cscibox: an intelligent assistant for dang ice and sediment cores …lizb/cscience/agu2015.pdf ·...
TRANSCRIPT
AI-geoscien+stinterac+on• Thescien+stiskept“intheloop”ateachdecision
point,andcanguidetheprocessassheseesfit.Ateachstep,shecancomparetheresultstoherintui+onortootheragemodels.
• Hobbesexploresalterna+ves,performscomputa+ons,andtracksuncertainty.Itexploresthespaceofpossiblemodelsthoroughlyandquickly,andalwaysexplainsitsreasoningtothescien+st,includingstepsthatareconflic+ng,uncertain,orotherwiseproblema+c.
Hobbes’sKnowledgeBaseHobbesisaimingtoencodethevastknowledgegeoscien+stshavebuiltupthroughyearsofexperienceda+ngcores.Hobbes’scollec+onofrulesspringsoutoflong,detaileddiscussionsbetweentheAIscien+stsandthegeoscien+stsontheteam.Theseconversa+onsfocuson1. Themethodsofreasoninggeoscien+stsusetocreate&
evaluateagemodels2. Specificrulesofinferencegeoscien+stsuseinpar+cular
situa+onswhilecrea+nganagemodelExamplesofrulesobtainedfromtheseconversa+onsinclude:
• Themoremeasurementswehaveofacore,themorecertainwe’llbeabouttheresul+ngagemodel.
• Stra+Counterworkswellforlayercoun+ngifneighboringlayershaveroughlythesamethickness.
• Ifasedimentlayercontainsfossilsofterrestriallife,itmusthavebeendepositedonland.
• Ifthesedimenta+onratechangesabruptly,theagemodelisprobablywrong.
TheAIscien+ststhenencodeeachruleofinferenceintoanalgorithmHobbescanunderstand.
EvidenceFORsnowmeltTemperaturefrequentlyabovefreezing(strongerevidence)
EvidenceAGAINSTsnowmeltAveragetemperaturebelowfreezing(weakerevidence)
Example:Da+ngNEEM-S1icecoreStep1:UseStra:Counter[Winstrup15]forautomatedlayercoun:ng
Hobbes:Ineedaparameter\mu,fortheaveragelayerthickness.Iguessed\mu=7cm,basedonasimilarcore.Dotheselayerboundarieslookcorrect?
Geocien+st:No,therearetoomanylayerboundaries.Use\mu=13cm.Hobbes:Okay,using\mu=13cm.Dotheselayerboundarieslookcorrectnow?
Geocien+st:Yes,con+nuelayercoun+ng.
Step2:AlignthegasrecordandtheicerecordHobbes:Weneedafirndensifica+onmodeltoes+mateagedifferencebetweengasbubblesandsurroundingice.ThestandardmodelisbyHerron&Langway,butitassumestherewasnosnowmelt.Iwillevaluatewhetherthismodelisappropriate.
ArgumentFORsnowmelt(onlymoderatelycertain)
Hobbes:Judgingbythecurrenttemperaturevaluesattheloca+onoftheNEEM-S1core(average-3.34Candstandarddevia+on22.97),theHerron&Langwaymodelmaynotbeappropriateforthiscore.ShouldIusethe<nameothermodelhere>instead?
Step3:Stoplayercoun:ngaDerannualsignalsfadeHobbes:Weareatdepth3525m.IwillcheckwhethertheannualsignalshavefadedtoseeifIshouldkeeplayercoun+ng.
Argument:numberofpeaksperseriesisnormalpartlyapplicablesound
Argument:averagelayerthicknessisnormalmostlyapplicableplausible
Conclusion:con+nuelayercoun+ngmostlyapplicablesound
Hobbes:Wecancon+nuelayercoun+ngajerdepth3525m.
Finaloutput:Reasoningbehindagemodels:• Performedautomatedlayercoun+ng
betweendepth50mand4020m<thesearemadeup!>
• UsedDansgaard-Johnsenflowmodelforda+ngdepthsbelow4020mIdon’tseethatabove.Ithinkyou’remissinga4thstep
• Usedtwodifferentfirndensifica+onmodels:Herron&Langway(orangecurve)and?(bluecurve)
• Whataboutthegasrecordalignment?• Etc(moredetailswouldbeincludedinthe
fullreport)
ReproducibilityReproducibilityisakeycomponentofthescien+ficmethod.Ascomputa:onbecomesmorecentraltothescien+ficenterprise,itisurgenttoaddressconcernsregardingreproducibilityofcomputa:onalresults.CSciBoxaddressesreproducibilityconcernsby:• Tracksallac+onstakenwiththecore,forcomplete
reproducibility• Includescita+onnotesevery+meyouuseanother
scien+st’stool• Metadataisalwaysbundledwiththecoredata• WeuseLinkedPaleoclimateDateformat• Itiseasy(andmandatory)inCSciBoxtosaveallthe
dataandmetadatathatproducedanagemodel(ifyouwanttosavetheagemodel).Bymetadata,wemeanspecificparameterchoicesandsuch-everythingthat’srequiredtoreplicatethecomputa+onexactly.
• Itisalsoeasy(butop+onal)tosaveinforma+onabouttheothermodelsthatthescien:stconsideredbutrejected.<Somethingaboutwhythisisimportantforsciencetoo>
• Hobbes’sknowledgebaseencodesthereasoningbehindgeoscienceresearch.Thisreasoningmaynotalwaysbereportedintradi+onalscien+ficjournaling(?)Theprocessof“algorithmizing”therulesofinferenceforcesscien+ststocodifyandreportalloftheirreasoning.That’scool!
• CSciBoxspecificallyaddressesmanyoftherecommenda+onsoftheYaleLawSchoolRoundtableonDataandCodeSharing(cite).
Argumenta+vereasoningHobbesconsidersargumentsforandagainsteachtheory,andweighstheirstrength.• Thismimicsthewayscien:ststhink,andthewaythey
communicateaboutitwithothers—andthusstreamlinescommunica+onbetweenHobbesanditsusers
• Italsoallowsthescien+sttoconsidermul:pleagemodelsatonce,andenablesHobbestohandleconflic+ngarguments.
• Atthispointintheprocess,forinstance,thescien+stcanchooseoneortheotherofthesuggestedfirnmodels—ormaychoosetodirectHobbestoexploreboth.
• Thisistheplacetoputthatmaterialaboutscien+ficdisagreementandconflic+ngrules
Twocomponentsofconfidenceinthereasoning• Confidenceinthescien:ficknowledge(validity):Isthisruleofinference
valid?Validityrangesfromauniversallyacceptedrule(accepted),toagut-levelfeeling(plausible).
Accepted>Sound>Probable>PlausibleAvaliditylevelisaqachedtoeachrule,andisdeterminedbytheexpertgeoscien:stsontheteam.Thediscretelevelsofvaliditycorrespondtotheweightsscien+stsgivetotheirreasoning,andareeasiertoobtainfromexpertsthancon+nuousvalueswouldbe.
• Confidenceintheevidence(applicability):Howwelldoesoursitua+onmatchtheruleofinference,asshownbyourdataset?
Highlyapplicable>mostlyapplicable>partlyapplicableHobbesdeterminestheapplicabilityofapieceofevidence,byanalyzingthedata.• Whenweighingtheevidencefororagainstatheory,Hobbesisdesignedto
intelligentlyconsiderthevalidityandapplicabilityofeachargument..Hobbescombinestheindividualargumentsintoasingleconclusion,describinghislevelofbeliefinthetheory.
I’mstrugglingabitwithwri+ngaboutthetwocomponentsofconfidence.Theconceptisrathersubtle,andhardtoexplainwellinashortblurbonaposter.I’dliketotalkaboutit,butIdon’twanttomen+onitandnotexplainitwell.S+llworkingonthat…Alsonotethatthearrowshouldpointtothesec+on“stoplayercoun+ngajerannualsignalsfade”,butIcan’teasilymakepowerpointdothatrightnow.
Hobbes:Okay[saves\mu=13cminthemetadataandproceeds]
PP11A-2203
Elizabeth Bradley1, Kathleen Finlinson2, Tyler Jones3, Colin Lindsey3, Brett Israelsen1
Kenneth A. Anderson1, Thomas M. Marchitto3, Laura Rassbach de Vesine1, and James W. C. White3, 1 Department of Computer Science, University of Colorado, Boulder, Colorado, USA
2 Department of Applied Mathematics, University of Colorado, Boulder, Colorado, USA 3 Institute for Alpine and Arctic Research (INSTAAR), University of Colorado, Boulder, Colorado, USA
CSciBox:Anintelligentassistantforda+ngiceandsedimentcores
Themainidea:
CSciBox!
Geoscience
Computerscience
Image&courtesy&of&Dorthe&Dahl2Jensen,&via&Jim&White&Photo:DortheDahl-Jensen
Geoscien+st:• Hascore• Wantsagemodel• Bringsscien+ficexper+se
CSciBox
Output:• Oneormoreproposedagemodelsforthecore• Sophis+cateduncertaintyes+mates• Summaryofthereasoningbehindeachmodel
Hobbes
CSciBox:• Helpsapplyscien+ficreasoningandalgorithms• Exploresalterna+vesautoma+cally• Keepstrackofallthereasoninginvolvedinthemodel
crea+on• Needtoexplaintherela+onshipbetweenHobbes
andCSciBox
Thisisold;willneeda
newscreenshot
ThismaterialisbaseduponworksponsoredbytheNa+onalScienceFounda+on.Anyopinions,findings,andconclusionsorrecommenda+onsexpressedinthismaterialarethoseoftheauthor(s)anddonotnecessarilyreflecttheviewsoftheNSF.
www.cs.colorado.edu/~lizb/cscience.html!
• Sourcecode(python)availableongithub:github.com/ldevesine/cscience• Butyoudon’thavetoknowpythontorunit;wealsohaveone-clickinstallerstheretoo• GNUpubliclicense;freetomodify/extend/useasyouseefit
Wearetryingtobuild—andsupport—ausercommunity.Pleasejoinus!
CSciBoxisopensource
Intheworks:• Saywhatthefutureworklandscapelookslike• Herearesomesugges+ons:
• MakingStra+Counter&Baconbothwork• HelpingHobbesmakesmartdecisionsaboutwhentorunexpensivecomputa+ons• AddingextensivelytoHobbes’rulebase
Winstrup15cita+on