Automated Oracles

Prepa­ra­tion is the buzz­word for effi­cient action. If you know about a traf­fic jam, you can drive around it, when a drought is immi­nent, you can save water. The corona­virus revealed a world unpre­pared. By mid-2020, more than half a mil­lion peo­ple had died from Covid-19 infec­tion. And the ques­tion aris­es: Why is there no early warn­ing sys­tem for pan­demics?

Early warn­ing sys­tems are already stan­dard in a num­ber of other con­texts. On Decem­ber 26, 2004, the ocean floor off the coast of north­ern Suma­tra shift­ed. This under­wa­ter earth­quake was one of the strongest in a cen­tu­ry, mea­sur­ing 9.3 on the Richter scale. Its tremors caused waves to surge and slam ever high­er against the island and many other coast­lines of the Indi­an Ocean. Nobody was pre­pared, and more than 25,000 peo­ple died. A con­sor­tium led by the Ger­man Research Cen­tre for Geo­sciences (GFZ) in Pots­dam then set up an early warn­ing sys­tem in the Indi­an Ocean. A net­work of seis­mome­ters detects earth­quake epi­cen­ters, and satel­lites mea­sure move­ments of the earth’s sur­face via GPS. If the sys­tem were to respond to every vibra­tion, there would be many false alarms. GPS buoys and pres­sure sen­sors on the ocean floor help mea­sure each wave fol­low­ing a quake. The data are trans­mit­ted to a pro­cess­ing cen­ter and com­pared with pre­vi­ous records. This wealth of data yields mod­els that can inform oper­a­tors with­in min­utes of the speed, direc­tion, and mag­ni­tude of poten­tial tsunamis—in other words, whether they are harm­less or haz­ardous.

Early warn­ing sys­tems can also help com­pa­nies. One exam­ple lets them pre­pare for dis­rup­tions in their sup­ply chains. Resilience360 Sup­ply Watch is the name of a sys­tem used by the Ger­man logis­tics com­pa­ny DHL. The pro­gram defines around 140 risk cat­e­gories, includ­ing finan­cial, envi­ron­men­tal, and social fac­tors. Are media out­lets report­ing crim­i­nal activ­i­ty in a cer­tain region? What is the inci­dence rate of qual­i­ty com­plaints? How do inven­to­ry lev­els look—do ware­hous­es have suf­fi­cient stock? The DHL sys­tem ana­lyzes data from up to 30 mil­lion online and social media posts and makes the results of its risk analy­ses avail­able to clients. The pro­gram itself con­tin­u­ous­ly eval­u­ates the rel­e­vance and poten­tial con­se­quences of its data. It can do this with the help of machine learn­ing.

Machine learn­ing essen­tial­ly means that a sys­tem auto­mat­i­cal­ly com­piles new infor­ma­tion and eval­u­ates it in rela­tion to what it already knows. For exam­ple, a com­muter might nor­mal­ly need twen­ty min­utes to drive from A to B. But the route now has a con­struc­tion site. Because other dri­vers have already had to wait, an app can inform the com­muter before leav­ing home that today’s drive will take an addi­tion­al thir­ty min­utes. The sys­tem ana­lyzes thou­sands of con­struc­tion sites in this man­ner. Based on the over­all pool of data it col­lects, at some point it might deter­mine that when­ev­er an acci­dent blocks one lane of a par­tic­u­lar route, the drive will be delayed by approx­i­mate­ly forty min­utes. Then the app can report this dis­rup­tion imme­di­ate­ly, with­out wait­ing to ana­lyze the results from the first dri­vers to com­plete the route.

Early warn­ing sys­tems often draw on mil­lions or even bil­lions of data sets. Depend­ing on the field in ques­tion, that might be any­thing: from his­tor­i­cal records to tree rings to Twit­ter posts. Machine learn­ing helps orga­nize these files. It mod­els the inter­play of dif­fer­ent fac­tors.

Dirk Hel­bing, a pro­fes­sor of com­pu­ta­tion­al social sci­ence at ETH Zurich, stud­ies how ana­lyz­ing large vol­umes of data can help us do things like improve the flow of traf­fic. The goal is for sys­tems to be sta­ble. “If we make per­fect use of a road’s capac­i­ty, or have the max­i­mum num­ber of cars com­plete a route with­in an hour, that is an unsta­ble sys­tem,” he says. “The flow will break down. The road’s capac­i­ty will drop, and the result will be con­ges­tion.” It is bet­ter for the road to be used at a con­stant rate, but not at its max­i­mum capac­i­ty. This approach makes the sys­tem less sus­cep­ti­ble to dis­rup­tion. “Every access road or exit, every lane change, every accel­er­a­tion or brake action—all of these are poten­tial dis­rup­tions.” Machine learn­ing can be used to ana­lyze these dis­rup­tive fac­tors and derive a col­lec­tive pic­ture of dri­ving behav­ior. “It won’t do you any good to fur­nish just one car with a pro­gram,” says Hel­bing. Effi­cien­cy comes from a swarm structure—namely, when many cardrivers make their dri­ving pat­terns avail­able for data analy­sis. “Ana­lyz­ing about 40 per­cent of the cars could sig­nif­i­cant­ly improve traf­fic flow.”

Data sci­en­tist Dr. Avaré Stew­art also used swarm intel­li­gence when she led the M‑Eco project at Leib­niz Uni­ver­si­ty Han­nover. Her team want­ed to ana­lyze data from social media to devel­op an early warn­ing sys­tem for pan­demics. “Peo­ple often post about things like fever symp­toms or vis­its to the doc­tor,” says Stew­art. “A sin­gle post is of no use for a pro­jec­tion. But if fever posts clus­ter in a cer­tain loca­tion, that can be an indi­ca­tion.”

The advan­tage of social media lies in the imme­di­ate nature of the data, which can help flag an out­break early on. But it is dif­fi­cult to estab­lish a solid basis for these data. “Peo­ple post about their dog, their niece, or their next vaca­tion. Things that don’t nec­es­sar­i­ly have any­thing to do with ill­ness. To increase the qual­i­ty of your data, you have to fil­ter it.” If the fil­ter is too fine, you will lose impor­tant infor­ma­tion. If it’s too coarse, the sys­tem is unus­able. “Find­ing the right bal­ance here is the biggest chal­lenge,” says Stew­art.

Anoth­er prob­lem lies in deter­min­ing how early a warn­ing should be. “We’re com­par­ing the present with knowl­edge about the past,” says Stew­art. “So we can only issue a warn­ing when some­thing has already hap­pened.” In the case of a pan­dem­ic, that means when peo­ple are already infect­ed. “For us, it was too early to pre­dict poten­tial pan­demics sole­ly on the basis of our data.” That was a fac­tor in dis­con­tin­u­ing the project when its EU fund­ing ran out.

Germany’s corona­virus alert app is an early warn­ing sys­tem as well. It doesn’t actu­al­ly pre­vent infec­tions, but does help to slow their spread. Now that a num­ber of prob­lems in con­nec­tion with this com­plex project have been solved, the app is more reli­able at gen­er­at­ing new numer­i­cal codes every few min­utes, which it sends to all near­by smart­phones via Blue­tooth. Unlike such apps in other coun­tries, no cen­tral­ized com­put­er knows which phones have been in con­tact with other devices. The codes are only saved on the smartphones—which can­not iden­ti­fy any other users. Google and Apple, the two largest providers of smart­phone oper­at­ing sys­tems, have access to all the codes and can link them to the respec­tive devices. If a user reg­is­ters a pos­i­tive test result, this infor­ma­tion is sent to the two IT giants, which then send alerts to all smart­phones whose users have con­tact­ed or been in close spa­tial prox­im­i­ty to the infect­ed indi­vid­ual.

Because Ger­many does not col­lect data in a cen­tral­ized loca­tion or track GPS sig­nals and because it guar­an­tees anonymi­ty, it need­ed more time to devel­op its app than other coun­tries such as China, whose tech giant Aliba­ba took only three days. How­ev­er, it is unclear how well the app worked in China dur­ing the first few weeks. An inves­ti­ga­tion by the Berlin-based Mer­ca­tor Insti­tute for China Stud­ies reports more than 50,000 com­plaints of tech­ni­cal glitch­es in the first week of app usage alone.

Exten­sive debate about data pri­va­cy before the app was intro­duced in Ger­many seems to have had a favor­able effect on its accep­tance in the coun­try. The num­ber of users is a cru­cial fac­tor in how well this sys­tem devel­oped by SAP and Telekom can func­tion. In the week fol­low­ing its release, the corona­virus alert app was down­loaded by around 15 per­cent of the Ger­man pop­u­la­tion. In Aus­tria, a sim­i­lar app called Stopp Coro­na was avail­able a month ear­li­er, but was only being used by around 8 per­cent of the pop­u­la­tion three months after its release. Aus­tria has since improved its warn­ing app and guar­an­teed the same level of data pri­va­cy as in Ger­many.

In Ger­many, every infec­tion report­ed on the app is ver­i­fied. Peo­ple who reg­is­ter a pos­i­tive test result have to pro­vide con­fir­ma­tion. This elim­i­nates the pos­si­bil­i­ty of false alarms. How­ev­er, it also takes more time for alerts to be sent. And with every minute that goes by, the risk of infect­ing more peo­ple increas­es. What is more impor­tant, safe­ty or speed? This dilem­ma keeps appear­ing in con­junc­tion with early warn­ing sys­tems. And even the best sys­tems can­not pro­vide cer­tain­ty about the future, but instead only show options. By lay­ing the foun­da­tions for knowl­edge early on, they buy time. That allows peo­ple to pre­pare, and to take pre­cau­tions suf­fi­cient­ly early to influ­ence the future. The corona­virus app can give warn­ings. But the users them­selves then need to decide to limit their con­tact with oth­ers.