Info Tech in Global Economy
Sudha_Ghanta123
Chapter5.docxChapter 5
From Building a Model toAdaptive Robust Decision Making Using Systems Modeling
Erik Pruyt
Abstract Starting from the state-of-the-art and recent evolutions in the field of system dynamics modeling and simulation, this chapter sketches a plausible near term future of the broader field of systems modeling and simulation. In the near termfuture,differentsystemsmodelingschoolsareexpectedtofurtherintegrateand accelerate the adoption of methods and techniques from related fields like policy analysis, datascience, machinelearning, andcomputerscience.Theresultingfuture state of the art of the modeling field is illustrated by three recent pilot projects. Each of these projects required further integration of different modeling and simulation approaches and related disciplines as discussed in this chapter.These examples also illustratewhichgapsneedtobefilledinordertomeettheexpectationsofrealdecision makers facing complex uncertain issues.
5.1 Introduction
Many systems, issues, and grand challenges are characterized by dynamic complexity, i.e., intricate time evolutionary behavior, often on multiple dimensions of interest. Many dynamically complex systems and issues are relatively well known, buthavepersistedforalongtimeduetothefactthattheirdynamiccomplexitymakes them hard to understand and properly manage or solve. Other complex systems and issues—especially rapidly changing systems and future grand challenges—are largely unknown and unpredictable. Most unaided human beings are notoriously bad at dealing with dynamically complex issues—whether the issues dealt with are persistent or unknown. That is, without the help of computational approaches, most humanbeingsareunabletoassesspotentialdynamicsofcomplexsystemsandissues, and are unable to assess the appropriateness of policies to manage or address them.
Modelingandsimulationisafieldthatdevelopsandappliescomputationalmethods to study complex systems and solve problems related to complex issues. Over the past half century, multiple modeling methods for simulating such issues and for advising decision makers facing them have emerged or have been further developed. Examples include system dynamics (SD) modeling, discrete event simulation (DES), multi-actor systems modeling (MAS), agent-based modeling (ABM), and complex adaptive systems modeling (CAS).All too often, these developments have taken place in distinct fields, such as the SD field or theABM field, developing into separate “schools,” each ascribing dynamic complexity to the complex underlying mechanismstheyfocuson,suchasfeedbackeffectsandaccumulationeffectsinSDor heterogenousactor-specific(inter)actionsinABM.Theisolateddevelopmentwithin separate traditions has limited the potential to learn across fields and advance faster andmoreeffectivelytowardsthesharedgoalofunderstandingcomplexsystemsand supporting decision makers facing complex issues. Recent evolutions in modeling and simulation together with the recent explosive growthincomputationalpower,data,socialmedia,andotherevolutionsincomputer science have created new opportunities for model-based analysis and decision making. These internal and external evolutions are likely to break through silos of old, open up new opportunities for social simulation and model-based decision making, and stir up the broader field of systems modeling and simulation. Today, different modeling approaches are already used in parallel, in series, and in mixed form, and several hybrid approaches are emerging. But not only are different modeling traditions being mixed and matched in multiple ways, modeling and simulation fields have also started to adopt—or have accelerated their adoption of—useful methods andtechniquesfromotherdisciplinesincludingoperationsresearch,policyanalysis, data analytics, machine learning, and computer science. The field of modeling and simulation is consequently turning into an interdisciplinary field in which various modeling schools and related disciplines are gradually being integrated. In practice, the blending process and the adoption of methodological innovations have just started. Although some ways to integrate systems modeling methods and many innovationshavebeendemonstrated, furtherintegrationandmassiveadoptionarestill awaited. Moreover, other multi-methods and potential innovations are still in an experimental phase or are yet to be demonstrated and adopted. Inthischapter,someofthesedevelopmentswillbediscussed,apictureofthenear future state of the art of modeling and simulation is drawn, and a few examples of integratedsystemsmodelingarebrieflydiscussed.TheSDmethodisusedtoillustrate these developments. Starting with a short introduction to the traditional SD method in Sect. 5.2, some recent and current innovations in SD are discussed in Sect. 5.3, resulting in a picture of the state of modeling and simulation in Sect. 5.4. A few examplesarethenbrieflydiscussedinSect.5.5toillustratewhatthesedevelopments couldresultinandwhatthefuturestate-of-the-artofsystemsmodelingandsimulation could look like. Finally, conclusions are drawn in Sect. 5.6.
5.2 System Dynamics Modeling and Simulation of Old
System dynamics was first developed in the second half of the 1950s by Jay W. Forresterandwasfurtherdevelopedintoaconsistentmethodbuiltonspecificmethodological choices1. It is a method for modeling and simulating dynamically complex systems or issues characterized by feedback effects and accumulation effects. Feedback means that the present and future of issues or systems, depend—through a chain of causal relations—on their own past. In SD models, system boundaries are set broadly enough to include all important feedback effects and generative mechanisms. Accumulation relates not only to building up real stocks—of people, items, (infra)structures, etc.,—but also to building up mental or other states. In SD models, stock variables and the underlying integral equations are used to group largely homogenous persons/items/... and keep track of their aggregated dynamics over time. Together, feedback and accumulation effects generate dynamically complex behavior both inside SD models and—so it is assumed in SD—in real systems. Other important characteristic of SD are (i) the reliance on relatively enduring conceptualsystemsrepresentationsinpeople’sminds,akamentalmodels(Doyleand Ford 1999, p. 414), as prime source of “rich” information (Forrester 1961; Doyle and Ford 1998); (ii) the use of causal loop diagrams and stock-flow diagrams to represent feedback and accumulation effects (Lane 2000); (iii) the use of credibility and fitness for purpose as main criteria for model validation (Barlas 1996); and (iv) theinterpretationofsimulationrunsintermsofgeneralbehaviorpatterns,akamodes of behavior (Meadows and Robinson 1985). In SD, the behavior of a system is to be explained by a dynamic hypothesis, i.e., a causal theory for the behavior (Lane 2000; Sterman 2000). This causal theory is formalizedasamodelthatcanbesimulatedtogeneratedynamicbehavior.Simulating the model thus allows one to explore the link between the hypothesized system structure and the time evolutionary behavior arising out of it (Lane 2000). Not surprisingly, these characteristics make SD particularly useful for dealing withcomplexsystemsorissuesthatarecharacterizedbyimportantsystemfeedback effects and accumulation effects. SD modeling is mostly used to model core system structures or core structures underlying issues, to simulate their resulting behavior, andtostudythelinkbetweentheunderlyingcausalstructureofissuesandmodelsand theresultingbehavior.SDmodels,whicharemostlyrelativelysmallandmanageable, thusallowforexperimentationinavirtuallaboratory.Asaconsequence, SDmodels are also extremely useful for model-based policy analysis, for designing adaptive policies (i.e., policies that automatically adapt to the circumstances), and for testing theirpolicyrobustness(i.e.,whethertheyperformwellenoughacrossalargevariety of circumstances).
In terms of application domains, SD is used for studying many complex social– technical systems and solving policy problems in many application domains, for example, in health policy, resource policy, energy policy, environmental policy, housing policy, education policy, innovation policy, social–economic policy, and other public policy domains. But it is also used for studying all sorts of business dynamics problems, for strategic planning, for solving supply chain problems, etc. At the inception of the SD method, SD models were almost entirely continuous, i.e.,systemsofdifferentialequations,butovertimemoreandmorediscreteandother noncontinuous elements crept in. Other evolutionary adaptations in line with ideas from the earliest days of the field, like the use of Group Model Building to elicit mental models of groups of stakeholders (Vennix 1996) or the use of SD models as engines for serious games, were also readily adopted by almost the entire field. But slightly more revolutionary innovations were not as easily and massively adopted. In other words, the identity and appearance of traditional SD was well established bythemid-1980sanddoes—atfirstsight—notseemtohavechangedfundamentally since then.
5.3 Recent Innovations and Expected Evolutions
5.3.1 Recent and Current Innovations
Looking in somewhat more detail at innovations within the SD field and its adoption of innovations from other fields shows that many—often seemingly more revolutionary—innovations have been introduced and demonstrated, but that they have not been massively adopted yet. Forinstance,intermsofquantitativemodeling,systemdynamicistshaveinvested inspatiallyspecificSDmodeling(RuthandPieper1994;Struben2005;BenDorand Kaza2012),individualagent-basedSDmodelingaswellasmixedandhybridABMSDmodeling(CastilloandSaysal2005;Osgood2009;Feolaetal.2012;Rahmandad and Sterman 2008), and micro–macro modeling (Fallah-Fini et al. 2014). Examples of recent developments in simulation setup and execution include model calibration andbootstrapping(Oliva2003;Dogan2007),differenttypesofsampling(Fiddaman 2002;Ford1990;Clemsonetal.1995;IslamandPruyt2014),multi-modelandmultimethodsimulation(PruytandKwakkel2014; Moorlag2014), anddifferenttypesof optimization approaches used for a variety of purposes (Coyle 1985; Miller 1998; Coyle1999;GrahamandAriza1998;Hamaratetal.2013,2014).Recentinnovations in model testing, analysis, and visualization of model outputs in SD include the developmentandapplicationofnewmethodsforsensitivityanduncertaintyanalysis (Hearne 2010; Eker et al.2014), formal model analysis methods to study the link between structure and behavior (Kampmann and Oliva 2008, 2009; Saleh et al. 2010), methods for testing policy robustness across wide ranges of uncertainties (Lempertetal.2003),statisticalpackagesandscreeningtechniques(FordandFlynn 2005; Taylor et al. 2010), pattern testing and time series classification techniques
(YücelandBarlas2011;Yücel2012;SuculluandYücel2014;IslamandPruyt2014), and machine learning techniques (Pruyt et al. 2013; Kwakkel et al. 2014; Pruyt et al. 2014c). These methods and techniques can be used together with SD models to identify root causes of problems, to identify adaptive policies that properly address theserootcauses,totestandoptimizetheeffectivenessofpoliciesacrosswideranges of assumptions (i.e., policy robustness), etc. From this perspective, these methods andtechniquesareactuallyjustevolutionaryinnovationsinlinewithearlySDideas. And large-scale adoption of the aforementioned innovations would allow the SD field,andbyextensionthelargersystemsmodelingfield,tomovefrom“experiential art” to “computational science.” Most of the aforementioned innovations are actually integrated in particular SD approaches like in exploratory system dynamics modelling and analysis (ESDMA), which is an SD approach for studying dynamic complexity under deep uncertainty. Deep uncertainty could be defined as a situation in which analysts do not know or cannotagreeon(i)anunderlyingmodel,(ii)probabilitydistributionsofkeyvariables andparameters,and/or(iii)thevalueofalternativeoutcomes(Lempertetal.2003).It isoftenencounteredinsituationscharacterizedbyeithertoolittleinformationortoo much information (e.g., conflicting information or different worldviews). ESDMA isthecombinationofexploratorymodelingandanalysis(EMA),akarobustdecision making, developed during the past two decades (Bankes 1993; Lempert et al. 2000; Bankes2002;Lempertetal.2006)andSDmodeling.EMAisaresearchmethodology fordevelopingandusingmodelstosupportdecisionmakingunderdeepuncertainty. Itisnotamodelingmethod,inspiteofthefactthatitrequirescomputationalmodels. EMA can be useful when relevant information that can be exploited by building computational models exists, but this information is insufficient to specify a single model that accurately describes system behavior (Kwakkel and Pruyt 2013a). In such situations, it is better to construct and use ensembles of plausible models since ensembles of models can capture more of the un/available information than any individualmodel(Bankes2002).Ensemblesofmodelscanthenbeusedtodealwith model uncertainty, different perspectives, value diversity, inconsistent information, etc.—in short, with deep uncertainty.2 InEMA(andthusinESDMA),theinfluenceofaplethoraofuncertainties,including method and model uncertainty, are systematically assessed and used to design policies: sampling and multi-model/multi-method simulation are used to generate ensemblesofsimulationrunstowhichtimeseriesclassificationandmachinelearning techniques are applied for generating insights. Multi-objective robust optimization (Hamaratetal.2013,2014)isusedtoidentifypolicyleversanddefinepolicytriggers, and by doing so, support the design of adaptive robust policies. And regret-based approaches are used to test policy robustness across large ensembles of plausible runs (Lempert et al. 2003). EMA and ESDMA can be performed with TU Delft’s
EMA workbench software, which is an open source tool3 that integrates multimethod, multi-model, multi-policysimulationwithdatamanagement, visualization, and analysis. Thelatterisjustoneoftherecentinnovationsinmodelingandsimulationsoftware and platforms: online modeling and simulation platforms, online flight simulator andgamingplatforms,andpackagesformakinghybridmodelshavebeendeveloped too. And modeling and simulation across platforms will also become reality soon: the eXtensible Model Interchange LanguagE (XMILE) project (Diker and Allen 2005; Eberlein and Chichakly 2013) aims at facilitating the storage, sharing, and combination of simulation models and parts thereof across software packages and acrossmodelingschoolsandmayeasetheinterconnectionwith(real-time)databases, statisticalandanalyticalsoftwarepackages,andorganizationalinformationandcommunicationtechnology(ICT)infrastructures.Notethatthisisalreadypossibletoday with scripting languages and software packages with scripting capabilities like the aforementioned EMA workbench.
5.3.2 Current and Expected Evolutions
Threecurrentevolutionsareexpectedtofurtherreinforcethisshiftfrom“experiential art” to “computational science.” The first evolution relates to the development of “smarter” methods, techniques, and tools (i.e., methods, techniques, and tools that provide more insights and deeper understanding at reduced computational cost). Similar to the development of formal model analysis techniques that smartened the traditional SD approach, new methods, techniques, and tools are currently being developed to smarten modeling and simulation approaches that rely on “brute force” sampling, for example, adaptive output-oriented sampling to span the space of possible dynamics (Islam and Pruyt 2014)orsmartermachinelearningtechniques(Pruytetal.2013;Kwakkeletal.2014; Pruyt et al. 2014c) and time series classification techniques (Yücel and Barlas2011; Yücel 2012; Sucullu andYücel 2014; Islam and Pruyt 2014), and (multi-objective) robust optimization techniques (Hamarat et al. 2013, 2014). Partly related to the previous evolution are developments relates to “big data,” datamanagement,anddatascience.AlthoughtraditionalSDmodelingissometimes calleddata-poormodeling,itdoesnotmeanitis,norshouldbe.SDsoftwarepackages allow one to get data from, and write simulation runs to, databases. Moreover, data are also used in SD to calibrate parameters or bootstrap parameter ranges. But more could be done, especially in the era of “big data.” Big data simply refers here to much more data than was until recently manageable. Big data requires data science techniques to make it manageable and useful. Data science may be used in
modelingandsimulation(i)toobtainusefulinputsfromdata(e.g.,fromreal-timebig data sources), (ii) to analyze and interpret model-generated data (i.e., big artificial data),(iii)tocomparesimulatedandrealdynamics(i.e.,formonitoringandcontrol), and (iv) to infer parts of models from data (Pruyt et al. 2014c). Interestingly, data science techniques that are useful for obtaining useful inputs from data may also be made useful for analyzing and interpreting model-generated data, and vice versa. Online social media are interesting sources of real-world big data for modeling and simulation,bothasinputstomodels,tocomparesimulatedandrealdynamics,andto informmodeldevelopmentormodelselection.Therearemanyapplicationdomains in which the combination of data science and modeling and simulation would be beneficial. Examples, some of which are elaborated below, include policy making with regard to crime fighting, infectious diseases, cybersecurity, national safety and security, financial stress testing, energy transitions, and marketing. Another urgently needed innovation relates to model-based empowerment of decisionmakers.Althoughexistingflightsimulatorandgamingplatformsareusefulfor developing and distributing educational flight simulators and games, and interfaces canbebuiltinSDpackages,usingthemtodevelopinterfacesforreal-worldreal-time decisionmakingandintegratingthemintoexistingICTsystemsisdifficultandtime consuming. In many cases, companies and organizations want these capabilities inhouse, even in their boardroom, instead of being dependent on analyses by external or internal analysts. The latter requires user-friendly interfaces on top of (sets of) models possibly connected to real-time data sources. These interfaces should allow for experimentation, simulation, thoroughly analysis of simulation results, adaptive robust policy design, and policy robustness testing.
5.4 Future State of Practice of Systems Modeling and Simulation
Theserecentevolutionsinmodelingandsimulationtogetherwiththerecentexplosive growthincomputationalpower,data,socialmedia,andotherevolutionsincomputer sciencemayheraldthebeginningofanewwaveofinnovationandadoption,moving the modeling and simulation field from building a single model to simultaneously simulating multiple models and uncertainties; from single method to multi-method and hybrid modeling and simulation; from modeling and simulation with sparse data to modeling and simulation with (near real-time) big data; from simulating and analyzing a few simulation runs to simulating and simultaneously analyzing wellselected ensembles of runs; from using models for intuitive policy testing to using models as instruments for designing adaptive robust policies; and from developing educational flight simulators to fully integrated decision support. For each of the modeling schools, additional adaptations could be foreseen too. In case of SD, it may for example involve a shift from developing purely endogenous to largely endogenous models; from fully aggregated models to sufficiently spatially explicit and heterogenous models; from qualitative participatory modeling
to quantitative participatory simulation; and from using SD to combining problem structuring and policy analysis tools, modeling and simulation, machine learning techniques, and (multi-objective) robust optimization. Adoption of these recent, current, and expected innovations could result in the future state of the art4 of systems modeling as displayed in Fig. 5.1. As indicated by (I) in Fig. 5.1, it will be possible to simultaneously use multiple hypotheses (i.e., simulationmodelsfromthesameordifferenttraditionsorhybrids),fordifferentgoals includingthesearchfordeeperunderstandingandpolicyinsights,experimentationin avirtuallaboratory,future-orientedexploration,robustpolicydesign,androbustness testing under deep uncertainty. Sets of simulation models may be used to represent differentperspectivesorplausibletheories, todealwithmethodologicaluncertainty, or to deal with a plethora of important characteristics (e.g., agent characteristics, feedback and accumulation effects, spatial and network effects) without necessarily havingtointegratetheminasinglesimulationmodel.Themainadvantagesofusing multiple models for doing so are that each of the models in the ensemble of models remains manageable and that the ensemble of simulation runs generated with the
ensemble of models is likely to be more diverse which allows for testing policy robustness across a wider range of plausible futures. Some of these models may be connected to real-time or near real-time data streams, and some models may even be inferred in part with smart data science tools from data sources (see (II) in Fig. 5.1). Storing the outputs of these simulation models in databases and applying data science techniques may enhance our understanding, may generate policy insights, and may allow for testing policy robustness across large multidimensional uncertainty spaces (see (III) in Fig. 5.1). And userfriendly interfaces on top of these interconnected models may eventually empower policy makers, enabling them to really do model-based policy making. Note,however,thattheintegratedsystemsmodelingapproachsketchedinFig.5.1 may only suit a limited set of goals, decision makers, and issues. Single model simulation properly serves many goals, decision makers, and issues well enough for multi-model/multi-method, data-rich, exploratory, policy-oriented approaches not to be required. However, there are most certainly goals, decision makers, and issues that do.
5.5 Examples
Although all of the above is possible today, it should be noted that this is the current state of science, not the state of common practice yet. Applying all these methods and techniques to real issues is still challenging, and shows where innovations are most needed. The following examples illustrate what is possible today as well as what the most important gaps are that remain to be filled. The first example shows that relatively simple systems models simulated under deep uncertainty allow for generating useful ensembles of many simulation runs. Using methods and techniques from related disciplines to analyze the resulting artificialdatasetshelpstogenerateimportantpolicyinsights.Andsimulationofpolicies acrosstheensemblesallowstotestforpolicyrobustness.Thisfirstcasenevertheless shows that there are opportunities for multi-method and hybrid approaches as well as for connecting systems models to real-time data streams. The second example extends the first example towards a system-of-systems approachwithmanysimulationmodelsgeneratingevenlargerensemblesofsimulation runs. Smart sampling and scenario discovery techniques are then required to reduce the resulting data sets to manageable proportions. The third example shows a recent attempt to develop a smart model-based decision-support system for dealing with another deeply uncertain issue. This example shows that it is almost possible to empower decision makers. Interfaces with advancedanalyticalcapabilitiesaswellaseasierandbetterintegrationwithexisting ICT systems are required though. This example also illustrates the need for more advanced hybrid systems models as well as the need to connect systems models to real-time geo-spatial data.
5.5.1 Assessing the Risk, and Monitoring, of New Infectious Diseases
Thefirstcase,whichisdescribedinmoredetailin(PruytandHamarat2010;Pruytet al. 2013), relates to assessing outbreaks of new flu variants. Outbreaks of new (variants of) infectious diseases are deeply uncertain. For example, in the first months afterthefirstreportsabouttheoutbreakofanewfluvariantinMexicoandtheUSA, muchremainedunknownaboutthepossibledynamicsandconsequencesofthispossible epidemic/pandemic of the new flu variant, referred to today as new influenza A(H1N1)v.Table 5.1 shows that more and better information became available over time,butalsothatmanyuncertaintiesremained.However,evenwiththeseremaining uncertainties, it is possible to model and simulate this flu variant under deep uncertainty, for example with the simplistic simulation model displayed in Fig. 5.2, since flu outbreaks can be modeled. Simulating this model thousands of times over very wide uncertainty ranges for each of the uncertain variables generates the 3D cloud of potential outbreaks displayed in Fig. 5.3a. In this figure, the worst flu peak (0–50 months) is displayed on the X-axis, the infected fraction during the worst flu peak (0–50%) is displayed on the Y-axis, and the cumulative number of fatal cases in the Western world (0– 50.000.000)isdisplayedontheZ-axis.This3Dplotshowsthatthemostcatastrophic outbreaks are likely to happen within the first year or during the first winter season followingtheoutbreak. Usingmachinelearningalgorithmstoexplorethisensemble of simulation runs helps to generate important policy insights (e.g., which policy levers to address). Testing different variants of the same policy shows that adaptive policies outperform their static counterparts (compare Fig. 5.3b and c). Figure 5.3d finally shows that adaptive policies can be further improved using multi-objective robust optimization. However,takingdeepuncertaintyseriouslyintoaccountwouldrequiresimulating more than a single model from a single modeling method: it would be better to simultaneouslysimulateCAS,ABM,SD,andhybridmodelsunderdeepuncertainty and use the resulting ensemble of simulation runs. Moreover, near real-time geospatial data (from twitter, medical records, etc.) may also be used in combination with simulation models, for example, to gradually reduce the ensemble of modelgenerated data. Both suggested improvements would be possible today.
5.5.2 Integrated Risk-CapabilityAnalysis under Deep Uncertainty
The second example relates to risk assessment and capability planning for National Safety and Security. Since 2001, many nations have invested in the development of all-hazard integrated risk-capability assessment (IRCA) approaches. All-hazard IRCAsintegratescenario-basedriskassessment, capabilityanalysis, andcapabilitybased planning approaches to reduce all sorts of risks—from natural hazards, over technical failures to malicious threats—by enhancing capabilities for dealing with
them. Current IRCAs mainly allow dealing with one or a few specific scenarios for a limited set of relatively simple event-based and relatively certain risks, but not for dealing with a plethora of risks that are highly uncertain and complex, combinationsofmeasuresandcapabilitieswithuncertainanddynamiceffects, anddivergent opinions about degrees of (un)desirability of risks and capability investments. Thenextgenerationmodel-basedIRCAsmaysolvemanyoftheshortcomingsof theIRCAsthatarecurrentlybeingused. Figure5.4displaysanextgenerationIRCA for dealing with all sorts of highly uncertain dynamic risks. This IRCA approach, described in more detail in Pruyt et al. (2012), combines EMA and modeling and simulation, both for the risk assessment and the capability analysis phases. First, risks—like outbreaks of new flu variants—are modeled and simulated many times acrosstheirmultidimensionaluncertaintyspacestogenerateanensembleofplausible risk scenarios for each of the risks. Time series classification and machine learning techniques are then used to identify much smaller ensembles of exemplars that are representativeforthelargerensembles.Theseensemblesofexemplarsarethenused as inputs to a generic capability analysis model. The capability analysis model is subsequently simulated for different capabilities strategies under deep uncertainty (i.e., simulating the uncertainty pertaining to their effectiveness) over all ensembles ofexemplarstocalculatethepotentialofcapabilitiesstrategiestoreducetheserisks.
Fig. 5.4 Model-based integrated risk-capability analysis (IRCA)
Finally, multi-objective robust optimization helps to identify capabilities strategies that are robust. Not only does this systems-of-systems approach allow to generate thousands of variants per risk type over many types of risks and to perform capability analyses across all sorts of risk and under uncertainty, it also allows one to find sets of capabilities that are effective across many uncertain risks. Hence, this integrated model-basedapproachallowsfordealingwithcapabilitiesinanall-hazardwayunder deep uncertainty. This approach is currently being smartened using adaptive output-oriented sampling techniques and new time-series classification methods that together help to identify the largest variety of dynamics with the minimal amount of simulations. Covering the largest variety of dynamics with the minimal amount of exemplars is desirable,forperformingautomatedmulti-hazardcapabilityanalysisovermanyrisks is—due to the nature of the multi-objective robust optimization techniques used— computationally very expensive. This approach is also being changed from a multimodel approach into a multi-method approach. Whereas, until recently, sets of SD models were used; there are good reasons to extend this approach to other types of systemsmodelingapproachesthatmaybebettersuitedforparticularrisksor—using multipleapproaches—helptodealwithmethodologicaluncertainty.Finally,settings of some of the risks and capabilities, as well as exogenous uncertainties, may also be fed with (near) real-world data.
5.5.3 Policing Under Deep Uncertainty
The third example relates to another deeply uncertain issue, high-impact crimes (HIC).An SD model and related tools (see Fig. 5.5) were developed some years ago inviewofincreasingtheeffectivenessofthefightagainstHIC,morespecificallythe fightagainstrobberyandburglary.HICsrequireasystemicperspectiveandapproach:
Thesecrimesarecharacterizedbyimportantsystemiceffectsintimeandspace,such as learning and specialization effects, “waterbed effects” between different HICs and precincts, accumulations (prison time) and delays (in policing and jurisdiction), preventive effects, and other causal effects (ex-post preventive measures). HICs are also characterized by deep uncertainty: Most perpetrators are unknown and even though their archetypal crime-related habits may be known to some extent at some pointintime, accuratetimeandgeographicallyspecificpredictionscannotbemade. At the same time, is part of the HIC system well known and is a lot of real-world information related to these crimes available. ImportantplayersintheHICsystembesidesthepoliceand(potential)perpetrators are potential victims (households and shopkeepers), partners in the judicial system (the public prosecution service, the prison system, etc.). Hence, the HIC system is dynamically complex, deeply uncertain, but also data rich, and contingent upon external conditions. Themaingoalsofthispilotprojectweretosupportstrategicpolicymakingunder deep uncertainty and to test and monitor the effectiveness of policies to fight HIC. The SD model (I) was used as an engine behind the interface for policy makers (II) to explore plausible effects of policies under deep uncertainty and identify realworld pilots that could possibly increase the understanding about the system and effectivenessofinterventions(III),toimplementthesepilots(IV),andmonitortheir outcomes (V). Real-world data from the pilots and improved understanding about the functioning of the real system allow for improving the model.
Today, a lot of real-world geo-spatial information related to HICs is available online and in (near) real time which allows to automatically update the data and model, and hence, increase its value for the policy makers. The model used in this project was an ESDMA model. That is, uncertainties were included by means of sets of plausible assumptions and uncertainty ranges. Although this could already be argued to be a multi-model approach, hybrid models or a multi-method approach would really be needed to deal more properly with systems, agents, and spatial characteristics. Moreover, better interfaces and connectors to existing ICT systems and databases would also be needed to turn this pilot into a real decision-support system that would allow chiefs of police to experiment in a virtual world connected to the real world, and to develop and test adaptive robust policies on the spot.
5.6 Conclusions
Recent and current evolutions in modeling and simulation together with the recent explosive growth in computational power, data, social media, and other evolutions in computer science have created new opportunities for model-based analysis and decision making. Multi-method and hybrid modeling and simulation approaches are being developed to make existing modeling and simulation approaches appropriate for dealing withagentsystemcharacteristics,spatialandnetworkaspects,deepuncertainty,and otherimportantaspects. Datascienceandmachinelearningtechniquesarecurrently beingdevelopedintotechniquesthatcanprovideusefulinputsforsimulationmodels as well as for building models. Machine learning algorithms, formal model analysis methods, analytical approaches, and new visualization techniques are being developedtomakesenseofmodelsandgenerateusefulpolicyinsights.Andmethodsand tools are being developed to turn intuitive policy making into model-based policy design. Some of these evolutions were discussed and illustrated in this chapter. Itwasalsoarguedandshownthateasierconnectorstodatabases, tosocialmedia, toothercomputerprograms,andtoICTsystems,aswellasbetterinterfacingsoftware need to be developed to allow any systems modeler to turn systems models into real decision-support systems. Doing so would turn the art of modeling into the computational science of simulation. It would most likely also shift the focus of attention from building a model to using ensembles of systems models for adaptive robust decision making.
References
Bankes SC (1993) Exploratory modeling for policy analysis. Operat Res 41(3):435–449 BankesSC(2002)Toolsandtechniquesfordevelopingpoliciesforcomplexanduncertainsystems. Proc NatlAcad Sci U SA 99(3):7263–7266
5 From Building a Model toAdaptive Robust Decision Making Using Systems Modeling 91
BarlasY(1996)Formalaspectsofmodelvalidityandvalidationinsystemdynamics.SystDynRev 12(3):183–210 BenDor TK, Kaza N (2012)A theory of spatial system archetypes. Syst Dyn Rev 28(2):109–130 BryantBP,LempertRJ(2010)Thinkinginsidethebox:aparticipatory,computer-assistedapproach to scenario discovery. Technol Forecast Soc Change 77(1):34–49 Castillo D, SayselAK (2005) Simulation of common pool resource field experiments: a behavioral model of collective action. Ecol Econ 55(3):420–436 ClemsonB,TangY,PyneJ,UnalR(1995)Efficientmethodsforsensitivityanalysis.SystDynRev 11(1):31–49 Coyle GR (1985) The use of optimization methods for policy design in a system dynamics model. Syst Dyn Rev 1(1):81–91 Coyle GR (1999) Simulation by repeated optimisation. J Operational Res Society 50:429–438 DikerVG,Allen RB (2005) XMILE: towards an XML interchange language for system dynamics models. Syst Dyn Rev 21(4):351–359 Dogan G (2007) Bootstrapping for confidence interval estimation and hypothesis testing for parameters of system dynamics models. Syst Dyn Rev 23(4):415–436 Doyle JK, Ford DN (1998) Mental models concepts for system dynamics research. Syst Dyn Rev 14:3–29 Doyle JK, Ford DN (1999) Mental models concepts revisited: some clarifications and a reply to Lane. Syst Dyn Rev 15:411–415 Eberlein RL, Chichakly KJ (2013) XMILE: a new standard for system dynamics. Syst Dyn Rev 29(3):188–195 Eker S, Slinger JH, Daalen EC, Yücel G (2014) Sensitivity analysis of graphical functions. Syst Dyn Rev 30(3):186–205 Feola G, Gallati JA, Binder CR (2012) Exploring behavioural change through an agent-oriented system dynamics model: the use of personal protective equipment among pesticide applicators in Colombia. Syst Dyn Rev 28(1):69–93 Fallah-FiniS,RahmandadR,ChenHJ,WangY(2014)Connectingmicrodynamicsandpopulation distributions in system dynamics models. Syst Dyn Rev 29(4):197–215 FiddamanTS(2002)Exploringpolicyoptionswithabehavioralclimate–economymodel.SystDyn Rev 18(2):243–267. Ford A (1990) Estimating the impact of efficiency standards on the uncertainty of the northwest electric system. Oper Res 38(4):580–597 Ford A, Flynn H (2005) Statistical screening of system dynamics models. Syst Dyn Rev 21(1): 273–303 Forrester JW (1961) Industrial dynamics. MIT Press, Cambridge (Waltham, MA: Pegasus Communications.AND Productivity Press: Portland, OR) Forrester JW (1968) Principles of systems.Wright-Allen Press, Cambridge Forrester JW (1969) Urban dynamics. MIT Press, Cambridge Forrester JW (1991) The beginning of system dynamics. D-Note D-4165–1 Forrester JW (2007) System dynamics—a personal view of the first fifty years. Syst Dyn Rev 23(2–3):345–358 GrahamAK,ArizaCA(2003)Dynamic,hardandstrategicquestions:usingoptimizationtoanswer a marketing resource allocation question. Syst Dyn Rev 19(1):27–46 Hamarat C, Kwakkel JH, Pruyt E (2013) Adaptive robust design under deep uncertainty. Technol Forecast Soc Change 80(3):408–418 Hamarat C, Kwakkel JH, Pruyt E, Loonen ET (2014) An exploratory approach for adaptive policymaking by using multi-objective robust optimization. Simul Model Pract Theory 46:25–39 HearneJW(2010)Anautomatedmethodforextendingsensitivityanalysistomodelfunctions. Nat Resour Model 23(2):107–120 HomerJB(2012)Modelsthatmatter:selectedwritingsonsystemdynamics1985–2010.Grapeseed Press, Barrytown
92 E. Pruyt
