




版權說明:本文檔由用戶提供并上傳,收益歸屬內容提供方,若內容存在侵權,請進行舉報或認領
文檔簡介
生成式AI和仿真建模使用ChatGPT等大型語言模型FASTGenerativeAIandsimulationmodeling:howshouldyou(not)uselargelanguagemodelslike
生成式AI和仿真建模:使用ChatGPT等大型語言模型growthintheUnitedStates.However,ourguidelinesaregenericandcanbeappliedtoabroaderCopyright?2024TheAuthors.SystemDynamicsReviewpublishedbyJohnWiley&SonsLtdonbehalfofSystemDynamicsSociety.Syst.Dyn.Rev.AdditionalSupportingInformationmaybefoundonlineinthesupportinginformationtabforthisGenerativeAIholdsimmensepotentialforresearchwhileposingchallengesplyprovide“endlessentertainment”(Thorp,2023,p.313).IntheearlydaysofChatGPT,itgeneratedandreferencedscientificstudiesthatdidnotexisteralmajorjournals,includingbutnotlimitedtoNatureandScience,releasedguidanceontheresponsibleuseofgenerativeAI,andNatureaMGHInstituteforTechnologyAssessment,HarvardMedicalSchool,Boston,Massachusetts,bSloanSchoolofManagement,MassachusettsInstituteofTechnology,Cambridge,Massachusetts,*Correspondenceto:MohammadS.Jalali,MGHInstituteforTechnologyAssessment,125NashuaSt.,Boston,MA\hAcceptedbyAndreasGr??ler,Received21February2024;Revised2April2024;Accepted23AprilSystemDynamicsSystemDynamicsPublishedonlineinWileyOnlineLibrary()DOI:10.1002/sdr.1773ThisisanopenaccessarticleunderthetermsoftheCreativeCommonsAttribution-NonCommercialLicense,whichpermitsuse,
生成式人工智能(AI)工具,如大型語言模型(LLM)和ChatGPTAI?領域的潛力以及缺乏其有效部署指南方面,仍然存在差距。本文旨在彌合這些差距。我們通過一個建模COVID?19對美國經濟增長影響的例子來討論ChatGPT的應用。然而,我們的指南是通用的,可以應用于更廣泛的生成式AI工具。我們的工作提出了一種系統(tǒng)的方法,將生成式AI整合到仿真研究連續(xù)體中,從問題闡述到見?2024·SonsLtdSyst.Dyn.Rev.生成式人工智能(AI)和大型語言模型(LLM),例如ChatGPT,已經得到了研審查(例如,Ariyaratne等人,2023;Castellanos?Gomez,2023),但它們對實際上很少產生研究產出。他們指出,像ChatGPT這樣的工具會產生“看似可信但實際上錯誤”的回應(Shen等人,2023)以及僅僅提供“無盡的娛樂”(Thorp,2023,第313頁)。在ChatGPT的早期階段,它生成并引用了不存在的科學文獻(Thorp,2023),其他平臺也存在類似問題。Meta的Galactica產生了不準確和具有種族歧視的內容(Stokel?Walker和VanNoorden,2023)。在我們的經驗中,我們觀察到Google的Bard在聲稱從GoogleScholar發(fā)布了關于生成式人工智能負責任使用的指南,以及《自然》(NatureaMGHInstituteforTechnologyAssessment,blhlfhfhlbdh)。通訊作者:MohammadS.Jalali,MGHInstituteforTechnologyAssessment,125NashuaSt.,Boston,MA02114USA\hmsjalali@AcceptedbyAndreasGr??ler,Received21February2024;Revised2April2024;Accepted23April于WileyOnlineLibrary(wileyonlinelibrary.comDOI:10.1002/sdr.1773
2Editorial,2023),amongseveralothers,prohibitedtheinclusionofgenerativeintegrityofthework(Flanaginetal.,2023;Thorp,2023).Ontheotherhand,somearguethatgenerativeAIcancontributetotheresearchprocess.TheysuggestthatgenerativeAIoffersgreatpotentialtoimproveaca-demicworkefficiency(Meyeretal.,2023),andtoolslikechatbotsenabletheexplorationof“unchartedscientificterritories”(ThapaandAdhikari,2023,tionmodeling.Inthisarticle,weaimtoclarifyanddemonstratehowtousegenerativeAIBackgroundandofLLMs.Thesemodelsaretrainedonvastamountsoftextdatatopredictthenextwordinasequence,andtheycanbeusedtogeneratecoherentandcontextu-allyrelevantcontent.WhileLLMsareasubsetofgenerativeAIthatfocusesonorotherformsofcontent.DOI:
作品的原創(chuàng)性和完整性負責(Flanagin等人,2023;Thorp,2023)。智能具有提高學術工作效率的巨大潛力(Meyer等人,2023),而像聊天機?人這fi(ThapaAdhikari,20232647Ghaffarzadegan等人,2024)。例如,ChatGPT可以生成研究想法、文獻綜述和數(shù)據(jù)摘要(Dowling和Lucey,2023)。這些工具有助于研究人員修改和審閱稿件以及開發(fā)或檢查代碼(Stokel?Walker和VanNoorden,2023)。也有人提出,它關于生成式人工智能應用的價值及其對研究過程直接(例如,一手研究過程)本文旨在闡明并演示如何在仿真建模研究中使用生成式AI。我們還討論了如何利用生成式AI將口頭和書面提示轉化為輔助研究過程的結果。我們以COVID?19爆發(fā)及其對美國GDP的影響的簡化案例進行研究。特別是,我們考慮了COVID?19作為一個外生因素影響簡化的消費乘數(shù)模型(薩繆爾森,1939;洛,1980)。最后,該案例提供了一個現(xiàn)實的環(huán)境,可以在實際中測試生成式AI的應用。在這篇文章中,我們展示并強調,生成式AI不應取代思考;相反,它是一種有大型語言模型(LLMs)是經過深度學習訓練以理解和生成人類語言的模型。ChatGPT和Bard,它們使用生成式Transformer架構,是這類模型的例子,屬于更廣泛的LLMs家族。這些模型在大量文本數(shù)據(jù)上訓練,以預測序列中的下一個單詞,并且可以用來生成連貫且與上下文相關的文本內容。雖然LLMs是關注文本的生成式AI的一個子集,但生成式AI還可以包括生成圖像、音樂、視頻或其他形式內容的最近,不同的AI平臺開始提供組合功能。谷歌的Bard、微軟的Bing和OpenAI的ChatGPT等平臺,被稱為對話式AI模型,它們提供LLMs和其他生成式AI工具(如圖像處理和高級數(shù)據(jù)分析)的混合,使它們成為研究過程中的有用10991727,0,Downloaded10991727,0,Downloadedfrom/doi/10.1002/sdr.1773byMassachusettsGeneralHospital,WileyOnlineLibraryon[22/07/2024].SeetheTermsandConditions(/terms-and-conditions)onWileyOnlineLibraryforrulesofuse;OAarticlesaregovernedbytheapplicableCreativeCommonsesneciLsnommoCevitaerCelbacilppaehtybdenrevogeraselcitraAO;esufoselurrofyrarbiLenilnOyeliWno)snoitidnoc-dna-smret/moc.yeliw.yrarbilenilno//:sptth(snoitidnoCdnasmreTehteeS.]4202/70/22[noyrarbiLenilnOyeliWlatipsoHlareneGsttesuhcassaMyb3771.rds/2001.01/iod/moc.yeliw.yrarbilenilno//:sptthmorfdedaolnwoD,0
A.Akhavan和M.S.Jalali:生成式AI和仿真建模writingthisarticle.Therefore,wefocusontheapplicationsofGPT-4whileacknowledgingthatmoreevolvedtechnologiesmaybeintroducedinthenearDavisetal.,2007;Lawetal.,2007;Sterman,2000)haveproposedstepsofthecesshasbeenwidelyemphasizedintheliterature(e.g.Harrisonetal.,2007;andmaintenancetoenhancetransparencyandreplicabilityofthesimulationresearch,anareathatstillhassignificantroomforimprovement(Jalalietal.,2021;JalaliandBeaulieu,2023;Monksetal.,2019;RahmandadandSterman,2012).study.Webuildupontheframeworksonhowtoconductsimulationmodelingresearchanddeveloptheoriesfromcasestudies(Davisetal.,2007;Sterman,2000).Weconsideredgeneralstepsthatareapplicabletomostsimulationmodelingcanhelpshapeandrefinetheprocess.Itisimportanttonotethatsimulation(i.e.heterogeneityinhowsentencesareformed),theyshouldreceiveclearinstructions(prompts)fromusers.However,theabsenceofestablishedguidelinessuchcases,werevisedpromptstoachievesatisfactoryresponses,knownaslikeChatGPT.ChatGPT’sresponsesareoftenlong;therefore,tokeepthisarticleconcise,wecreationandenhancementofprompts,forexample,iterativeprocesstorefineUsinggenerativeAIinsimulationmodeling—caseDOI:
截至本文撰寫之時,GPT?4(最新版本)提供了最強大的功能。因此,我們專注于GPT?4的應用,同時承認在不久的將來可能會出現(xiàn)更先進的技術。者(例如,Banks,2005Davis2007Law2007Sterman,2000)提出了模擬研究中的建模步驟,這一過程的迭代性質在文獻中得到了廣泛強調(例如,Harrison等人,2007;Homer,1996)。我們遵循同樣的方法,將自己置于進行模擬研究的學者的位置,并從ChatGPT這一領域仍有很大的改進空間(Jalali等人,2021;Jalali和Beaulieu,2023;Monks等人,2019;Rahmandad和Sterman,2012)。行模擬建模研究和從案例研究中發(fā)展理論(Davis等人,2007;Sterman,2000)方法,并展示了ChatGPT如何幫助塑造和改進這一過程。需要注意的是,模擬建模此外,由于生成式AI工具,如ChatGPT,依賴于隨機模型(即句子形成的異質是提供精心制作的提示示例,這些示例說明了如何與生成式AI進行有效溝通。這些示例為用戶提供了實際指南,幫助他們了解如何制定能夠從ChatGPT等工具中獲得以下,我們首先展示這些提示和ChatGPT的相應響應。ChatGPT的響應通常AI在此,我們通過一種交互式對話方法展示了生成式AI在仿真研究中的應用。我們的10991727,0,Downloaded10991727,0,Downloadedfrom/doi/10.1002/sdr.1773byMassachusettsGeneralHospital,WileyOnlineLibraryon[22/07/2024].SeetheTermsandConditions(/terms-and-conditions)onWileyOnlineLibraryforrulesofuse;OAarticlesaregovernedbytheapplicableCreativeCommonsesneciLsnommoCevitaerCelbacilppaehtybdenrevogeraselcitraAO;esufoselurrofyrarbiLenilnOyeliWno)snoitidnoc-dna-smret/moc.yeliw.yrarbilenilno//:sptth(snoitidnoCdnasmreTehteeS.]4202/70/22[noyrarbiLenilnOyeliWlatipsoHlareneGsttesuhcassaMyb3771.rds/2001.01/iod/moc.yeliw.yrarbilenilno//:sptthmorfdedaolnwoD,0
4intheeducationalprocessofrefiningthevariousstepsinvolvedinmodeling.
對于剛開始在仿真建模領域探索的學生或初級研究人員來說,這種方法旨在展示這些工具的利用,并協(xié)助他們在建模過程中各個步驟的教育過程。雖然我們故意引入了一些錯誤([]標記這些故意錯誤以避免混淆),我們的方法不僅限于錯誤糾正。通過這種基于對話的方法,我們旨在為學習者提供見AIProblemdefinitionandresearchGenerativeAIcanhelprefineresearchquestionsandexploreproblemback-withresearchboundaryconsiderationsbysuggestingaspectsresearchersmaynotResearchquestionandIamstartingsimulationmodelingresearchandwanttocross-checkmyresearchquestion.Ihaveconsideredthisquestion:HowhaseconomicgrowthimpactedCOVID-19?Youshouldbecritical,findpotential?awsinmyresearchquestion,andsuggestimprovements.[WeintentionallyreversedthecausalityandResponseChatGPTprovidedfeedbackonresearchquestionclarityanddirectionality,theresearchquestion’sclarityanddirectionality,itnotedthatour“researchques-economicgrowthhasbeenaffectedbythepandemicorhoweconomicgrowthonyourintention,youmightneedtorewordforclarity.”ityandcorrelation.Itsuggestedthat“Sayingeconomicgrowth‘impacted’COVID-19assumesacausalrelationship.However,inreality,manyvariablesringtothegrowththatoccurredbeforethepandemic,duringit,orafter?Thisneedstobedefined.”DOI:
Problemdefinitionandresearch生成式AI可以幫助細化研究問題和探索問題背景。它可以理解研究問題的廣度和范圍,并分析研究問題變量之間的語義和因果關系。連接到互聯(lián)網的生成式AI可以搜COVID?19在缺陷,并提出改進建議。我們故意顛倒了因果關系,并考慮了廣泛的術語而不是Responsesummar(輸入內容未提供足夠信息,無法進行翻譯。ChatGPT了反饋。關于研究問題的清晰度和方向性,它指出我們的研究問題似乎暗示經濟關于因果關系,ChatGPT“‘’COVID?19ChatGPT還是大流行后的增長?這需要定義。”GDP10991727,0,Downloaded10991727,0,Downloadedfrom/doi/10.1002/sdr.1773byMassachusettsGeneralHospital,WileyOnlineLibraryon[22/07/2024].SeetheTermsandConditions(/terms-and-conditions)onWileyOnlineLibraryforrulesofuse;OAarticlesaregovernedbytheapplicableCreativeCommonsesneciLsnommoCevitaerCelbacilppaehtybdenrevogeraselcitraAO;esufoselurrofyrarbiLenilnOyeliWno)snoitidnoc-dna-smret/moc.yeliw.yrarbilenilno//:sptth(snoitidnoCdnasmreTehteeS.]4202/70/22[noyrarbiLenilnOyeliWlatipsoHlareneGsttesuhcassaMyb3771.rds/2001.01/iod/moc.yeliw.yrarbilenilno//:sptthmorfdedaolnwoD,0
A.Akhavan和M.S.Jalali:生成式AI和仿真建模RevisedresearchThesearegoodpoints.IseethatImadeamistakeinthedirectionofcausalityandthatIwasnotspecificenough.ItshouldhavebeentheimpactofCOVID-19trendfrom2013to2023,andIwanttofocusontheU.S.economy.Withthoseaspects,myrevisedresearchquestionis:HowdidchangesinconsumerbehaviorduringtheCOVID-19pandemicimpactU.S.GDPgrowth?Howdoesthatsound?Responseestablishabaselinetrendbeforethepandemicandthenassessanydeviationsfrombackoncounterfactualscenarios,confoundingfactors,anddatasources.Forexam-Canyousearchforsimilarpeer-reviewedjournalarticlesusingsimulationresearch?CanyoualsointroduceafewotherreliablesourcessothatIcangetdataonGDPgrowthandrelevanteconomicmeasures?ResponserecentversionsofChatGPTproviderelevantliteraturewithlinkstopublisheddata.ChatGPTalsoprovidedabriefoverviewofthecontentofthestudies.Forexample,itsaidthatthegivenarticle“discussesamodelthatinterlinksthelikeCOVID-19,”andaboutthedatabaseitsaid,“St.LouisFederalReservereportedreferencethatwedonotreporthere.WhatChatGPTprovidesisnotexhaus-targetedliteraturesearchandcomplementmanualsearchesindatabases.DOI:
市場表現(xiàn)、就業(yè)創(chuàng)造,還是其他指標?您說的“影響”,是指對病例數(shù)、死亡率、COVID?19關注GDP增長。關于時間動態(tài),我正在考慮2013年至2023年的GDP趨勢,并想專注于美國經濟。有了這些方面,我修改后的研究問題是:COVID?19大流行期間消費者行為的變化是如何影響美國GDP增長的?這個想法怎么樣?ResponseChatGPT承認了我們對研究問題的修改,并提供了如下反饋通過指定2023COVID?19對此趨勢的偏離。此外,它還提供了關于反事實情景、混雜因素和數(shù)據(jù)來源的反饋。例如,它建議考慮在沒有大流行的情況下GDP增長會是什么樣子,或者像財政政策GDP來源,以便我獲取關于GDP增長和相關的經濟指標的數(shù)據(jù)嗎?Responsesummar(輸入內容未提供足夠信息,無法進行翻譯。與早期的聊天機?人不同,它們編造了不存在的引用,ChatGPT的最新版本提供了相關的文獻,并附有已發(fā)表材料的鏈接。針對我們的查詢,ChatGPT提供了一篇相關期刊文章(Xiang等人,2021)。此外,它還提到了美國經濟分析局和圣路易斯聯(lián)邦儲備銀行(FRED)等數(shù)據(jù)庫,這些數(shù)據(jù)庫報告了GDP增長數(shù)據(jù)。ChatGPT還簡要概述了研究的內容。例如,它說,所提供的文章“討論了一個將經濟中勞動力供應的動態(tài)與COVID?19等傳染病的傳播相互關聯(lián)的模型”以及關于數(shù)告。ChatGPT提供的內容并不全面;因此,可以更系統(tǒng)地執(zhí)行此類過程,以協(xié)助有10991727,0,Downloaded10991727,0,Downloadedfrom/doi/10.1002/sdr.1773byMassachusettsGeneralHospital,WileyOnlineLibraryon[22/07/2024].SeetheTermsandConditions(/terms-and-conditions)onWileyOnlineLibraryforrulesofuse;OAarticlesaregovernedbytheapplicableCreativeCommonsesneciLsnommoCevitaerCelbacilppaehtybdenrevogeraselcitraAO;esufoselurrofyrarbiLenilnOyeliWno)snoitidnoc-dna-smret/moc.yeliw.yrarbilenilno//:sptth(snoitidnoCdnasmreTehteeS.]4202/70/22[noyrarbiLenilnOyeliWlatipsoHlareneGsttesuhcassaMyb3771.rds/2001.01/iod/moc.yeliw.yrarbilenilno//:sptthmorfdedaolnwoD,0
6Fig.1.InitialmodelofprovidedinaChatGPT
1由ChatGPT提示提供 Modeltualmodelsandidentifytherelationshipsamongtheidentifiedvariables.Wethechoiceofvariablesandtheirnaming,directionofcausality,polarityoftheConceptualIwantyourfeedbackonmyattempttoidentifyvariablesanddevelopacon-ceptualmodelforincludingCOVID-19shockinasimplifiedmacroeconomicmodeloftheconsumptionmultiplier.Ihaveputtogetherthekeyrelationshipsformymodel.Canyouchecktheattachedphotoandgivemeyourfeedback?Becrit-Responsemadeintentionallywhileacknowledgingthecorrectaspects.Itrecognizedtheditureandaggregatedemandbyspecifying:“Thisisunusual.Generally,anincreaseingovernmentexpenditure,allelsebeingequal,wouldleadtoDOI:
AIAI能夠分析概念模型的復圖像(圖1);然而,圖像的內容也可以簡單地用文字來呈現(xiàn)。生成式AI可以提供的反饋。生成式AI還可以檢查模型變量的定義并建議新的變量。ConceptualCOVID?19附上的照片并給我反饋嗎?請?zhí)岢雠u意見。我們故意在概念模型中刪除了一些變ResponseChatGPT極性。然后,它識別了我們故意犯的錯誤,同時承認了正確的方面。它通過指出“隨著GDP的增長,我們預計人口的平均收入或預期收入將上升”,識別了GDP和預期收入之間錯誤的因果極性。它還通過指定這是不尋常的。通常情況下,在其”,識別了政府支出和總此外,ChatGPT還提出了額外的變量,建議了一些在GDP宏觀經濟模型中常見的因10991727,0,Downloaded10991727,0,Downloadedfrom/doi/10.1002/sdr.1773byMassachusettsGeneralHospital,WileyOnlineLibraryon[22/07/2024].SeetheTermsandConditions(/terms-and-conditions)onWileyOnlineLibraryforrulesofuse;OAarticlesaregovernedbytheapplicableCreativeCommonsesneciLsnommoCevitaerCelbacilppaehtybdenrevogeraselcitraAO;esufoselurrofyrarbiLenilnOyeliWno)snoitidnoc-dna-smret/moc.yeliw.yrarbilenilno//:sptth(snoitidnoCdnasmreTehteeS.]4202/70/22[noyrarbiLenilnOyeliWlatipsoHlareneGsttesuhcassaMyb3771.rds/2001.01/iod/moc.yeliw.yrarbilenilno//:sptthmorfdedaolnwoD,0
A.Akhavan和M.S.Jalali:生成式AI和仿真建模feedbackloop:“Inreality,anincreaseinaggregatedemandmightleadtoantiontrade-off,suchthatwhileitisimportanttokeepthemodelsimple,weshouldensurethatitstillcapturesthenecessarycomplexityofthereal-worldAsnotedearlier,ChatGPT’sresponsesareoftenlong.WeprovideditsfullIhaverevisedmyconceptualmodelandhaveattachedtheupdatedversion[-Figure2].Ifixedthepolarityerrors.IconsideredtheimpactofNetExportsonGDPinadditiontoAggregateDemand.IalsoaddedthecausallinkbetweenCon-sumptionandAggregateDemand.Forsimplicity,Ihavenotincludedparameterssuchasinitialvaluesandadjustmenttimes,andtheconceptualmodelonlycap-turesthebasicdynamics.IwanttomoveforwardwiththisversioniftherearenoResponserelationshipsonebyone,andprovidedfeedbackaboutthemodel’salignmentbeunderstandabletothosefamiliarwithbasicmacroeconomicprinciples.”Itnotedanimportantobservationthatthemodelmissestimelags:“Whilenotrepre-
加可能會導致GDP的增加,進而影響預期收入,從而影響消費,形成循環(huán)效應。”如前所述,ChatGPTS1我考慮了凈出口對GDP的影響,除了總需求之外。我還添加了消費和總需求之間的ResponseChatGPT與已建立的經濟理論的吻合度的反饋。它指出,“該模型以清晰的方式呈現(xiàn),具有易理解的。它還指出一個重要的觀察結果,即模型缺少時間滯后雖然簡化模型中沒有表示,但請注意這一點。”它還提醒我們,該模型沒有捕捉到所有潛在的因果關系,這與我們縮小這個簡單模型邊界的意圖相符。在這方面,它還指出,“雖然你的模型是簡化的,但請注意,在一個更動態(tài)的模型中,GDP的變化可以反饋到政府Fig.2.Revisedconceptualmodel
圖2.在ChatGPT提示DOI:
10991727,0,Downloaded10991727,0,Downloadedfrom/doi/10.1002/sdr.1773byMassachusettsGeneralHospital,WileyOnlineLibraryon[22/07/2024].SeetheTermsandConditions(/terms-and-conditions)onWileyOnlineLibraryforrulesofuse;OAarticlesaregovernedbytheapplicableCreativeCommonsesneciLsnommoCevitaerCelbacilppaehtybdenrevogeraselcitraAO;esufoselurrofyrarbiLenilnOyeliWno)snoitidnoc-dna-smret/moc.yeliw.yrarbilenilno//:sptth(snoitidnoCdnasmreTehteeS.]4202/70/22[noyrarbiLenilnOyeliWlatipsoHlareneGsttesuhcassaMyb3771.rds/2001.01/iod/moc.yeliw.yrarbilenilno//:sptthmorfdedaolnwoD,0
v18v2exports,creatingcomplexinteractions.”Thisisagreatpoint;however,wedidthreevariables,andassuch,theirhistoricaleffectsarealreadycaptured.IftGPTremindedusthatallmodelsaresimplificationsofreality,andweshouldditionsorpolicyregimes.
出口,創(chuàng)建復雜交互。這是一個很好的觀點;然而,鑒于我們使用外生數(shù)據(jù)來表示這超出了當前簡單模型的范圍。最后,ChatGPT提醒我們,所有模型都是現(xiàn)實的一Modelformulationandetersareinreasonableranges.Additionally,generativeAItoolsaregreatfortrans-latingwrittenmathematicalequationsandverbalcommunicationsintothecodes.Iwanttodevelopasimulationmodelbasedonmyconceptualmodelandobservethebehaviorofthevariablesovertimeatanaggregatelevel.BasedonwhatIhavelearnedinmyclassesandsomeinitialsearches,Ithinkacompart-mentalmodelingapproachwouldbeareasonablechoice.Whatdoyouthink?ResponseChatGPT’sresponseemphasizedthesuitabilityofusingacompartmentalmodelingapproach,typicallyseeninepidemiologyandsystemsdynamics,foricsofaggregatedemand,supply,income,andconsumption;thecapacitytounderstandthedynamicbehaviorofeconomicsystems,includingdelaysandfeedbackloops;itsutilityinsimulatingpolicyinterventionsandtheirimpactstations:thepotentialoversimplificationofcomplexeconomicinteractions;thereal-worldpolicy.ChatGPTconcludedthatwhileacompartmentalmodelisagoodstartingpointforobservingeconomicvariablebehaviorovertime,espe-ciallyinresponsetoshockslikeCOVID-19,amoreextensiveanalysismayDOI:
AIAI這些工具通過提供編程指南,可以促進代碼開發(fā)過程,并將代碼從一種語言翻譯成另AIResponsesummar(輸入內容未提供足夠信息,無法進行翻譯。ChatGPT的回答強調了使用分室建模方法(重要。然而,ChatGPT也指出了局限性:復雜經濟互動可能被過度簡化;準確參數(shù)挑戰(zhàn);以及將模擬結果謹慎地轉化為現(xiàn)實世界政策的需要。ChatGPT總結說,雖然COVID?1910991727,0,Downloaded10991727,0,Downloadedfrom/doi/10.1002/sdr.1773byMassachusettsGeneralHospital,WileyOnlineLibraryon[22/07/2024].SeetheTermsandConditions(/terms-and-conditions)onWileyOnlineLibraryforrulesofuse;OAarticlesaregovernedbytheapplicableCreativeCommonsesneciLsnommoCevitaerCelbacilppaehtybdenrevogeraselcitraAO;esufoselurrofyrarbiLenilnOyeliWno)snoitidnoc-dna-smret/moc.yeliw.yrarbilenilno//:sptth(snoitidnoCdnasmreTehteeS.]4202/70/22[noyrarbiLenilnOyeliWlatipsoHlareneGsttesuhcassaMyb3771.rds/2001.01/iod/moc.yeliw.yrarbilenilno//:sptthmorfdedaolnwoD,0
A.Akhavan和M.S.Jalali:生成式AI和仿真建模ModelformulationsandparameterThanks!Ifeelmoreconfidentaboutmydecisiontochooseacompartmentalmodelingapproach.Ihavedefinedtheequationsbelowandamtryingtouseordinarydifferentialequationstomodelthedynamicrelationshipsamongthevariables.Forthisdraft,Ihaveconsideredahypotheticalcaseandassignedarbi-trarynumberstoparameters.Becriticalandgivemefeedbackontheequationsandparametervalues.[Weintentionallyusedawrongequationfor“Aggregate Income=INTEGRAL inExpected ExpectedIncome).Consumption=MarginalPropensitytoConsume×ExpectedtionTime.×NetAdjustmentTime.GovernmentExpenditure=Investment=pensitytoConsume.Alpha=ResponseOverall,ChatGPTacknowledgedthatusingordinarydifferentialequations(ODEs)isagoodstartandremindedustoensurethatoursimulationenvironmentaggregatedemand,indicatingthat“theaggregatedemandistheproductofitsthatparametervaluesseemarbitraryandcommentedontheirintuitions.ForquicklyorslowlyagentsupdatetheirexpectationsinresponsetochangesinGDP.Italsorecognizedtheothererror,suggesting“notethattheMarginalPropen-haveMPC=80,whichsuggeststhatforeveryunitofexpectedDOI:
并為參數(shù)分配了任意數(shù)值。請批評并就方程和參數(shù)值給我反饋。我們故意使用了錯誤的“總需求”方程,并為“邊際消費傾向”分配了一個異常高的值。預期收入積分(預期收入變化,初始預期收入)。?消費消費邊際傾向預期收入。?(GDP預期收入)GDP積分(GDP變化,初始GDP)。?總需求=消費投資政府支出×GDP(總需求+—GDP)?2初始預期收入250。?GDP250。?政府支出80。?10。?凈出口=10。?消費邊際傾向=IF2020時間≤2020.25THEN參考消費邊際傾向×(1—Alpha)ELSE參考消費邊際傾向。?生產調整時間1。?參考消費邊際傾向=80。Alpha=0。Responsesummar(輸入內容未提供足夠信息,無法進行翻譯??偟膩碚f,ChatGPT承認使用常微分方程(ODEs)有方程,并指出了我們在總需求定義中犯的錯誤,指出總需求是其三個組成部分加法模型。它還指出參數(shù)值似乎任意,并對其直覺進行了評論。例如,它指出,預期形成時間的值應該反映代理人對GDP變化的反應速度。它還認識到其他錯誤,建議請注意,邊際消費傾向通常是介于0和1之間的分數(shù)。在你的參數(shù)中,你MPC8010991727,0,Downloaded10991727,0,Downloadedfrom/doi/10.1002/sdr.1773byMassachusettsGeneralHospital,WileyOnlineLibraryon[22/07/2024].SeetheTermsandConditions(/terms-and-conditions)onWileyOnlineLibraryforrulesofuse;OAarticlesaregovernedbytheapplicableCreativeCommonsesneciLsnommoCevitaerCelbacilppaehtybdenrevogeraselcitraAO;esufoselurrofyrarbiLenilnOyeliWno)snoitidnoc-dna-smret/moc.yeliw.yrarbilenilno//:sptth(snoitidnoCdnasmreTehteeS.]4202/70/22[noyrarbiLenilnOyeliWlatipsoHlareneGsttesuhcassaMyb3771.rds/2001.01/iod/moc.yeliw.yrarbilenilno//:sptthmorfdedaolnwoD,0
10valuebetween0and1.”Italsoaddedthatoverlysimplisticmodelsmightnotfitwithinthebroaderframeworkofthemodel.RevisedequationsandvariableImadeamistakeindefiningAggregateDemandandsettingtheparametervalueofMarginalPropensitytoConsume.IhaverevisedthoseandconsideredtheadditiveformulationforAggregateDemandand0.80forMPC.Ialsoaddedtheunitsofmeasureforeachvariablebelow.Theunitsseemfinetome.Canyoucheckmyrevisedequationsandforunitconsistency?[Weintentionallyused
8001它還內容,并考慮了總需求的加法公式和0.80的MPC。我還為下面的每個變量添加了度位一致性嗎?GDP]
[在本文中,我們僅展示修 ed+NetReferenceMarginalPropensitytoConsume=ExpectedIncome:Consumption:ChangeinExpectedIncome:$/(Year×GDP:AggregateDemand:ChangeinGDP:ExpectationFormationTime:InitialGDP:GovernmentExpenditure:Investment:NetExports:MarginalPropensitytoConsume:ResponseChatGPTreviewedandverifiedtherevisedequations.Itcheckedforconsis-fiedinconsistentunits:“TheunitofGovernmentExpenditureshouldbe$/YeartoDOI:
需求總量消費+投資++凈出口。?參考邊際消費傾向0.8預期收入:$/年。?消費:$/年。?預期收入變化:$/(第×年)。?GDP:$/年。?總需求:$/年。?GDP變化:$/季度。?預期形成時間:年。?初始預期收入:$/年。?初始GDP:$/年。?政府際消費傾向:無量綱。?生產調整時間:年。AlphaResponseChatGPT的一致性。它識別出不一致的單位$保持一致。它還建議“GDP變化單位的調整應與方程匹配。它還?2024·SonsLtd10991727,0,Downloaded10991727,0,Downloadedfrom/doi/10.1002/sdr.1773byMassachusettsGeneralHospital,WileyOnlineLibraryon[22/07/2024].SeetheTermsandConditions(/terms-and-conditions)onWileyOnlineLibraryforrulesofuse;OAarticlesaregovernedbytheapplicableCreativeCommonsesneciLsnommoCevitaerCelbacilppaehtybdenrevogeraselcitraAO;esufoselurrofyrarbiLenilnOyeliWno)snoitidnoc-dna-smret/moc.yeliw.yrarbilenilno//:sptth(snoitidnoCdnasmreTehteeS.]4202/70/22[noyrarbiLenilnOyeliWlatipsoHlareneGsttesuhcassaMyb3771.rds/2001.01/iod/moc.yeliw.yrarbilenilno//:sptthmorfdedaolnwoD,0
A.Akhavan和M.S.Jalali:生成式AI和仿真建模butyou’velistedtheunitsas$/Quarter.EitheradjusttheunitsinyourlistorIfixedtheincorrectunits.IthendevelopedthePythoncodebelowformyordi-narydifferentialequations[forpresentationinthisarticle,wereportthecodeinAppendixS2intheonlinesupportinginformation].Iwantthemodeltostartfrom2013totheendof2022andthemodelupdatestohappenattimeintervalsof0.25.Irunthemodel;itworks,andIcanseetheresults.Canyoucheckifthereareanyissueswiththecode?Anythingtoimprove?Donotproviderevisedcode;instead,tellmehowtoimproveitpiecebypiece,asIwouldliketolearnhowtoimprove.Responsetheplotwecreatedtoshowtheresults.Itmentionedthatourcodelackscom-mentsexplainingthepurposeoffunctionsandvariables,makingithardtounder-standthelogicandflow.Italsomentionedseveralvariablenamesareabbreviationsthatmaynotbeimmediatelycleartosomeonereadingthecodetionandsuggesteddefiningseparatefunctionsforeachequationorgroupoftionedthatwespecifiedthetimerangeasnp.arange(2013,2023,0.25),whichdoesnotincludetheendofthelastquarterof2022.Weshouldinsteadextendtherangeto2023.25tomodeltheentireperioduptoandincludingtheendoftion.Italsonotedthattheinitialconditionsshouldmatchthoseof2013andemphasizedtheimportanceofcheckingthemodelagainstreal-worlddata
$$我修復了錯誤的單位。然后,我開發(fā)了下面的Python代碼來展示我的常微分方程[,我們將在附錄S2的在線補充信息中報告該代碼]。我希望模型從2013年開始20220.25ResponseChatGPT及我們創(chuàng)建的用于展示結果的圖表中的單位提出了評論。它提到,我們的代碼缺少解釋函數(shù)和變量目的的注釋,這使得理解邏輯和流程變得困難。它還提到,幾個變量名是縮寫,可能對閱讀代碼的人來說并不立即清楚,并建議使用全描述性名稱。此外,ChatGPT“model”np.arange(2013,2023,0.25)20222023.2520222023Python2013ModeltestingandGenerativeAIcanhelpfacilitateevaluatingandassessingsimulationmodels.Theycanalsoprovidedataanalysisandinterpretation,suggestmodelrefine-ments,facilitatemodelcalibrationandoptimization,andhelpwithsensitivityanalysis.Whiletheycansimulateandanalyzecode,theycaninterpretmodelBelow,Iincludetheupdatedversionofmycodebasedonyourfeedback[seeAppendixS2intheonlinesupportinginformation].Now,I’mtryingtoassessDOI:
生成式AI可以幫助促進評估和評估仿真模型。它們還可以提供數(shù)據(jù)分析與解釋,提和分析代碼,但它們可以解釋模型結果與數(shù)據(jù)的關系(3)并提供增強S2]?,F(xiàn)在,10991727,0,Downloaded10991727,0,Downloadedfrom/doi/10.1002/sdr.1773byMassachusettsGeneralHospital,WileyOnlineLibraryon[22/07/2024].SeetheTermsandConditions(/terms-and-conditions)onWileyOnlineLibraryforrulesofuse;OAarticlesaregovernedbytheapplicableCreativeCommonsesneciLsnommoCevitaerCelbacilppaehtybdenrevogeraselcitraAO;esufoselurrofyrarbiLenilnOyeliWno)snoitidnoc-dna-smret/moc.yeliw.yrarbilenilno//:sptth(snoitidnoCdnasmreTehteeS.]4202/70/22[noyrarbiLenilnOyeliWlatipsoHlareneGsttesuhcassaMyb3771.rds/2001.01/iod/moc.yeliw.yrarbilenilno//:sptthmorfdedaolnwoD,0
12Fig.3.Usingdata
圖3.在ChatGPT提示 model.Withmymodelvariablesbeingbasedontheirhistoricaldata,Iexpectthatmymodeloutcome,thatis,simulatedGDP,bereasonablyclosetothehistor-icaltrendsofGDP.ButwhenIplottheSimulatedGDPagainsthistoricaldata,Idon’tseeaclosereplicationofthehistoricaltrend.Iamsureaboutthedataqual-ity,though,andIalreadycheckedforunitconsistency.Givemefeedbackbasedonthecodeandtheattachedresultsfigure.Iaimtolearnhowtoimprovethemodeltoreplicatethehistoricaltrendclosely.ResponsetoensuretheCOVID-19shockisappropriatelymodeledandtheassumptionsment,governmentexpenditure,andnetexports.(Inourinitialmodelensurethattheextrapolationbeyondthedatasetisreasonableandalignswitheco-tionFormationTime,ProductionAdjustmentTime,andReferenceMarginalexaminingthesystemofOrdinaryDifferentialEquationstoensuretheycapturethecomplexdynamicsofaneconomyduringshockperiods.Italsosuggestedconsider-DOI:
模型。我的模型變量基于其歷史數(shù)據(jù),我預計我的模型結果,即模擬GDP與GDP的歷史趨勢相當接近。但當我將模擬GDP與歷史數(shù)據(jù)對比時,我沒有看到歷ChatGPT對COVID?19析提供了反饋。首先,它要求我們確保COVID?19沖擊得到適當?shù)慕?,并且假設適如投資、政府支出和凈出口。(在我們的初始模型概念化中,如圖1所示,我們只提到了這三個變量,但它們的名稱中沒有包含“數(shù)據(jù)”這樣的字樣;然而,在收到這一10991727,0,Downloaded10991727,0,Downloadedfrom/doi/10.1002/sdr.1773byMassachusettsGeneralHospital,WileyOnlineLibraryon[22/07/2024].SeetheTermsandConditions(/terms-and-conditions)onWileyOnlineLibraryforrulesofuse;OAarticlesaregovernedbytheapplicableCreativeCommonsesneciLsnommoCevitaerCelbacilppaehtybdenrevogeraselcitraAO;esufoselurrofyrarbiLenilnOyeliWno)snoitidnoc-dna-smret/moc.yeliw.yrarbilenilno//:
溫馨提示
- 1. 本站所有資源如無特殊說明,都需要本地電腦安裝OFFICE2007和PDF閱讀器。圖紙軟件為CAD,CAXA,PROE,UG,SolidWorks等.壓縮文件請下載最新的WinRAR軟件解壓。
- 2. 本站的文檔不包含任何第三方提供的附件圖紙等,如果需要附件,請聯(lián)系上傳者。文件的所有權益歸上傳用戶所有。
- 3. 本站RAR壓縮包中若帶圖紙,網頁內容里面會有圖紙預覽,若沒有圖紙預覽就沒有圖紙。
- 4. 未經權益所有人同意不得將文件中的內容挪作商業(yè)或盈利用途。
- 5. 人人文庫網僅提供信息存儲空間,僅對用戶上傳內容的表現(xiàn)方式做保護處理,對用戶上傳分享的文檔內容本身不做任何修改或編輯,并不能對任何下載內容負責。
- 6. 下載文件中如有侵權或不適當內容,請與我們聯(lián)系,我們立即糾正。
- 7. 本站不保證下載資源的準確性、安全性和完整性, 同時也不承擔用戶因使用這些下載資源對自己和他人造成任何形式的傷害或損失。
最新文檔
- 土地租賃合同范文匯編(20篇)
- 武館入學合同協(xié)議書范本
- 員工分紅入股合同協(xié)議
- 和平協(xié)議書格式
- 員工買賣合同協(xié)議書范本
- 商品混凝土勞動合同協(xié)議
- 唐山土地流轉合同協(xié)議
- 商務英文合同協(xié)議
- 品牌西服采購合同協(xié)議
- 荒山養(yǎng)殖承包合同(19篇)
- 級進模具畢業(yè)設計說明書
- 兒童呼吸機基本使用
- 手術替代治療方案
- 建筑工程項目管理績效評價指標體系
- T-CNPPA 3025-2023 藥包材質量協(xié)議管理指南
- 美洲印第安文明的發(fā)展與衰落-瑪雅、阿茲特克與印加
- 家庭教育講座活動流程
- 大學《思想道德與法治》期末考試復習題庫(含答案)
- 麥肯錫入職培訓第一課在線閱讀
- 省級課題結題報告范本
- 什么叫干槽癥課件
評論
0/150
提交評論