across the ensemble of future worlds.
Although data analysis techniques (step (iii)) could be used to investigate the effect of under-
lying mechanisms/(inter)actions/conditions, to separate different modes of behavior, to determine
the conditions that lead to these different modes of behaviors, to find bifurcation points and cri
1 variables, it may be even more interesting to define different (adaptive) policies/strategies
immediately test their (relative and absolute) effectivene: uncertainties
(step (iv). The effectiveness/robustness of policies can then be evaluated over the entire multi-
dimensional uncertainty space without needing to analyze/understand millions of outcomes. In
other words, the effectiveness /robustness of policies/strategies could be evaluated and compared
without reducing uncertainties related to the system of interest and without getting overwhelmed
by combinatorial complexity.
EMA is still under development: researchers of several institutes are currently improving,
extending and contributing to EMA theory, EMA methodology, and EMA tools, and are working
on a plethora of EMA applications, often in combination with adaptive policymaking’. Currently,
special attention is paid to the combination of ESD and EMA in order to ease their combined
application to deeply uncertain dynamically complex issues.
and
2.3. Exploratory System Dynamics modeling and Analysis (ESDMA)
In ‘Exploratory System Dynamics modeling and Analysis’ (ESDMA), the fast and relatively simple
models used to generate ensembles of future worlds are more specifically ESD models. Since EMA
is appropriate for systematically exploring deep uncertainty and testing the robustness of polic
and ESD models are particularly appropriate for generating plausible behaviors over time, it follows
that ESDMA is particularly appropriate for systematically exploring and analyzing thousands to
millions of plausible dynamic behaviors over time, and for testing the robustness of policies over
all these scenarios. As opposed to EMA, ESDMA therefore goes beyond the calculation of end
states or static values, which is in fact a very serious difficulty, that can be side-stepped in this
paper.
Since EMA is a quantitative uncertainty analysis approach and mainstream SD modeling con-
of making quantitative simulation models, it follows that ESDMA is mainly a quantitative
multi-method, but which leads —just like traditional SD- to qualitative interpretations, isions
2related to initial values, parameters, specific variable formulation, generative structures, model formulations,
model boundaries, different models, different modeling methods and paradigms, and different. preferences and
perspectives related to different world-views, and policies
3For RAND related EMA work see for
ample (Bankes 1993; Lempert and Schlesinger 2000),
(Lempert, Popper, and Bankes 2003; Lempert, Groves, Popper, and Bankes 2006; Bryant and Lempert 2009).
For TUD/RAND work, see (Bankes, Walker, and Kwakkel 2010). For Delft University of Technology related EMA
work see (Walker and Marchau 2003; Van der Pas, Agusdinata, Walker, and Marchau 2007; Agusdinata 2008),
(Van der Pas, Agusdinata, Walker, and Marchau 2008) (Agusdinata, Van der Pas, Walker, and Marchau 2009)
(Pruyt 2010b; Pruyt and Hamarat 2010a; Pruyt and Hamarat 2010b) (Kwakkel, Walker, and Marchau 2010a;
Kwakkel, Walker, and Marchau 2010b; Kwakkel and Slinger 2011), and (Pruyt, Kwakkel, Yucel, and Hamarat 2011;
Pruyt, Logtens, and Gijsbers 2011; Pruyt 2011a)
Pruyt & Kwakkel, 2011. To Radicalize or to Deradicalize? In: Proc. Int. Conf. SD Society — 5
and recommendations.
3 The First De/Radicalization Model
The model presented and used in this
section is just a combination of micro-hypotheses about
de/radicalization. The model is high-level, simple, generic, and most certainly ‘wrong’. This
model and other formulations of the same issue may however be useful. The model is pr
subsection 3.1 (one may want to skip this relatively dry model description), its behavior is
analyzed in subsection 3.2, and it is used for ESDMA in subsection 3.3.
nted in
quickly
3.1 Model Structure of the First De/Radicalization Model
frit vabe of
dering phenomena =F problem syntons ———— vsdity ofthe sitacenet
: problem“ the usbily ofthe
ita eadiness to take # — problem tough
action per contact re etc rte ofinerease of ~ radkal actions
ea the tdrh a phenomenon oy
“ renfirsenet ofthe
sbi ofthe problem
problemi hed se ei ppt ‘through radkal and
phenomenon evel 1 wetinerease ofthe aa ot-radical actions
ij aver phesomenon = s
ae rate of decrease _ aa ‘ _ Ay (cam \ = a
ataimblerateof___.yeuph sockual change Tourn fnconviced) etna mats
decrease rows ~ size | eravon_ L_stiess_| [Seirieed |“
societal change 4 te kre
Lg 2 OS somal
id Srstation due to y \ ~\prestasieness
percved magne nargnaization i /
‘ofthe problem y \ :
\ strength of coniact ate of presvasiness of
\ / compat comedies nomsaalactons
\ / selfinerest x
\ le / \
\ average degre of \ / \ \
\ noone orenboed vomalcantsaeor |
\ a Cee” | | /
\ a \ he radicalaction vel
\ average readiness
readiness to take action per \
‘ewy cominced cen take action
I/ S
rent readiness to take action of
Tastatin of the newky convinced citizens
commeed czens
iniialmnber of
cominced etizens
intial amber of
unconvinced citizens
Figure 1: Stock-Flow Diagram of the generic de/radicalization model
In this generic model, it is assumed that radicalization is brought about by a (real or perceived)
slumbering underlying phenomenon (see the Stock-Flow structure at the top left of Figure 1). In
this version of the model, it is also assumed that problem
ymptoms become visible when/if the
level of the underlying phenomenon increases above a particular level. Here, those problem symp-
toms are modeled simplistically as the non-negative difference between the underlying phenomenon
and the problematic phenomenon level, over the problematic phenomenon level.
But citizens that are convinced of the problem —even if there are no clear problem symptoms
visible to the rest of the population yet— perceive the problem: this is modeled here as the fraction
of the underlying phenomenon over the problematic phenomenon level. The bigger the perceived
problem is, the higher the frustration of those convinced citizens will be. The frustration of
convinced citizens -expressed as a percentage between 0% and 100%- is modeled here as the
product of the frustration due to marginalization, the frustration due to inertia, the average
degree of conviction of convinced citizens, and the perceived magnitude of the problem.
Pruyt & Kwakkel, 2011. To Radicalize or to Deradicalize? In: Proc. Int. Conf. SD Society 6
In this model, unconvinced citizens become convinced citizens if/when persuaded by convinced
citizens. However, there are also citizens that will never be convinced — they belong to the radicals
at the other end of the spectrum. This persuation is modeled here as the product of the contact rate
‘itizens, the fraction of convinced citizens, the number of unconvinced citizens,
sibility of the problem (see below), divided
by the strength of their incompatible self-interest. The frustration due to marginalization is a
function of the fraction of convinced citizens
More convinced citizens leads —ceteris paribus~ to more societal change. S
allow to stop the increase of, or even decrease, the underlying phenomenon
the rate of decrease through societal change equals the fraction of convinced citizens multiplied by
the maximum attainable rate of decrease through socictal change. The rate of decrease through
societal change mitigates the net increase of the underlying phenomenon. The net increase of the
underlying phenomenon is equal to the product of the underlying phenomenon with the difference
between the intrinsic rate of increase of the underlying phenomenon and the rate of decrease
through societal change.
The difference between the maximal attainable rate of decrease through soci
the rate of decrease through s tal change, divided by the maximum attainable rate of decreas
through sdeietal charge, is used Hard ad's proxy for thie fristration due te inertia — which fruateates
convinced citizens if they do not see sufficient change.
assumed that the contact rate of convinced citizens is equal
to the normal contact rate of the convinced citizens multiplied by the complement of the radicz
action level; with the radical action level equal to the average readiness to take action multiplied
by the frustration of the convinced citizens.
The non-radical action level is the complement of the frustration of the convinced citizens times
the average readiness to take action. The non-radical action level and the normal persuasiveness
determine the pe of non-radical actions. The average readiness to take action equals
the total readiness to take action of all convinced citizens divided by the number of convinced
citizens.
It is furthermore assumed that the total readiness to take action of all convinced citizens
initially equals the product of the number of convinced citizens and an initial readiness to take
action per convinced citizen. The total readiness to take action of all convinced citizens increas
by means of the flow of persuations multiplied by their readiness to take action — that is, per newly
convinced citizens. The average persuasiveness of convinced citizens in this version of the model
also equals the total readiness to take action of all convinced citizens divided by the number of
convinced citizens.
In the basic version of the model, it is assumed that the visibility of the problem equals the
problem symptoms times the reinforcement of the visibility of the problem through radical and
1 actions, and that the reinforcement of the visibility of the problem through radical
and non-radical actions is equal to (1 + non-radical action level) * (1 + radical action level *
reinforcement of the visibility of the problem through radical actions).
‘suasiveness of non-radical actions, and the v
tal change may
ssumed that.
al change and
In this version of the model, it i:
oS
Figure 2 shows a Causal Loop Diagram of this simulation model — this CLD cannot be aggre-
gated /simplified beyond the one shown here without losing the link between driving structure and
behavior.
3.2. Quick and Dirty Exploration of the First De/Radicalization Model
Initial values and parameter values used to generate figures in this subsection are displayed in
Table 1.
Figure 3 shows a ‘deradicalization’ mode of behavior. First, the underlying phenomenon keeps
on worsening, causing the frustration of the —at first- small group of convinced citizens to rise,
causing them to take radical and non-radical actions, hence, increasing the visibility of the problem
-after the first real problem symptoms have become visible~ to sc at large. The resulting
persuasions and drop in frustration (and marginalization) of convinced citizens causes a shift from
Pruyt & Kwakkel, 2011. To Radicalize or to Deradicalize? In: Proc. Int. Conf. SD Society — 7
__ tndering
7 7 phenomenon ——_
»
( che)
\+
- —_—_
perceived problem times
ff growth ate
Hk coninced
a
+4
\ contact ate
‘unconvinced. ‘
\ pes ae,
\ | / oa a demeeot =O
\ \/ average conviction and persunsiveness \ oy /
a _— willingness to act ‘
| A
+ pon radical action
evel /
Figure 2: Causal Loop Diagram of the generic de/radicalization model
radical actions to non-radical actions. This shift from radical to non-radical actions increases ~at
least in this model- the persuasiveness of non-radical actions and de-isolates convinced citizens,
allowing for an increase of their contact rate — jointly leading to more persuz and hence more
and more convinced citizens until all citizens except for the unconvincible, that is, the radical
individuals at the other end of the spectrum~ are convinced. More and more convinced citizens
ion is taken to fight the underlying phenomenon until it c
‘ions
also means that more and more
to be a problem.
Manual univariate sensitivity uncertainty analysis shows that another mode of behavior could
be obtained if starting from the base case values~ one of following changes is made:
e if the initial readiness to take action per convinced citizen 0.8 in the base case~ is larger
than 0.925;
e if the readiness to take action per newly convinced citizen —0.5 in the base case is larger
than 0.8;
1f-interest —2 in the base ca
e if the strength of incompatible s is larger than 16;
e if the normal contact rate of convinced citizens ~1000 in the base case~ is smaller than 400;
e if the normal persuasiveness ~1% in the ba: -— is smaller than 0.25%.
Further radicalization —instead of deradicalization— is obtained in all of these cas
4(a) for the behavior of the variable frustration of the convinced citizens ore
value of the reinforcement parameter does not change the mode of behavior of the model (not
displayed). From these quick qnd dirty analyses, it looks as though two radically different modes
of behavior can be obtained with this model. Figure 4(b) shows two different based on
these two modes of behaviors.
cenari
Pruyt & Kwakkel, 2011. To Radicalize or to Deradicalize? In: Proc. Int. Conf. SD Society — 8
20M. ctizens
4M ciizens/Vear
Perer rer Pete gefenipetol
10M citizens 4.
2M eitzens/Vear
Pe
ts
© citizens 4
0 ciizens/Year ee LL | all
1980” 1995" “20102025 "2040" 20552070.
Tine (Year)
1980 1990 2000 2010 2020 2030 2040 2050 2060 2b70 2bs0
‘Time (Year)
Sotltg eieprele mater
“ay ey ot
Sapp meno
1,000 contzets(Year*eitzen)
1 Dani
0.006 ciizensicontacts
is
500 coniacts/(Year* citizen) a
0S Dam l hep pee FSS
0.003. etzensicontacts Se 5
\ Ns
0 contacts/(Year* citizen) 4
© Dani 1
0. ctizescontacts 4) halen
19891995 20102025 2080 20: o 4
Tame (Vex) 1980 1990 2000 2010 2050 2080 2040 2050" "2060 “Sov oso
cp cso cies nation ——y——-— conc Tinw (Yeas)
fevini ofconrnced Cas indeanton POST pre tetany pp
Stemwiattavelva) mdeaatee = — =
indies action Jvel meécalaton
Spemunstess ofson-adralactons” smaicazatin | §-—~—- 5
Figure 3: The deradicalization mode
1980 1990 2000 2010 2020 2030 2040 2080 2060 2070 2080 1980 1990 2000 2010 2020 2030 2049 2050 2060 2070 2080
Time (Year) Tine (Year)
(a) Behavior mode sensitivity (frustration of convinced (b) Two distinct scenarios: deradicalization (red) versus
citizens) for 5 out of 6 parameters: deradicalization (dark further radicalization (blue)
blue) versus radicalization (all other)
Figure 4: Deradicalization mode versus radicalization mode of the de/radicalization model
Pruyt & Kwakkel, 2011. To Radicalize or to Deradicalize? In: Proc. Int. Conf. SD Society 9
parameter, initial value, lookup value(s) units
underlying phenomenon 50 dmnl
problematic phenomenon level 60 dmal
initial number of convinced citizens 1.000 itizen.
initial number of unconvinced citizens 12.999.000 citizen
initial number of citizens that cannot be convinced 3.000.000 citizen
total population 16.000.000 citizen
maximum attainable rate of decrease through societal change 0.05 1/year
intrinsic rate of increase of the underlying phenomenon 0.01 1/year
normal contact rate of the convinced citizens 1000 contact /person/year
‘iveness 0.01 person/contact
s to take action per convinced citizen 0.80 dmnl
readiness to take action per newly convinced citizens 0.50 dmnl
sibility of the problem thr. radical actions 10 dmnl
normal persue
initial readines
reinforcement of the
strength of incompatible self-interest 2) dm
frustration due to marginalization (0, 1), (0.025, 0.50), (0.05, 0.20), (0.075, 0.05), (0.1, 0), (1,0)
Table 1: Initial values, parameter values and lookup function used in the simulation in Figure 3
3.3 ESDMA of the First De/Radicalization Model
In this case, ESDMA is used for broader and deeper exploration and additional analys
parametric uncertainties. Normally ESDMA for policy testing purposes about such a topic would
require exploration of model uncertainties, structural uncertainties, etc. However, only this generic
model is explored here’. This narrow focus allows us to focus here on analysis techniques and
sualization techniques for bifurcations.
s of some
parameter/initial value minimum maximum
normal contact rate convinced T 2000.0
maximum attainable rate of decrease through societal change 0.01 0.2
problematic phenomenon level 20 100.0
intrinsic rate of increase of the underlying phenomenon 0.005 0.1
strength of the incompatible self-interest 1 20.0
normal persuasiveness 0.001 0.1
readiness to take action per newly convinced citizen 0.2 0.8
Table 2: Parameter values used in the ESDMA
Table 2 contains the parameter ranges used in the ESDMA on this gencric model. Figure
5 shows the behaviors of 1000 Latin Hypercube samples for the (a) perceived or real underlying
phenomenon, (d) the radical action level, (c) the frustration of convinced citizens, and (d) the
it allows to show the
number of convinced citizens. Although 1000 samples is a rather small set
same patterns without much distortion. Often more samples are better. But in this case, too
many samples may result in a distorted picture of the scenario space. The end state histograms of
these variables show that non-white space does not necessarily mean many runs. Three of these
ibutions indicate a clear bifurcation in the uncertainty space. In thi
that two typical modes of behaviors are generated.
Although the set of behaviors is clearly bifurcated, it is not cl
for these uncertainties lead to one or the other outcome. Classification and regression trees allow
finding combinations of values splitting the data set: the tree in Figure 6 has been created using
the C4.5 algorithm (Quinlan 1993) on an ensemble of 1000 scenarios, with the constraints that
minimally 80% of a class needs to belong to one or the other end state and that the minimum
t sizes need to exceed 20 else they are pruned.
end state di
case, it means
ar which combinations of values
sul
4We use a case based development strategy to extend our ESDMA toolkit.
Pruyt & Kwakkel, 2011. To Radicalize or to Deradicalize? In: Proc. Int. Conf. SD Society 10
SNS
12 bee 1e6
Sacer
a
Hi
ie
é
Gar om = om -_ —
EA@AAAAZZA
ANE I IN
2
bso 2000
Figure 5: Behaviors of LH 1000 samples for following variables (top to bottom): (a) perceived or
real underlying phenomenon, (b) the frustration of convinced citizens, (c) the radical action level,
and (d) the convinced citizens
This tree is fairly simple to read and interpret. For example, 269 out of 1000 scenarios had a
problematic phenomenon level of less than 41.524 and 269 out of those 269 cases lead to radical-
ization (=0), not to deradicalization (=1). Out of the remaining 731 scenarios with a problematic
phenomenon level of more than 41.524, 28 had a normal persuasiveness of less than 0.005. None
of these 28 cases lead to deradicalization. From the remaining 703 scenarios (= 1000 - 269 - 28),
63 had a normal contact rate of convinced citizens of less than 185 person/person/year. In 59
cases out of these 63 low contact rate cases result in further radicalization. And so on.
Reading this three the other way around, we find two more subsets that mostly lead to further
radicalization: scenarios characterized by rather high intrinsic rates of increase of the underlying
phenomenon ~ either above 0.027, or above 0.02 in combination with a strong incompatible self-
interest (>3.983) and a moderate normal persuasiveness (0.005 < ...< 0.033). All other subsets
result in most cases in deradicalization.
At this point we should warn against probabilistic interpretation (e.g. 269/1000 or 26.9%)
and/or attaching too much value to the ‘discrete splits’ (e.g. a problematic phenomenon level
of 41.524) and even to the identified subsets: given deeply uncertain and dynamic issues and
exploratory model(s) thereof, classification trees could be used in ESDMA to identify key un-
certainties and to get a rough idea of specific subspaces. Trees generated with slightly different
ttings may already lead to different trees. Analyzing (a set of) tr ka random forests~ may
nevertheless provide useful information for further explorations (c.g. directed searches) and for
designing adaptive policies and associated dynamic monitoring mechanisms.
Pruyt & Kwakkel, 2011. To Radicalize or to Deradicalize? In: Proc. Int. Conf. SD Society 11
problematic phinomenon level
ieee t209 > sot
0.008 Nae
{G8 ou of) >) woul oat te comes ens
re ot of 8) readies to take ation per newly corwineed citizen
oon Ne
normal persuasiveness —_intinsic rate of micrease of the underlying phenomenon
Z —
sea ofimcompaleseFataest C0 ew of 28) > C_oasoutetz0) Ca for eatotos) >
/ =3.983, SS 983
~/
7 out oF 30) >) ininsc rte of increase ofthe mudeing phen
~0020
20s ou jot2) >
Figure 6: Classification tree: 1 = radicalization ; 0 = deradicalization
1 (01 out of 109)
The two different modes are also clearly distinguishable in the 4D-plot in Figure 7 (1000 sce-
narios). Following four variables aled between 0 and 10 and plotted: the perceived or
real underlying phenomenon, the number of convinced citizens, the frustration of those convinced
citizens, and the radical action level (in color). This 4D-plot shows that —with this version of the
model and with these uncertainties— two opposite end states are reached: one end state consists of
many convinced citizens who deal with the underlying phenomenon and, hence, their frustrations
are res
— the other end state consists of few highly frustrated and ‘isolated’ individuals who are unable
to convince others —with more radical actions~ and, hence, are unable to deal with the underlying
phenomenon.
Time series clustering on the underlying phenomenon with a simplistic Atomic Behavior Pat-
tern concatenation and classification algorithm on these 10000 runs leads to 15 clusters of respec-
tively 8505, 640, 512, 111, 69, 36, 32, 31, 22, 21, 9, 7, 3, 1, and 1 runs (see Figure 8 for eight
interesting clusters).
3.4 Conclusions Related to the First De/Radicalization Model
The model p: just one among many different models about deradi
tion and radicalization. This particular model leads to counter-intuitive policy recommendations:
it is implicitly assumed that isolation leads to radicalization, hence, the adv’
calization would be to integrate radical groups/individuals or to manage society at large through
a transition process towards the good aspects of (some) ideas shared by (potential) radical indi-
viduals/groups. Moreover, this model is generic: it could be tailor-made for specific situations
but does not —in its current form- fit any situation in particular. Hence, it does not have any
predictive value for specific real-world de/radicalization processes. But even then is it useful for
several reasons:
nted and used above
an
in view of deradi-
Pruyt & Kwakkel, 2011. To Radicalize or to Deradicalize? In: Proc. Int. Conf. SD Society 12
frustration of
convinced \
\
oo” \ _-~ 0.000
\
convinced _— underlying
citizens phenomenon
Figure 7: 4D-plot (3 spatial dimensions + 1 color dimension) of 1000 runs
e The model is from a methodological point of view— an interesting toy model for developing
and testing ESDMA analysis techniques and visualization tools for cases characterized by
bifurcations (see above): in this particular case, classifying the end states means classifying
the modes of behaviors, not necessarily the root causes (uncertainties) — classifying modes
of behavior under deeply uncertain dynamic complexity is much more difficult®.
The model is from an applied point of view~ an interesting starting point for developing
rather different exploratory models about specific de/radicalization processes;
e The model is -from an educational point of view— an interesting model for illustrating how
one and the same endogenous model can generate a few different —almost opposite- modes
of behavior.
5Glassification algorithms for more complex cases (in terms of modes of behaviors) are currently under develop-
ment in our research lab.
Pruyt & Kwakkel, 2011. To Radicalize or to Deradicalize? In: Proc. Int. Conf. SD Society 13
MA
(a) 8505 runs (b) 640 runs (c) 512 runs (a) 111 runs
(h) 7 runs
4 The Second De/Radicalization Model
The model pre ction is another combination of micro-hypoth
de/radicalization — similar to the previous model, but different enough to be called a different
model. The model is also high-level, simple, generic, and ‘wrong’, but may be useful. The model
is presented in subsection 4.1 and used for ESDMA in subsection 4.2.
ented and used in this about
4.1 Model Structure of the Second De/Radicalization Model
Changing the previous generic model slightly already leads to a new de/radicalization model (see
Figure 9). Additional structures were added (part of the variables in red) to distinguish activists
from extremists from terrorists (AIVD 2010). They influence the system in different ways. Other
functions and structures were changed too (the other part of the variables in red).
4.2 ESDMA of the Second De/Radicalization Model
Figure 10 shows the outputs of 1000 simulations with this model for the parametric uncertainties
discussed above (as well as uncertainty ranges on the initial values of the subpopulations in the
model). The graphs of the individual traces seem to suggest that the results of the end states
are more or less evenly spread over the range. The corresponding envelopes and distributions tell
a slightly different story, especially for homegrown extremists and homegrown terrorists, namely
that the full spectrum is plausible, but not likely. Different cuts would allow the exploration
of the development over time. Again it should be stated that (i) quantitative interpretation of
probability distributions or histograms is irrelevant from an EMA point of view — plausi
what matters, and (ii) only some parametric uncertainties are explored here using a simplistic toy
model ~ real analysis requires at least the consideration of structural and model uncertainties too.
Direct searches in and analysis applied to the ensemble of scenarios may help to explore causes of
the low-probability high-risk cases and may suggest potential solutions.
But here it suffices to note that the ensemble of behaviors generated by this model and these
uncertainties is plausible and that it differs from the ensemble generated with the first model —
sity is good.
remember: divi
to Deradicalize? In: Proc. Int. Conf. SD Society 14
Pruyt & Kwakkel, 2011. To Radicalize or
Figure 9: Stock Flow Diagram of an alternative De/Radicalization model
Pruyt & Kwakkel, 2011. To Radicalize or to Deradicalize? In: Proc. Int. Conf. SD Society 15
40000 - =
200000)
000000}
‘nooo
‘00000
00000
200000
hw 7000 70
2800
2a)
g
Figure 10: Individual traces LH1000 (left) and the corresponding envelopes and end state distri-
butions for the second SD model for following variables (from top to bottom): convinced citizens,
homegrown activists, homegrown extremists, and homegrown terrorists
5 ESDMA on 2 x 2 Models
Both models dis
from convinced citizens to unconvinced citizens) are used in this section as an illustration of an
ESDMA with multiple models. The different models should, in real-world multi-model ESDMAs,
be plausible and also as divers as possible — which case in this illustration. Only the
core set of uncertainties should be shared (model specific uncertainties can be dealt with too).
In a real-world ESDMA, models should not be pre-analyzed in detail as in subsection 3.3.
Testing the plausibility of behavior ~as in subsection 4.2- is recommended though. The reason for
not performing detailed analysis on cach of the models’ output relates to the aim of ESDMA: to
generate an ensemble of —ideally all- plausible behaviors and to analyze the ensemble as a whole.
Moreover, real-world EMA should also include models that are used in the field but are black-box
models (e.g. undisclosed for proprietary reasons or black-box models from a technical point of
view). Ex-post, that is to say, during the analysis phase, time-series clustering and classification
may still show that a particular type of behavior is only generated by a particular model or
structure, although that is not guaranteed. Subsection 3.3 does not make much sense from an
ESDMA point of view — it does from a SD point of view and an educational point of view (first
uni-model ESDMA before multi-model ESDMA).
ed above as well a
two structural variants (the two models with return flows
5.1 The Ensemble of Scenarios
Figure 11 displays 1000 individual runs generated with the same experimental design (Latin Hy-
percube) uniformly changing the 12 uncertainties displayed in Table 3 for each of the four different
models (thus, simulating cach model 250 times). Figure 11(b) shows the same runs, but in a dif.
ferent color for each of the four models, both for envelopes and lines.
Pruyt & Kwakkel, 2011. To Radicalize or to Deradicalize? In: Proc. Int. Conf. SD Society 16
12 le6
if fmderying phenomenon
&
06
83
“bso 2000 3020 2040 2060 2080"
80) 2000 2040 2060 2080
(a) lines for all models (no distinction)
3 [underlying phenomenon
8
“fbso 2000 2020 2040 2060 2080
so 2000 2020 2040 2060 2080,
80 2000 2020 2040 2060 2080
(b) lines color-coded for each of the models
Figure 11: Lines and kernel density estimates without distinction between the models (left) and
with distinction between the different models (right)
Pruyt & Kwakkel, 2011. To Radicalize or to Deradicalize? In: Proc. Int. Conf. SD Society 17
parameter/initial value minimum maximum
ens 1 3650
normal contact rate of convinced ci
maximum attainable rate of decrease through societal change 0.001 0.1
problematic phenomenon level 20 200
intrinsic rate of increase of the underlying phenomenon 0.005 0.1
strength of incompatible self interest 0.8 20
normal persuasiveness 0.0001 0.01
readiness to take action per newly convinced citizen 0 1
initial number of convinced citizens 50000 150000
initial number of activ: 500 1500
initial number of extrem 0 10
initial number of terroris 0 10
citizens that cannot be convinced 1000000 10000000
Table 3: 12 random uniformly sampled uncertainties (Latin Hypercube) and the corresponding
uncertainty ranges
parameter/initial value minimum maximum
normal contact rate of convinced citizens 365 3650
maximum attainable rate of decrease through societal change 0.01 01
normal contact rate of convinced citizens 365 3650.
maximum attainable rate of decrease through societal change 0.01 01
intrinsic rate of increase of the underlying phenomenon 0.005 01
normal persuasiveness 0.0001 0.01
readiness to take action per newly convinced citizen 0 1
citizens that cannot be convinced 1000000 10000000
Table 4; 8 random uniformly sampled uncertainties (full factorial) and the corresponding uncer-
tainty ranges from which 4 values are sampled
A Full Factorial (FF) design may be needed to control the experimental design and perform
particular analysis. Starting with the same type of analysis as in the LH simulations helps to
appreciate differences and similarities between the designs for the case at hand. Applying a full
factorial design to each of these 4 models with 4 values per uncertainty for the uncertainties listed
in Table 3 leads to 16384 runs. Figure 12 shows 1190 out of 16384 full factorial runs with end
states for radical action level (RAL) above 1, and the remainder of the runs with a RAL end
state value below 1. Concatenation and time-series clustering of this ensemble already leads to 16
clusters for the subset with RALs above 1.
One type of analysis and visualization for which a Full Factorial (FF) design is (currently)
needed is robustness analysis based on relative regret. Normally one would want to use that
type of analysis to compare different policies over the entire uncertainty space. With different
models (and without different policies) this type of analysis could also be used to detect which
model performs worst in parts of the uncertainty space. Figure 13 shows the performance on
a single key output indicator (the radical action level) for all FF uncertainties, visualized for
two uncertainties (the normal contact rate of convinced citizens between 365 and 3650 and the
maximum attainable rate of decrease through societal change between 0.01 and 0.1 (white = zero
regret, which is desirable — for more information about this type of analysis and visualization,
see (Lempert, Popper, and Bankes 2003)). These graphs indicate that the first model does not
perform well for a below-average maximum attainable rate of decrease through societal change
and that the behavior of the other models mainly differs in terms of the normal contact rate of
convinced citizens, with the second model performing badly in the lower range, and the other
two models performing badly in the higher range of the normal contact rate of convinced citizens.
These results could be used to develop policies to remedy worst performance.
Pruyt & Kwakkel, 2011. To Radicalize or to Deradicalize? In: Proc. Int. Conf. SD Society 18
ke les
Fi
q
B
Ee
q
: :
———— aan
ZB : =
am 00 220 240 00 ‘2800
is
bso 200 m0, M0, 7 OSLO 709
(a) 1190 out of 16384 with radical action level end states above 1
(b) rest of the runs with radical action level end states below 1
Pruyt & Kwakkel, 2011. To Radicalize or to Deradicalize? In: Proc. Int. Conf. SD Society 19
Figure 13: Visualization of the relative regret for the radical action level for the same Full Factorial
design applied to each of the four models (white is no regret which is preferable)
6 Conclusions
Although Exploratory System Dynamics Modeling and Analysis is a model-based methodology, it
actually requires a mental switch from traditional modelers. In ESDMA, there is no single base
case: all plausible cases are base cases.
The main uses of ESDMA are exploration to broaden the horizon and testing of deep robust-
s. If it reverts to prediction at all, then only to ‘ensemble prediction of (modes of) behaviors’
over the entire uncertainty space and claims about ‘policy robustness’. It starts and ends with
uncertainties. And the design of adaptive policies and resilient systems that lead to robust behav-
ior/outcomes are pursued.
ESDMA greatly enhanc
nes!
SD. ESDMA is currently under rapid development in our ESDMA
lab. New versions of the software are continuously developed (the preliminary version of this
paper was developed using version 1.0 ~a single model version~ and the final version using ver-
sion 2.1 -a multi-model version~ and version 3.0 -a multi-model multi-policy version will be
implemented soon) and many algorithms are under development, from time ification
algorithm (System Dynamicist surely understand that the analysis of end-state is insufficient) to
powerful interactive data visualization algorithms.
This paper shows the application of ESDMA to radicalization versus deradicalization. How-
ever, that is not the real issue addressed in this paper. The main issue addressed is methodological
(almost philosophical): current computing power allows to deal with deep uncertainty and dynamic
complexity. And if we can, then we should. The specific point made in this paper is the need for
multi-model exploration to deal with structural and model uncertainty.
The application domain is of course interesting too: some
models about de/radicalization and activism were presented and
models generate fascinating dynamics. The de/rad
-from an applied decision support point of view — .
The first model illustrated both ESD and ESDMA for cases characterized by bifurcations. The
other models showed different ensemble behaviors — nevertheless all plausible. Multiple models
should preferably be used in ESDMA since the set of all plausible behaviors ~aimed for in ESDMA~
cannot be generated with just one model. The models considered in this paper are from that point
of view too similar: in reality, a much wider model diversity should be aimed for.
The uncertainties dealt with in this paper included parametric uncertainties, but also different
models. Not dealt with were uncertainties related to functions and formulations, generative struc-
tures, loops, and submodels within one and the same model, different model formulations and
model boundaries, and hence, different world-views (and hence, different models and preference
systems, and different modeling methods and paradigms). Real-world ESDMA of de/radicalization
would also require the consideration of structural, uncertainties, model uncertainties, different
world-views, etc. Only then would the ESDMA be sufficiently broad for testing the robustness of
adaptive policies (the effectiveness of policies over the ensemble of scenarios).
simplistic versions of all plausible
(jointly) analyzed. These simple
s dealt with in this paper are
Pruyt & Kwakkel, 2011. To Radicalize or to Deradicalize? In: Proc. Int. Conf. SD Society 20
References
Agusdinata, D. (2008). Exploratory Modeling and Analysis. A Promising Method to
Deal with Deep Uncertainty. Phd dissertation, Delft University of Technology, Delft.
http: //www.nextgenerationinfrastructures.eu/index. php?pageID=17#itemID=433706.
24
Agusdinata, D., J. Van der Pas, W. Walker, and V. Marchau (2009). An innovative multi-
criteria analysis approach for evaluating the impacts of intelligent speed adaptation. Journal
of Advanced Transport systems 43(4), 413-454. 4
AIVD (2010, February). Dierenrechtenextremisme in Nederland. Technical report, ATVD. Down-
loaded on 21/04/2010. 13
Bankes.
S. (1993, May-June). Exploratory modeling for policy analysis.
ch 41(3) 449. 4
Bankes, $., W. Walker, and J. Kwakkel (2010). Encyclopedia of OR (upcoming 3th edition
ed.)., Chapter Exploratory Modeling and Analysis. Boston: Kluwer Academic Publishers.
[forthcoming]. 4
Operations Re-
Bryant, B. and R. Lempert (2009). Thinking inside the box: A participatory, computer-assisted
approach to scenario discovery. Technological Forecasting € Social Change 77(2010), 34-49.
4
Kwakkel, J. and J. Slinger (2011, July). A system dynamics model-based exploratory analysis of
salt water intrusion in coastal aquifers. In Proceedings of the 29th International Conference
of the System Dynamics Society, Washington, USA. System Dynamics Society. 2, 4
Kwakkel, J., W. Walker, and V. Marchau (2010a). Assessing the efficacy of adaptive airport
strategic planning: Results from computational experiments. Transport Policy. (accepted
with revisions). 4
Kwakkel, J., W. Walker, and V. Marchau (2010b, 2010). From predictive modeling to ex-
ploratory modeling: How to use non-predictive models for decisionmaking under deep un-
certainty. In Proceedings of the 25th Mini-EURO Conference on Uncertainty and Robust-
ness in Planning and Decision Making (URPDM2010), 15-17 April, Portugal. University of
Coimbra. 4
Lempert, R., D. Groves, $. Popper, and S. Bankes (2006). A general, analytic method for
generating robust strategies and narrative scenarios. Management Science 52(4), 514-528.
4
Lempert, R., §. Popper, and $. Bankes (2003). Shaping the next one hun-
dred years: New methods for quantitative, long-term policy _ analysis.
RAND report MR-1626, The RAND Pardee Center, Santa Monica, CA.
http: //www.rand. org/pubs/monograph_reports/2007/MR1626. pdf. 1, 2, 4, 17
Lempert, R. and M. Schlesinger (2000). Robust strategies for abating climate change. Climatic
Change 45(3-4), 387-401. 4
Pruyt, E. (2007). Dealing with Uncertainties? Combining System Dynamics with Multiple
Criteria Decision Analysis or with Exploratory Modelling. In Proceedings of the 25th Inter-
national Conference of the System Dynamics Society, Boston, MA, USA. System Dynamics
Society. www. systemdynamics . org/conferences/2007/proceed/papers/PRUYT386. pdf. 2
Pruyt, E. (2009, July). Making System Dynamics Cool? Using Hot Testing
& Teaching Cases. In Proceedings of the 27th International Conference of
the System Dynamics Society, Albuquerque, USA. System Dynamics Society.
http: //www.systemdynamics. org/conferences/2009/proceed/papers/P1167 . pdf.
22
Pruyt & Kwakkel, 2011. To Radicalize or to Deradicalize? In: Proc. Int. Conf. SD Society 21
Pruyt, E. (2010a, July). Making System Dynamics Cool II: New hot testing and teach-
ing cases of increasing complexity. In Proceedings of the 28th International Con-
ference of the System Dynamics Society, Scoul, Korea, System Dynamics Society.
http: //systemdynamics . org/conferences/2010/proceed/papers/P1026.pdf. 22
Pruyt, E. (2010b, July). Using Small Models for Big Issues: Exploratory System Dy-
namics for Insightful Crisis Management. In Proceedings of the 28th International Con-
ference of the System Dynamics Society, Scoul, Korea. System Dynamics Society.
http: //systemdynamics. org/conferences/2010/proceed/papers/P1266.pdf. 1, 2, 4
Pruyt, E. (2011a, July). Heart meets mind: Smart transition management to smarten energy
systems in a deeply uncertain world. In Proceedings of the 29th International Conference of
the System Dynamics Society, Washington, USA. System Dynamics Society. 2, 4
Pruyt, E. (2011b, July). Making System Dynamics cool III: New hot teaching & testing cases
of increasing complexity. In Proceedings of the 29th International Conference of the System
Dynamics Society, Washington, USA. System Dynamics Society. 22
Pruyt, E. and C. Hamarat (2010a). The concerted run on the DSB Bank: An Ex-
ploratory System Dynamics Approach. In Proceedings of the 28th International Con-
ference of the System Dynamics Society, Scoul, Korea. System Dynamics Society.
http: //systemdynamics . org/conferences/2010/proceed/papers/P1027 .pdf. 2, 4
Pruyt, E. and C. Hamarat (2010b). The Influenza A(HIN1)v Pandemic: An Ex-
ploratory System Dynamics Approach. In Proceedings of the 28th International Con-
ference of the System Dynamics Society, Scoul, Korea. System Dynamics Society.
http: //systemdynamics . org/conferences/2010/proceed/papers/P1389.pdf. 2, 4
Pruyt, E., J. Kwakkel, G. Yucel, and C. Hamarat (2011, July). Energy transitions towards
sustainability: A staged exploration of complexity and deep uncertainty. In Proceedings
of the 29th International Conference of the System Dynamics Society, Washington, USA.
System Dynamics Society. 2, 4
Pruyt, E., T. Logtens, and G. Gijsbers (2011, July). Exploring demographic shifts: Aging and
migration ~ Exploratory Group Model Specification and Simulation. In Proceedings of the
29th International Conference of the System Dynamics Society, Washington, USA. System
Dynamics Society. 2, 4
Quinlan, J. R. (1993). C4.5: programs for machine learning. San Francisco, CA: Morgan Kauf-
mann Publishers Inc. 9
Senge, P. (1990). The fifth discipline: the art and practice of the learning organization. London.
424 p. 1, 2
Sterman, J. (2002). The 2001 Jay W. Forrester Award. System Dynamics Review 18(1), 71-72.
2
Van der Pas, J., D. Agusdinata, W. Walker, and V. Marchau (2007). Exploratory modelling
to support multi-criteria analysis to cope with the uncertainties in implementing isa. In
R. L. of Highway Ministry of Communications (Ed.), 14th World Congress on Intelligent
Transport Systems, ITS For a Better Life, Beijing. 4
Van der Pas, J., D. Agusdinata, W. Walker, and V. Marchau (2008). Developing robust in-
telligent speed adaption policies within a multi-stakcholder context: An application of ex-
ploratory modelling. In P. Herder, P. Heijnen, and A. Nauta (Eds.), JBEE ~ NGInfra Sci-
entific Conference 2008: Building Networks for a Brighter Future, Delft. 4
Walker, W. and V. Marchau (2003). Dealing with uncertainty in policy anal:
making. Integrated Assessment 4(1), 1-4. 4
, and policy-
Pruyt & Kwakkel, 2011. To Radicalize or to Deradicalize? In: Proc. Int. Conf. SD Society 22
A Teaching and Testing Cases
The case description of one of two hot teaching/testing —in line with other teaching/testing pub-
lished in (Pruyt 2011b), (Pruyt 2010a) and (Pruyt 2009)- the one corresponding to the first model
used in this paper, is available in this appendix. The second model is only available to SD lecturers
upon request.
Case Description: The De/Radicalization Case I (_ /25)
A new international research institute dedicated to security and terrorism has recently been
founded in The Hague. The institute asks you to make a generic, explorative SD simulation
model related to (de) radicalization (that could at a later stage be used to study animal rights
activism, climate terrorism, etc). From discussions with leading security experts, you learn that:
Radicalization is often brought about by a (real or perceived) slumbering underlying phe-
nomenon. Suppose for example that this underlying phenomenon initially equals 50 and increases
(or decreases) through a net-increase of the underlying phenomenon. Problem symptoms become
visible when/if the level of this underlying phenomenon increases above a problematic phenomenon
level of, say, 60. These problem symptoms could be modeled simplistically as the difference be-
tween the underlying phenomenon and the problematic phenomenon level, divided by the problem-
atic phenomenon level. Problem symptoms can (mathematically speaking) only be non-negative.
But even if there are no problem symptoms yet, there is —for convinced citizens~ a perceived
magnitude of the problem, equal to the underlying phenomenon divided by the problematic phe-
nomenon level. The higher the perceived magnitude of the problem is, the higher the frustration of
the convinced citizens will be. Security experts suggest to model the frustration of the convinced
citizens as a percentage between 0% and 100%, equal to the product of the frustration due to
marginalization, the frustration due to inertia, the average degree of conviction of the convinced
citizens, and the perceived magnitude of the problem.
Unconvinced citizens could become convinced citizens if they are persuaded by already con-
vinced citizens. However, there are also Suppose that the
initial number of convinced citizens equals 1.000, the initial number of unconvinced citizens equals
12.999.000 and that the initial number of citizens that cannot be convinced equals 3.000.000 — on
a total population of 16.000.000 citizens. The convinced citizens could for example be modeled as
the product of the contact rate of the convinced citizens, the fraction of convinced citizens, the un-
convinced citizens, the persuasiveness of non-radical actions, and the visibility of the problem (see
last paragraph), divided by the strength of incompatible self-interest. The fraction of convinced
citizens equals of course the number of convinced citizens divided by the sum of all citizens.
Suppose for a start that: if the fraction of convinced citizens equals 0 then the frustration due
to marginalization equals 100%, if the fraction equals 0.025 then the frustration due to marginal-
ization equals 50%, if the fraction equals 0.05 then the frustration due to marginalization equals
20%, if the fraction equals 0.075 then the frustration due to marginalization equals 5%, and that
if the fraction is equal or greater than 0.1, the frustration due to marginalization equals 0%.
More convinced citizens implies that more socictal change may be expected. Societal change
may allow to stop the increase of, or even decrease, the underlying phenomenon: assume that the
rate of decrease through societal change equals the fraction of convinced citizens multiplied by the
maximum attainable rate of decrease through societal change of for example 5% per year. This
rate of decrease through societal change mitigates the net increase of the underlying phenomenon.
This net increase of the underlying phenomenon equals the product of the underlying phenomenon
with the difference between the intrinsic rate of increase of the underlying phenomenon and the
rate of decrease through societal change. Suppose in this generic model that the intrinsic rate of
increase of the underlying phenomenon is equal to 1% per year.
The difference between the mazimal attainable rate of decrease through societal change and
the rate of decrease through societal change, divided by the maximum attainable rate of decrease
itizens that will never be convin
Pruyt & Kwakkel, 2011. To Radicalize or to Deradicalize? In: Proc. Int. Conf. SD Society 23
through societal change, can be used as a proxy for the frustration due to inertia which frustrates
the convinced citizens if they do not see sufficient change.
Suppose that the contact rate of the convinced citizens equals the normal contact rate of
the convinced citizens of, say, 1000 (different and sufficiently close) contacts per person per year
multiplied (100% - radical action level); with the radical action level equal to the average readiness
to take action multiplied by the frustration of the convinced citizens.
The non-radical action level is the complement of the radical action level, in other words (100%
- frustration of the convinced citizens) * average readiness to take action. The non-radical action
level and the normal persuasiveness of 0.01 persons per contact determine the persuasiveness of
non-radical actions. The average readiness to take action equals the total readiness to take action
of all the convinced citizens divided by the number of convinced citizens.
Suppose that the total readiness to take action of all the convinced citizens initially equals the
product of the number of convinced citizens and an initial readiness to take action per convinced
citizen of, say, 80%. The total readiness to take action of all the convinced citizens increases
through adding readiness to take action of the newly convinced citizens, equal to the flow of the
newly convinced citizens multiplied by the readiness to take action per newly convinced citizens
of, say, 50%. The average persuasiveness of the convinced citizens also equals the total readiness
to take action of all the convinced citizens divided by the number of convinced citizens.
Suppose, for now, that the visibility of the problem is equal to the problem symptoms times
the reinforcement of the visibility of the problem through radical and non-radical actions, and that
the reinforcement of the visibility of the problem through radical and non-radical actions is equal
to (1 + non-radical action level) * (1 + radical action level * reinforcement of the visibility of the
problem through radical actions). Set the reinforcement of the visibility of the problem through
radical actions equal to 10, and the strength of incompatible self-interest to 2.
1. ( /10) Model this i
starting in 1980.
ue. Verify and simulate the model over a time horizon of 100 years,
2. (/1) Make graphs of the convinced citizens and the dynamic factors that directly influence
the newly convinced citizens.
3. (/2) Validate the model: describe and perform 2 different validation tests (except sensitivity
analysis — see the next question) and the conclusions of these tests. What would/should be
the goal of validation in this explorative case?
4. (/3) Test the
ness to take a
sensitivity of the model for changes in following parameters: the initial readi-
ion per convinced citizen, the readiness to take action per newly convinced
citizens, the strength of incompatible self-interest, the reinforcement of the visibility through
radical actions, the normal contact rate of the convinced citizens, and the normal persua-
siveness. Briefly describe your conclusions.
5. ( /3) Make two interesting and consistent exploratory scenarios. Simulate the scenarios and
save the outputs (graphs). What could be concluded from those exploratory scenarios?
6. (/2) Make a strongly aggregated CLD of this simulation model. And use the CLD to explain
the link between structure and behavior.
. (/2) Formulate policy advice —based on this modeling exercise- related to deradicalization.
won
. (/1) Give examples of 2 ‘soft variables’ in this model. Why could they be called ‘soft’?
°
. (/1) This is of course a preliminary generic model, and the an is explorative at best.
Give advice related to future refinements and extensions. How do you think those refinements
and extensions would change the model behavior?
