Impossible by Conventional Means: Ten Years on from the DARPA Red Balloon Challenge
IImpossible by Conventional Means: Ten Years onfrom the DARPA Red Balloon Challenge
Alex Rutherford , Manuel Cebrian , Inho Hong , and Iyad Rahwan Center for Humans and Machines, Max Planck Institute for Human Development, Berlin, Germany * To whom correspondence should be addressed: [email protected], [email protected] + These authors contributed equally to this work
ABSTRACT
Ten years ago, DARPA launched the ‘Network Challenge’, more commonly known as the ‘DARPA Red Balloon Challenge’. Tenred weather balloons were fixed at unknown locations in the US. An open challenge was launched to locate all ten, the first todo so would be declared the winner receiving a cash prize. A team from MIT Media Lab was able to locate them all within 9hours using social media and a novel reward scheme that rewarded viral recruitment . This achievement was rightly seen asproof of the remarkable ability of social media, then relatively nascent, to solve real world problems such as large-scale spatialsearch. Upon reflection, however, the challenge was also remarkable as it succeeded despite many efforts to provide falseinformation on the location of the balloons. At the time the false reports were filtered based on manual inspection of visualproof and comparing the IP addresses of those reporting with the purported coordinates of the balloons. In the ten yearssince, misinformation on social media has grown in prevalence and sophistication to be one of the defining social issues ofour time. Seen differently we can cast the misinformation observed in the Red Balloon Challenge, and unexpected adverseeffects in other social mobilisation challenges subsequently, not as bugs but as essential features. We further investigate therole of the increasing levels of political polarisation in modulating social mobilisation. We confirm that polarisation not onlyimpedes the overall success of mobilisation, but also leads to a low reachability to oppositely polarised states, significantlyhampering recruitment. We find that diversifying geographic pathways of social influence are key to circumvent barriers ofpolitical mobilisation and can boost the success of new open challenges. The DARPA Red Balloon Challenge and its Progeny
The DARPA Network Challenge began a series of open challenges (see Table 1) that explored different facets of socialmobilisation. The challenge conditions were able to incentivise researchers to focus on particular problems in sectors as diverseas intelligence, health and problem solving with the lure of prestige or financial reward. In each case the time constraints andhigh-profile nature of the challenge pushed the parameters of the task to extremes uncovering sometimes unsettling adversarialeffects.Following the Red Balloon Challenge, DARPA sponsored the Shredder Challenge to understand vulnerabilities in thecrowd-sourced reconstruction of fragments of shredded documents. A team led by the member of the MIT team which won theDARPA Network Challenge (by then at the University of California, San Diego) was the top crowdsourcing solution, yet failedto complete the challenge, partly due to a coordinated series of attacks that sabotaged the progress made by other contributors .One year later the US State Department sponsored the Tag Challenge, in which teams were invited to locate 5 mobilehuman ‘targets’ in 5 different cities on a given day . This challenge was won by a team led from Masdar Institute in theUAE in collaboration with same member of the MIT team which won the DARPA Network Challenge. Despite its success, theteam’s progress was hindered by several adversarial activities such as sabotage of the team’s platform and impersonation of theteam’s identity on social media.In 2017, the Viral Communications research group in MIT Media Lab launched the FiftyNifty challenge . The objective ofwhich was to mobilise voters in each of the 50 states to contact their representatives through viral recruitment. Building on thesuccessful modeling of social mobilisation phenomena we show that the success of such a recruitment process is susceptibleto political polarisation. A higher degree of polarisation not only decreased the overall success rate of the challenge (Fig. 1a)but also made it harder to reach oppositely polarised states (see Figs. S2 and S3 in the Supplementary Information).These adversarial activities were unexpected outcomes of these challenges, but they were surprisingly prescient of modernday adversarial activities using social media and digital platforms. While it might be tempting to favour fewer such challengesthat might encourage such side-effects, we conclude the opposite. These examples have shown that we need more of thesechallenges to bring possible side effects to prominence . Given the unprecedented complexity of AI and human ecologies, wesee a need for open challenges focusing on deployment of AI systems into society. This is consistent with recent calls for a a r X i v : . [ phy s i c s . s o c - ph ] A ug hallenge Year Theme Unexpected Lesson DARPA Red Network Challenge 2009 Social Search Misinformation is inevitable
DARPA Shredder Challenge 2011 Problem Solving Sabotage has asymmetric power
Nexus 7 (More Eyes) 2011 Intelligence Privacy concerns restrict uptake My HeartMap 2011 Health Social media needs mass Media State Department Tag Challenge 2012 Social Search Misinformation and sabotage are in-evitable
CLIQR Quest challenge 2012 Social Search Social media needs mass media &Misinformation and sabotage are in-evitable Langley Castle Challenge 2014 Social Search More similar (homophilous) friends mobilizefaster
WeHealth 2015 Health Privacy concerns restrict uptake FiftyNifty 2017 Political Mobilisation Political polarisation undermines bi-partisanmobilization Black Rock Atlas 2018 Social Search Social media needs mass media Table 1.
Open challenges and the lessons.dedicated study of the sociology of AI agents . Mobilisation impeded by political polarisation
Based on the negative experience in the Fifty-Nifty challenge, we now investigate the effect of polarisation on mobilisation.Political polarisation was not present at the current level for early challenges, so it is important to understand how it couldaffect future ones. For that, we simulate a social mobilisation process to show how political polarisation of individuals mayimpede mobilisation over the entire country. In this simulation, we show (1) the success rate impeded by polarisation, (2) thereachability to each state by polarisation and (3) good and bad seed states for mobilisation.We reconstruct mobilisation in the FiftyNifty challenge by simulating the branching recruitment process that spreadsthrough the county-wise friendship network given by the Facebook Social Connectedness Index (SCI) dataset on the politicallandscape of the Unities States . This novel dataset of Facebook SCI represents the normalized counts of friendship pairsbetween the entire US counties, and the political landscape is based on the result of the 2016 Presidential Election. In thesimulation, a political polarisation parameter α ranging from 0 (no polarization) to 1 (full polarization) controls the successprobability of recruits between a recruiter and a recruited as p = − α for opposite polarisation in contrast to p = α where some level of mobilisation in each state isrequired for success. As a result, we observe a monotonic decrease of the mobilisation success with increasing polarisation (seeFig. 1a). This finding shows that political polarisation between individuals hampers mobilisation across different regions. The hardest place to reach
So far, we observe the overall decrease of the success rate by polarisation. Then, how does polarisation affect mobilisation overdifferent states? We first check the relation of the mobilisation size and the population size of states. Fig. S1 shows a highcorrelation between mobilisation and population. Political polarisation decreases the mobilisation size in every state. Then,does polarisation bring equal impacts to states?To see the polarisation effect in different regions, we compare the mobilisation size of each state with its political makeup.Fig. S2 shows that Democratic states have a larger mobilisation size compared to Republican states in general. Therefore,both polarisation and population affect the success of mobilisation. This combination leads to a low reachability to small andRepublican state; For example, Wyoming is the hardest place to reach for a campaign started from Democrats in Massachusetts.By decomposing the population trend from the mobilisation size, we show the effect of polarisation on the success ofmobilisation in each state. We subtract the population trend in Fig. S1 from the mobilisation size. As a result, Fig. S3shows a clear separation between states in the mobilisation size by the political makeup. Also, higher polarisation makes a .0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0
Polarisation M ob ili s a t i on s u cc e ss ac b A K U T C O M T N M A Z M N W A I D W Y O R S D I A M E N V H I V T N E C A N D T X O K I L I N O H W I V A A R M O N H L A K S M A K Y C T W V M I P A
D C F L S C N C M D N Y G A A L R I M S N J T N D E Seed state S u cc e ss r a t e Fraction of democrats
Figure 1. (a) Rate of success in the simulation as a function of political polarisation. Our simulation reconstructs themobilisation process in the FiftyNifty challenge. The success rate is calculated from 1,000 simulations for each polarisationparameter α . (b) Friendship diversity and rate of success of seed states. The diversity measured from the entropy of friendshipweights to the other states shows high correlation (i.e., ρ = .
54) with the success rate of mobilisation from the seed state. (c)Size of mobilisation for different seed states for moderate polarisation (i.e., α = . What determines a good seed state?
The simulation so far was focused on reproducing the mobilisation process in the FiftyNifty challenge seeded from MiddlesexCounty, MA. If we launch a new challenge or campaign using the polarised friendship network, where is the best seed state forits success? To answer this question, we simulate the mobilisation process seeded from each state in turn. We assume thatDemocratic seeds are located in the most populated county of each state. Fig 1c shows the success rate of mobilisation fromdifferent seed states; Alaska is the best seed state while Delaware is the worst seed state. The variability of seed quality is notdetermined by the political makeup of states.Then, why are some states better than the others in seeding the mobilisation process? We find the answer in the connectiondiversity of each seed state. We define the friendship diversity of each state using the entropy of connection probability to otherstates as H i = − ∑ j p i j log p i j (1)where p i j = w i j / ∑ j w i j is the connection probability from state i to state j , and w i j is the aggregated weight of friendship onFacebook. As a result, Fig. 1b shows a high correlation (i.e., ρ = .
54 with p < − ) between connection diversity H i of eed state i and its success rate of mobilisation. This finding gives us two implications for successful mobilisation. First, thelocation of seeds is important in mobilisation. As the mean branching factor is less than 1 (i.e., the number of new recruits by arecruiter is less than 1 on average.), the depth of the recruitment network is finite. Thus, the first generation of recruits in therecruitment network plays a key role in overall mobilisation. Accordingly, the connection diversity of a seed state controlsthe spread of mobilisation to other states through a friendship network. When a seed state is evenly connected to the otherstates, mobilisation is easy to spread to many states. On the contrary, if a seed state has strongly biased connections to a fewstates, these states would take up most of the outgoing recruitment flows from the seed state. This finding demonstrates thatdiversification of the pathways of social influence would be the key to the success of new open challenges. Societal Testing
The allure of open challenges is clear when considering how other new technologies are developed, tested and launched. Inall cases, tests are first performed in a controlled environment and subsequently the test environment is gradually enlarged tosuccessively model a real world deployment more closely. For example, software is unit tested, integration tested and systemlevel tested, and subsequently, when in production, is subjected to user-driven bug reports and hacker-driven penetration testing.Likewise, new drugs may be tested on animals, small scale medical trials followed by open usage among certain demographics.For technologies such as drugs, regulation is well developed and the procedure for release is well understood. Thusfollowing medical trials, once it is established with reasonable certainty and noting reasonable caveats, that the drug’s benefitsoutweigh its harms when properly administered, it is freely released for usage within society with strict usage instructions. Themajority of testing is pre-deployment and limited testing is done post-deployment , for example looking at longitudinal effects orinteractions with other drugs.
Pre-deploymentmore prescribed
Post-deploymentless prescribedTECHNOLOGYPRE-DEPLOYMENT
Medical trialsDrugs System testingSoftware systems
Open challengesArtificial intelligence
Figure 2.
A prescriptive spectrum of technologies and testing methodologies.For software systems, there exists slightly more balance between pre- and post-deployment testing. Once technicalfunctionality has been tested and established, user testing takes place to see if the intended functionality persists under realisticusage patterns in a trial deployment. Following deployment into production, testing continues through continuous telemetry,A/B tests of new features and vulnerability tests through bug bounties and sanctioned white hat hacking. The reason whypost-production testing is more thorough, is that software is an example of a socio-technical system . Typically of suchsystems, the behaviour of computer software is determined by the coupled behaviour of its users and groups of users, in anon-trivial manner and such that testing of its behaviour cannot be conducted in isolation and without the realistic usage patternsof the user.Such is the complexity and versatility of social media platforms and other Internet mediated technologies, the expectedbehaviour of e.g. many-to-many instant messaging, Virtual Reality or DeepFakes cannot simply be interpolated from a smallsubsample, or a limited user testing. While some might object to open challenges as being reckless and unsanctioned societalexperimentation, we can ask what are the alternatives? Internet platform companies have resorted to closed focus groupsfollowed by the global release of largely unregulated tools in an arms race of engagement; a process that has been described as society as a beta test’.Reflecting on 10 years of social media challenges, we argue that emerging technologies represent socio-technical systemsof such complexity and with such far reaching effects that suitable testing should emphasise post-deployment testing as muchas pre-deployment testing. The challenges discussed above can be considered as a forum for transparent, white hat attacks onthese platforms operating at the limits of their parameters. Below we reflect upon the common characteristics of successfulchallenges. • Loosely defined and possibly tangential : The Red Balloon Challenge was not explicitly about social media, yet thisemerged as the most appropriate medium. It is also important that challenges are sufficiently loosely defined to allow thesystem freedom to unearth novel behaviours. • Non-prescriptive : Challenge platforms such as Kaggle incentivise incremental improvements in cases when we haveconverged upon a working solution. Many of the challenges described above failed or only partially succeeded. • Failure is likely (and fine) : following from the previous point, the challenge should shift emphasis from successfully(or not) completing the stated objective, but uncovering and reporting unexpected side-effects and adversarial efforts. • Has low barrier to entry : to maximise the freedom for solutions to explore the far reaches of the parameter spacechallenges should not impose onerous conditions or an entry fee. Diverse connectivity through open participation is thekey to successful mobilisation (see Fig. 1b). • Public : challenges should allow for public and transparent publications of findings to enable society to engage with thepotential risks and benefits. For example, the Red Balloon challenge findings were published in
Science and
PNAS aswell as receiving publicity on
The Colbert Show and in
The New York Times . • Prestigious : Successful challenges should be sanctioned by prestigious public institutions such as DARPA or the StateDepartment. This encourages uptake and participation of competitors, users and adversaries.These challenges that followed the DARPA balloon challenge brought to light some unsavoury faces of Internet mediatedtechnologies. Aspects that have become all too familiar: sabotage, polarisation and misinformation. Yet we are in a better placehaving illuminated and characterised these ‘edge cases’, as they were considered at the time, and better placed to deal withthem now that they are mainstream. Just as we began to understand and characterise the power and pitfalls of social media withthe Red Balloon Challenge a decade ago, we call urgently for a further decade of challenges investigating AI as it emerges as amainstream and general purpose technology.
Contributions
M.C. conceived the study. M.C., A.R. designed the study. A.R., I.H. developed the computational model. I.H. performedsimulations. M.C., A.R., I.H. analyzed data. A.R. wrote the manuscript. M.C., A.R., I.R., I.H. edited the manuscript.
Acknowledgements
We are grateful to May Al-Hazzani for early investigations of the FiftyNifty data, to Andrew Lippmann, Léopold Mebazaa,Travis Rich, Jasmin Rubinovitz, Simon Johnson, Jason Seibel and Penny Webb for the dataset, and to Jürgen Rossbach forgraphics editing.
References Cebrian, M., Coviello, L., Vattani, A. & Voulgaris, P. Finding red balloons with split contracts: robustness to individuals’selfishness. In
Proceedings of the forty-fourth annual ACM symposium on Theory of computing , 775–788 (2012). Tang, J. C. et al.
Reflecting on the DARPA Red Balloon Challenge.
Commun. ACM , 78–85 (2011). Pickard, G. et al.
Time-critical social mobilization.
Science , 509–512 (2011). Chen, H., Rahwan, I. & Cebrian, M. Bandit strategies in social search: the case of the DARPA red balloon challenge.
EPJData Sci. , 20 (2016). Stefanovitch, N., Alshamsi, A., Cebrian, M. & Rahwan, I. Error and attack tolerance of collective problem solving: TheDARPA Shredder Challenge.
EPJ Data Sci. , 13 (2014). . Naroditskiy, V., Jennings, N. R., Van Hentenryck, P. & Cebrian, M. Crowdsourcing contest dilemma.
J. The Royal Soc.Interface , 20140532 (2014). Oishi, K., Cebrian, M., Abeliuk, A. & Masuda, N. Iterated crowdsourcing dilemma game.
Sci. reports , 4100 (2014). Weinberger, S.
The imagineers of war: the untold history of DARPA, the Pentagon agency that changed the world (Knopf,2017). Rutherford, A. et al.
Targeted social mobilization in a global manhunt.
PloS one , e74628 (2013). Rahwan, I. et al.
Global manhunt pushes the limits of social mobilization.
Computer , 68–75 (2012). Naroditskiy, V., Rahwan, I., Cebrian, M. & Jennings, N. R. Verification in referral-based crowdsourcing.
PloS one ,e45924 (2012). Social media falls short in DARPA’s CLIQR challenge. https://gcn.com/articles/2012/03/20/agg-darpa-cliqr-challenge-no-winner.aspx. Accessed: 2020-03-20.
Alstott, J., Madnick, S. & Velu, C. Homophily and the speed of social mobilization: the effect of acquired and ascribedtraits.
PloS one , e95140 (2014). Wang, J., Madnick, S., Li, X., Alstott, J. & Velu, C. Effect of media usage selection on social mobilization speed: Facebookvs e-mail.
PloS one , e0134811 (2015). Epstein, Z. et al.
Towards a new social laboratory: An experimental study of search through community participation atburning man. arXiv preprint arXiv:1903.04125 (2019).
Rutherford, A. et al.
Limits of social mobilization.
Proc. Natl. Acad. Sci. , 6281–6286 (2013).
Rahwan, I. et al.
Machine behaviour.
Nature , 477–486 (2019).
Bailey, M., Cao, R., Kuchler, T., Stroebel, J. & Wong, A. Social connectedness: Measurement, determinants, and effects.
J.Econ. Perspectives , 259–80 (2018). Vespignani, A. Modelling dynamical processes in complex socio-technical systems.
Nat. physics , 32–39 (2012). Bakshy, E., Messing, S. & Adamic, L. A. Exposure to ideologically diverse news and opinion on facebook.
Science ,1130–1132 (2015). upplementary Information
Simulating polarised mobilisation
The overall mobilisation process adopts the branching recruitment process in Rutherford et al . Starting from a group of seedrecruiters, mobilisation occurs repeatedly over a social network until a goal of mobilisation is reached (i.e., success in openchallenges) or there is no more recruits. The prominent improvements of our simulation model from the original mobilisationmodel are (1) the implement of political polarisation modeled from literature , (2) an empirical social network from theFacebook Social Connectedness Index (SCI) dataset and (3) an empirical political landscape from the 2016 US PresidentialElection . The following description shows how these improvements are implemented in our simulation process of seeding,activation, recruitment and termination. Seeding.
The simulation starts from N s Democratic or Republican individual seeds in a specific county of US. The simulationfor the FiftyNifty challenge is seeded from Middlesex County (the location of Cambridge, MA from where the challenge waslaunched). When we simulate mobilisation for different seed states, we chose the most populated county of each state as theseed counties. The activation time ∆ t a ( i ) of seed i is determined by a log-normal distribution P ( ∆ t a ) with a mean of 1.5 day anda standard deviation of 5.5 days . The individual information is inserted into a priority queue, and is drawn one-by-one in theorder of the shortest activation time. Also, the number of friends k (i.e., branching factor) for recruitment is assigned to eachindividual seed following a Harris discrete distribution P ( k ) as, P ( k ) = H ab b + k a , (S1)where a is a power-law exponent, b is fitted to a given empirical mean value of branching factors, and H ab is a normalisationfactor. In the simulation, we use a = . (cid:104) k (cid:105) = . . Activation.
In each simulation step, one individual is drawn from the queue in the order of the shortest activation time. Thisactivated individual is mobilised with a probability of 1 if it has identical polarisation with its recruiter. In the case of oppositepolarisation with its recruiter, the mobilisation probability is given as 1 − α where α is the polarisation parameter ranging from0 to 1. The mobilisation probability of seeds is given to 1. If mobilisation is successful, the activated individual further recruitsits friends by the following recruitment process. Recruitment.
A mobilised individual recruits k friends following the branching factor k which was assigned in the previousrecruitment or seeding. Each friend has 4 stochastically chosen properties: activation time, branching factor, residence countyand the political orientation. • The waiting time for activation ∆ t a is chosen by the log-normal distribution in Seeding . If the current time is t , therecruited friend is actiavated at t + ∆ t a . • The branching factor is determined by the Harris distribution in Eq. (S1). As the mean of branching factors is less thanone, the branching recruitment process terminates eventually. • The residence county of each friend is determined by the strength of friendship given by the Facebook dataset. If arecruiter is in county i , the probability of recruiting a friend in county j is p i j = w i j / ∑ j w i j , where w i j is the SocialConnectedness Index (SCI) that represents the number of friend pairs in the Facebook dataset . As the goal of theFiftyNifty challenge was to mobilize every state, recruits within the same state are prohibited in the simulation. • The political orientation of a recruited friend is chosen by the combination of the political makeup of its residence countyand polarisation. Political polarisation between friends is known to be identical with a probability of 3 / / . Also, the probability of the political orientation is proportional to the politicalmakeup of the residence county given by the 2016 US Presidential Election . Combining them leads to ( p dem , p rep ) ∼ (cid:40) ( pol j , ( − pol j )) if the recruiter is Democratic , ( pol j , ( − pol j )) if the recruiter is Republican , (S2)where p dem = − p rep is the probability of being Democratic and pol j is the proportion of votes to the Democratic partyin the 2016 US Presidential Election in residence county j of the recruited. Termination.
The simulation terminates when we observe at least one mobilised individual in every state (i.e., “Success”), orthere is no remaining individual to be activated in the queue (i.e., “Failure”). .0 0.2 0.4 0.6 0.8 1.0
State population S i z eo f m ob ili s a t i on = 1.0= 0.5= 0.0MA Figure S1.
Size of mobilisation as a function of population size for different levels of polarisation ( α = { . , . , . } ). Theresult is the average of 500 simulations for 1,000 Democratic seeds in Middlesex County, MA. The error bars denote the 95%interval of the mobilisation size. MassachusettsNew YorkCaliforniaFloridaTexasPennsylvaniaIllinoisNew JerseyGeorgiaNorth CarolinaVirginiaNew HampshireConnecticutMarylandOhioWashingtonMichiganColoradoDistrict of ColumbiaArizonaTennesseeRhode IslandSouth CarolinaIndianaMissouriMaineMinnesotaWisconsinOregonLouisianaAlabamaNevadaVermontKentuckyMississippiIowaUtahKansasOklahomaHawaiiArkansasDelawareNew MexicoNebraskaWest VirginiaIdahoMontanaAlaskaNorth DakotaSouth DakotaWyoming S t a t e Polarisation = 1.0 F r a c t i on o f de m o c r a t s Size of mobilisation
MassachusettsCaliforniaNew YorkFloridaTexasPennsylvaniaIllinoisNew JerseyGeorgiaNorth CarolinaVirginiaOhioMarylandNew HampshireConnecticutMichiganColoradoWashingtonTennesseeArizonaSouth CarolinaIndianaMissouriWisconsinOregonDistrict of ColumbiaAlabamaMinnesotaLouisianaRhode IslandNevadaMaineKentuckyMississippiOklahomaKansasIowaUtahArkansasVermontHawaiiDelawareNew MexicoWest VirginiaNebraskaIdahoMontanaNorth DakotaAlaskaSouth DakotaWyoming
Polarisation = 0.5 MassachusettsCaliforniaNew YorkFloridaTexasGeorgiaIllinoisPennsylvaniaNorth CarolinaVirginiaNew JerseyOhioMarylandTennesseeMichiganColoradoWashingtonArizonaSouth CarolinaIndianaMissouriConnecticutAlabamaLouisianaNew HampshireWisconsinMinnesotaKentuckyOregonMississippiOklahomaDistrict of ColumbiaNevadaArkansasKansasIowaUtahRhode IslandMaineWest VirginiaNew MexicoNebraskaHawaiiIdahoVermontDelawareMontanaNorth DakotaSouth DakotaAlaskaWyoming
Polarisation = 0.0
Figure S2.
Size of mobilisation in each state for different levels of polarisation ( α = . , . , .
50 100
New YorkNew HampshireConnecticutDistrict of ColumbiaCaliforniaRhode IslandNew JerseyMarylandMaineVermontVirginiaGeorgiaHawaiiDelawareColoradoNevadaArizonaUtahAlaskaWyomingIllinoisNorth DakotaNew MexicoSouth DakotaMontanaSouth CarolinaWashingtonOregonIdahoWest VirginiaMississippiNebraskaNorth CarolinaKansasTennesseeArkansasPennsylvaniaIowaOklahomaLouisianaAlabamaIndianaKentuckyFloridaMissouriMinnesotaWisconsinMichiganOhioTexas S t a t e Polarisation = 1.0 F r a c t i on o f de m o c r a t s Residual
New YorkNew HampshireConnecticutDistrict of ColumbiaNew JerseyRhode IslandMarylandGeorgiaVirginiaCaliforniaMaineVermontArizonaColoradoNevadaHawaiiUtahDelawareSouth CarolinaTennesseeNorth CarolinaOregonAlaskaNorth DakotaWyomingMississippiKansasWest VirginiaIdahoSouth DakotaNew MexicoMontanaArkansasWashingtonNebraskaIllinoisOklahomaLouisianaFloridaAlabamaPennsylvaniaIowaKentuckyIndianaMissouriMinnesotaWisconsinTexasOhioMichigan
Polarisation = 0.5
100 0 100 200
New YorkGeorgiaNew HampshireDistrict of ColumbiaMarylandTennesseeConnecticutNew JerseyVirginiaArizonaRhode IslandSouth CarolinaColoradoNorth CarolinaMississippiNevadaUtahVermontHawaiiIllinoisMaineAlabamaArkansasDelawareKansasWest VirginiaOklahomaCaliforniaNorth DakotaIdahoAlaskaLouisianaWyomingNew MexicoSouth DakotaOregonWashingtonMontanaNebraskaFloridaKentuckyIndianaTexasMissouriIowaPennsylvaniaMinnesotaWisconsinOhioMichigan
Polarisation = 0.0
Figure S3.
Size of mobilisation without population trends for different levels of polarisation ( α = . , . , .0 (left to right).The result is the average of 500 simulations for 1,000 Democratic seeds in Middlesex County, MA. (blue for Democratic andred for Republican).