[PDF] ALLURE: A Multi-Modal Guided Environment for Helping Children

View - Download ALLURE: A Multi-Modal Guided Environment for Helping Children

Previous PDF

Next PDF

ALLURE

*: AMulti-Modal GuidedEn vironmentforHelpingChildr enLearnto Solvea Rubik"sCube withA utomaticSolving andInteractiveExplanations

Kausik Lakkaraju

1, ThahimumHassan 2, VedantKhandelwal3, PrathamjeetSingh 4,

Cassidy Bradley

5, RonakShah 6, ForestAgostinelli7, BiplavSriv astava8, DezhiW u9

AI Institute,Uni versityofSouthCarolina,Columbia, USA,

1,3,6,7,8

Department ofInte gratedInformationTechnology ,Uni versityofSouthCarolina,Columbia, USA

2,4,5,9

kausik@email.sc.edu,

1thassan@email.sc.edu,2vedant@mailbox.sc.edu,3psingh@email.sc.edu4, cmb11@email.sc.edu,5

rrshah@email.sc.edu,

Abstract

Modern artificialintelligence (AI)met hodsha vebeenused to solveproblemsthat many humansstruggle tosolve.This opens upne wopportunitiesforkno wledgedisco very anded- ucation. WedemonstrateALLURE, acollaborati ve educa- tional AIsystem forlearning tosolv ethe Rubik"s cubethat is designedto helpstudents improv etheir problemsolving skills. ALLUREcan bothfind itso wnstrate giesfor solving the Rubik"scubeand explain thosestrate giestohumans.In the future,ALLURE willalso beable tocollaborate withhu- mans byb uildingonuser-provided strategies forsolvingthe Rubik"scube andas wellas generalizeto othersearch andau- tomated planningproblems. Interactionbetween AIand user is facilitatedbyvisual andnatural languagemodalities.

Introduction

Weseek tocreate aplatform forkno wledgedisco very where humans cancollaborate withartificial intelligence(AI) to obtain newideasand totest theiro wnideas forsolving a problem(Agostinelli etal. 2021).This hasapplications to botheducation aswell asthe frontiersof research.Our project andits prototype,ALLURE, seeksto addressthe gap and rolethat today"s AItechnologycanplay ineducation through aRubik" scubeusecase. Ourtool also contributes to abroader scopeof today"s educationthrough interactive, multi-modal, ande xplainableAI(XAI)user interfaces to facilitateand dev elopstudents"twenty-firstcenturyskills, including communication,collaboration, problem-solving, critical, andcreati vethinkingtofosterlifelong learningand digital citizenship(Mayer andW ittrock2006; Ng2015;

Conn andMcLean 2019).

Wepick theRubik" scube givenits apparentability to capture theimagination ofpeople aroundthe world (Ru- bik 2020)as wellas itsrelationship tomathematics (Joyner

2008) andcomputer science(K orf1997; Rokickietal.2014;

Agostinelli etal. 2019).Its abilityto fosterproblem-solving abilities canbe seenby international"cubing" competitions, including competitionsfor solvingthe cubein thef astest time, solvingthe cubein thein thefe westnumber ofmo ves, and solvingmultiple cubesblindfolded. *

Name derivedfromtheunderlined letters.

Copyright© 2022,Association forthe Advancement ofArtificial Intelligence (www.aaai.org).Allrightsreserv ed.Figure 1:The Solved whitecross In thisdemonstration, wewill show how ALLUREisable to successfullyfind astrate gyfor achievingthewhite cross (shownin Figure1) anduse botha visualand naturallan- guage interfacetoe xplainits strategytothe user. Astep- by-step illustrationof thevisual interface issho wninFigure 2.

Explainable AIMethods

to thecomputer sciencecommunity fordecades (Korf 1997; Rokicki etal. 2014).Recently ,a methodcalledDeepCubeA (McAleer etal. 2018;Agostinelli etal. 2019)used AItech- niques tocreate adomain-independent algorithmthat learns to solvetheRubik" scube, amongotherpuzzles.This auto- mated problem-solvingalgorithm hassince beenapplied to cryptography(Jin andKim 2020)and quantumcomputing (Zhang etal. 2020).Ho wev er,likepreviousmethods,Deep- CubeA cannote xplaintoa humanitsov erallstrate gybehind its solutionsand cannotb uildon human-providedstrategies. Togi veDeepCubeAtheabilityto producee xplainableso- lutions, wedra woninspirationfrom how humanse xplain the manydifferent methodsforsolvingthe Rubik"s cube (Aggarwal2017).In almostall ofthese methods,there isa series ofsubgoals tobe achiev edand eachsubgoalhasaset of algorithmsused toachie ve it.Therefore,ALLUREwill

collaborate withthe userby communicatingsubgoals andTheTh irty-SixthAAAIConference onArtifi cialIntelligence(AAAI-22)

13185
(a)(b)(c)(d)(e)

Figure 2:Solving White Crossproblem inRubik" scube.(a)Initial state:White-OrangeEdge Pieceis notin place,(b) Step01:

perform D",(c) Step02: performF", (d)Step 03:perform: R,(e) Goalstate: performF toget White Cross. algorithms ithas discov eredwhilebuildingonsubgoals and algorithms suggestedby theuser .

Toallo wDeepCubeAtoquickly adaptto new subgoals,

we usehindsight experience replay(Andrychowiczet al.

2017) totrain av ersionof DeepCubeAthatcanachie ve any

validRubik" scubeconfiguration,including subgoalswhere only asubset ofthe stickers arespecified. DeepCubeAcan then beused topro videus withexamplesof how toachie ve anyparticular subgoal.F ore xample,ifwew ouldliketo achievesub goali, wefirst scramble thecubebytaki nghun- dreds ofrandom mov es,thenuseDeepCubeAachiev esub- goali-1, thencollect examples ofthestatesseen andac- tions takenwhengoing fromsubgoal i-1to subgoali.

From thesee xamples,wenow needto extracthuman-

understandable algorithms.T oaccomplishthis,we usefirst- order inductivelogicprogramming(ILP) (Muggleton1992). ILP givesustheability todefine relationalpredicates that ground knowledge,candescribe algorithmsin aconcise logic program.W eusethePopper ILPsoftw are(Cropper and Morel2021) thatalso allows usto includeexpressiv e inductivebiasesto simplifythe searchfor aprogram. For ex- ample, wecan giv etheprogramcommonsenserules, such as aspecifying thata sticker cannotbe inmorethanone lo- cation ata time.W ecan alsouseILPto discov era hierarchy of subgoalsby findingpatterns inthe orderin whichalgo- rithms areapplied andthe learnedpreconditions forthose algorithms. Wenow turntoILPand HCImethods forthe collaboration betweenhuman andAI.

CollaborativeInterface

User interface(UI)design ischallenging toef fectiv elydis- play solutionsto theRubik" sCube withunderstandable problem-solving stepsto demoand engage userswith real- time cuberotations driv enbyAIalgorithms.Moreov er, al- lowingthe userto communicatetheir own ideasadds an- other layerof difficulty forsuchadesign. To facilitate the solving, wedesign andimplement (1)a seriesof multi- modal userinterf acesthrough3Dmodel visualizationsto visually presentRubik" sCubeandreal-time cubemanipu- lations withe xplainablestepsobtainedfrom theAI system (this portionb uildsoncodefrom (cfopand Hui2020) and visualizations from(Y ouCanDotheRubik"sCube2021))(2)

an interactivechatbotsystemto provide learners customizedpedagogical hintsand guidance,af fordedby multi-modalUIs viate xt,image,animationand voice.

Weuse NLPtechniques whileha vingthe collaborativ eas- sistant (chatbot)interact withthe userto understandtheir gamingand learningintents, trackprogress andmanage in- teractions. Theuser canrefer topre viouscon versation as part oftheir context toinitiateag ame.The systemwill be able totrack user"s conversationand assessemotionslike sentiments todetect frustration,user engagedness andlearn- ing progress.The systemtranslates explanations fromfirst order logicto naturalte xtusing templatesthatarecustomiz- able todif ferentemotionstatesof theuser .W euse Rasa (Bocklisch etal. 2017)to implementthis NLPinterf ace. However,werecognizetheneed forcareful usere valuation to determinethe effecti venessofthesemethodsfordiffer - ent usergroups giv entheirsusceptibilitytotrustissues like bias (Srivastavaetal.2020;Kiritchenko andMohammad

2018). Furthermore,users willha ve theabilitytocommu-

nicate theiro wnideasforsubgoals usinga virtualcube that theycan edit.Users willbe ableto communicatetheir ideas for algorithmsby providing thesystemwithe xamplesof the algorithm andha vingtheILPsystem inducea logicpro- gram, whichwillthen betranslated tonatural languagefor verification.Alternati vely,theusercandirectlycommuni- cate theiralgorithm usingnatural language,with thesystem asking forclarification whenencountering any ambiguities.

Discussion andConclusion

In thedemonstration, wefirst visuallyintroduce theuser to the Rubik"scubeand conv entionsof moves.Then,theAI agent learnsthe crossand explainsthesolution withvisual cues. Theagent giv esanaturallanguageexplanation ofthe algorithm beingused andits effects. Next, astep-by-stepvi- sualization showsthemo ves requiredtoobtainthecross. Toour knowl edge,ALLUREisthefirst toolto bothl earn to solveRubik"s cubeandexplain itssolutions tohumans. In thefuture, wewill incorporatecollaborati ve learning,per - sonalized interventions,multi-lingualcapabilities, andcon- duct userstudies withstudents. Sincethe Rubik"s cubeis a well-studiedproblem inthe fieldsof searchand planning (Korf1987, 1997),ALLURE canpotentially begeneralized to widerange ofother problemsin thesefields. 13186

References

Aggarwal,A.2017. Listof methods.https://www .

speedsolving.com/wiki/index.php/Listofmethods. Ac- cessed: 2021-12-01. Agostinelli, F.;Mav alankar,M.;Khandelwal,V.; Tang, H.; Wu,D.; Berry, B.;Srivasta va, B.;Sheth,A.;andIrvin,M.

2021.Designing Children"sNewLearni ngPartner:Collab-

orativeArtificial Intelligence forLearningtoSolve theRu- bik"sCube , 610-614.Ne wYork,NY ,USA:Associationfor

Computing Machinery.ISBN9781450384520.

Agostinelli, F.;McAleer, S.;Shmakov ,A.; andBaldi,P.

2019. Solvingthe Rubik"s cubewithdeepreinforcement

learning andsearch. NatureMac hineIntelligence, 1(8):

356-363.

Andrychowicz,M.; Wolski, F.;Ray, A.;Schneider,J.; Fong, R.; Welinder,P.;McGre w,B.;T obin,J.;Abbeel,O.P.; and Zaremba, W.2017.Hindsight experience replay. InAd- vances inNeur alInformationProcessing Systems , 5048- 5058.
Bocklisch, T.;Faulkner ,J.;Pawlowski, N.;and Nichol,A.

2017. Rasa:Open sourcelanguage understandingand dia-

logue management.arXiv preprintarXiv:1712.05181. cfop; andHui, H.G. 2020.Three-rubiks-cube. https: 01. Conn, C.;and McLean,R. 2019.BulletproofPr oblemSolv- ing: TheOne Skillthat Changes Everything . JohnW iley& Sons.

Cropper,A.; andMorel, R.2021. Learningprograms by

learning fromf ailures.MachineLearning , 110(4):801-856.

Jin, J.;and Kim,K. 2020.3D CUBEAlgorithm forthe

KeyGenerationMethod: ApplyingDeep NeuralNetw ork

Learning-Based.IEEE Access, 8:33689-33702.

Joyner,D.2008. Adventuresin group theory:Rubik"sCube ,

Merlin"smac hine,andothermathematicaltoys

. JHUPress. Kiritchenko,S.; andMohammad, S.2018. ExaminingGen- der andRace Biasin Tw oHundred SentimentAnalysisSys-quotesdbs_dbs41.pdfusesText_41

[PDF] sassenage

[PDF] comment faire un rubik cube 3x3

[PDF] tous différents mais tous égaux

[PDF] solution rubik's cube 3x3x3 rapide

[PDF] notice rubik's cube 2x2

[PDF] symbole somme word

[PDF] rubik's cube pdf

[PDF] rubik cube 3x3 solution francais

[PDF] tous différents tous égaux cycle 3

[PDF] tous différents tous égaux cycle 2

[PDF] tous différents tous semblables

[PDF] acide oléique formule semi développée

[PDF] acide palmitoléique formule

[PDF] nous sommes une génération choisie accords

[PDF] je sais qui je suis pdf