⋅ Robert John Hani Hagras Oscar Castillo Editors Type-2 Fuzzy Logic and Systems Dedicated to Professor Jerry Mendel for his Pioneering Contribution 123 Editors RobertJohn Oscar Castillo LUCID Research Group,School Division of Graduate Studies ofComputer Science TijuanaInstitute of Technology TheUniversity of Nottingham Tijuana, BajaCalifornia JubileeCampus Mexico Nottingham, Nottinghamshire UK HaniHagras Schoolof Computer Science andElectronic Engineering TheComputational Intelligence Centre University of Essex Colchester UK ISSN 1434-9922 ISSN 1860-0808 (electronic) Studies in FuzzinessandSoft Computing ISBN978-3-319-72891-9 ISBN978-3-319-72892-6 (eBook) https://doi.org/10.1007/978-3-319-72892-6 LibraryofCongressControlNumber:2017962051 ©SpringerInternationalPublishingAG2018 Thisworkissubjecttocopyright.AllrightsarereservedbythePublisher,whetherthewholeorpart of the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation, broadcasting, reproduction on microfilms or in any other physical way, and transmission orinformationstorageandretrieval,electronicadaptation,computersoftware,orbysimilarordissimilar methodologynowknownorhereafterdeveloped. The use of general descriptive names, registered names, trademarks, service marks, etc. in this publicationdoesnotimply,evenintheabsenceofaspecificstatement,thatsuchnamesareexemptfrom therelevantprotectivelawsandregulationsandthereforefreeforgeneraluse. The publisher, the authors and the editors are safe to assume that the advice and information in this book are believed to be true and accurate at the date of publication. Neither the publisher nor the authorsortheeditorsgiveawarranty,expressorimplied,withrespecttothematerialcontainedhereinor for any errors or omissions that may have been made. The publisher remains neutral with regard to jurisdictionalclaimsinpublishedmapsandinstitutionalaffiliations. Preface ThisbookisdedicatedtoJerry,Prof.JerryMendel,forhispioneeringworksonthe type-2 fuzzy sets and systems. Jerry has had a long and distinguished academic careerinbothsignalprocessingandfuzzylogic,winningmanyawardsandhonours. However,sincehisfirstpaperin1998hehasbeentheleaderinthefascinatingarea ofthetype-2fuzzylogicfieldfornearly20years.Hehasover40journalarticlesin leadingjournalswithmanycitations.HisthreemostcitedpapersinGoogleScholar (as forJuly2017) are: “Type-2fuzzy setsmade simple”,J.M.Mendel,R.I.John, IEEE Transactions on Fuzzy Systems 10 (2), 117–127, 2002 (1788 citations), “Interval type-2 fuzzy logic systems: theory and design”, Q. Liang, J. M. Mendel, IEEE Transactions on Fuzzy Systems 8 (5), 535–555, 2000 (1380 citations) and “Type-2 fuzzy logic systems”, N. N. Karnik, J. M. Mendel, Q. Liang, IEEE trans- actionson Fuzzy Systems 7 (6), 643–658, 1999 (1200 citations). Jerry has worked with numerous Ph.D. students and colleagues from across the world,alwaysinacollaborativewaytomovethefieldforward.Wehavehadmany long hours discussing important research issues in type-2 fuzzy logic. Over that time he has become our friend and we are honoured to put together this invited collection of contributions. The chapters here cover a wide variety of topics—the type-2 fuzzy sets and the game Go, weighted averages, control of agricultural vehicles, challenges for the type-2fuzzycontrol,type-2fuzzycontrolingames,patternrecognitionandtherole of type-2 fuzzy sets in intelligent agents, just to mention a few. We would like to thank the authors for their interesting contributions. The diversity of topics covered and views and perspectives presented reflects the diversity in the type-2 community. If you are new to type-2 fuzzy logic, we hope you are inspired to read these and follow up on Jerry’s work. Nottingham, UK Robert John Colchester, UK Hani Hagras Tijuana, Mexico Oscar Castillo Spring 2017 v Contents From T2 FS-Based MoGoTW System to DyNaDF for Human and Machine Co-learning on Go . . . . . . . . . . . . . . . . . . . . . 1 Chang-Shing Lee, Mei-Hui Wang, Sheng-Chi Yang and Chia-Hsiu Kao Ordered Novel Weighted Averages . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25 Dongrui Wu and Jian Huang On the Comparison of Model-Based and Model-Free Controllers in Guidance, Navigation and Control of Agricultural Vehicles. . . . . . . . 49 Erkan Kayacan, Erdal Kayacan, I-Ming Chen, Herman Ramon and Wouter Saeys Important and Challenging Issues for Interval Type-2 Fuzzy Control Research . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75 Hao Ying Type-2 Fuzzy Logic in Pattern Recognition Applications. . . . . . . . . . . . 89 Patricia Melin Type-2 Fuzzy Logic Control in Computer Games . . . . . . . . . . . . . . . . . 105 Atakan Sahin and Tufan Kumbasar A Type-2 Fuzzy Model to Prioritize Suppliers Based on Trust Criteria in Intelligent Agent-Based Systems. . . . . . . . . . . . . . . 129 Mohammad Hossein Fazel Zarandi, Zohre Moattar Husseini and Seyed Mohammad Moattar Husseini vii From T2 FS-Based MoGoTW System to DyNaDF for Human and Machine Co-learning on Go Chang-Shing Lee, Mei-Hui Wang, Sheng-Chi Yang and Chia-Hsiu Kao Abstract ThischapterdescribestheresearchfromT2FS-basedMoGoTWsystem to DyNamic DarkForest (DyNaDF) open platform for human and machine co-learningonGo.AhumanGoplayer’sperformancecouldbeinfluencedbysome factors, such as the on-the-spot environment as well as physical and mental situ- ations of the day. In the first part, we used a sample of games played against machine to estimate the human’s strength (Lee et al. in IEEE Trans Fuzzy Syst 23 (2):400–420,2015[1]).TheType-2FuzzySets(T2FSs)withparametersoptimized by a genetic algorithm for estimating the rank was presented (Lee et al. in IEEE Trans Fuzzy Syst 23(2):400–420, 2015 [1]). The T2 FS-based adaptive linguistic assessmentsysteminferredthehumanperformanceandpresentedtheresultsusing thelinguisticdescription(Leeetal.inIEEETransFuzzySyst23(2):400–420,2015 [1]).InMarch2016,GoogleDeepMindchallengematchbetweenAlphaGoandLee Sedol in Korea was a historic achievement for computer Go development. In Jan. 2017, an advanced version of AlphaGo, Master, won 60 games against some top professional Go players. In May 2017, AlphaGo defeated Ke Jie, the top profes- sionalGoplayer,attheFutureofGoSummitinChina.Insecondpart,weshowed the development of computational intelligence (CI) and its relative strength in comparisonwithhumanintelligenceforthegameofGo(Leeetal.inIEEEComput Intell Mag 11(3):67–72, 2016 [2]). Additionally, we also presented a robotic pre- dictionagenttoinferthewinningpossibilitybasedontheinformationgeneratedby DarkForest Go engine and to compute the winning possibility based on the partial game situation inferred by FML assessment engine (Lee et al. in FML-based pre- diction agent and its application to game of Go, 2017 [3]). Moreover, we chose seven games from 60games toevaluatetheperformance (Lee etal.inFML-based prediction agent and its application to game of Go, 2017 [3]). In this chapter, we extract the human domain knowledge from Master’s 60 games for giving the desiredoutput.Then,wecombineParticleSwarmOptimization(PSO)andFMLto learn theknowledge baseandfurtherinfer thegame results ofGoogleAlphaGoin May2017.Theexperimentalresultsshowthattheproposedapproachisfeasiblefor C.-S.Lee(✉) ⋅ M.-H.Wang ⋅ S.-C.Yang ⋅ C.-H.Kao NationalUniversityofTainan,Tainan,Taiwan e-mail:[email protected] 2 C.-S.Leeetal. the application to human and machine co-learning on Go. In the future, powerful computer Go programs such as AlphaGo are expected to be instrumental in pro- moting Go education and AI real-world applications. 1 Introduction Many real-world applications are with a high-level of uncertainty. Type-2 FS (T2 FS) has the ability to capture the uncertainty about membership functions offuzzy sets [4, 5]. Moreover, Type-2 Fuzzy Logic System (T2 FLS) is used to handle the high uncertainties in the group decision-making process as it can model the uncertainties between expert preferences by using T2 FSs [4–7]. Because of the popularity of T2 FS [7], the Type-2 Fuzzy Markup Language (T2 FML), an extensionof the FML grammar, isdeveloped to allow system designers toexpress theirexpertisebyusinganIntervalType-2FuzzyLogicSystem(IT2FLS)tomodel type-2fuzzysetsandsystems[8–10].FuzzyMarkupLanguage(FML)hasbecome an IEEE Standard since May 2016 and provides designers of intelligent decision makingsystemswithaunifiedandhigh-levelmethodologyfordescribingsystems’ behaviorsbymeansofrulesbasedonhumandomainknowledge[8,10].FMLisa fuzzy-based markup language that can manage fuzzy concepts, fuzzy rules, and a fuzzy inference engine [8, 10]. Additionally, FML is with the following features: understandability,extendibility,andcompatibilityofimplementedprogramsaswell as efficiency of programming [10]. The main advantage of using FML is easy to understand and extend the implemented programs for other researchers [8, 10]. The game of Go is played by two players, Black and White. Two Go players alternatively play their stone at a vacant intersection of the board by following the rules of Go [11]. Additionally, Go is regarded as one of the most complex board games because of its high state-space complexity 10171, game-tree complexity 10360, and branching factor 250 [12]. The skill of amateur players in Go is ranked according to kyu (K) in the lower tier, where a smaller number stands for stronger playingskill(with1 Kbeingthehighestskilllevel),anddan(D)inthehighertier, wherealargernumberstandsforstrongerplayingskill.ProfessionalGoplayersare ranked entirely in dan, abbreviated with the letter P [2]. Go is typically played on 19 × 19 size boards, but 9 × 9 size boards are also common for beginners. The complexity of the 9 × 9 game is far less than the standard game, and the 9 × 9 game had been one of the interim goals for computer Go programs [2]. The handicapsforthehumanvs.computer19 × 19gamehavebeendecreasedfrom29 in 1998 to 0 in 2016 [2]. In May 2017, AlphaGo even defeated Ke Jie, the top professionalGoplayersintheworld, attheFuture ofGo Summit inWuzhen [13]. Owningtothequickadvanceinartificialintelligence,currentlypowerfulcomputer Go programs such as AlphaGo [14] and DeepZenGo are expected to give top professional humans a few handicap stones to make for an even match. Games have served as one of the best benchmarks for studying artificial intel- ligence [15, 16]. Over the last few years, Monte Carlo Tree Search (MCTS) has FromT2FS-BasedMoGoTWSystem… 3 already made a profound effect on artificial intelligence, especially in computer games[15].GellyandSilver[17]appliedRapidActionValueEstimation(RAVE) algorithm and Heuristic MCTS to a computer Go program, MoGo. Monte Carlo tree search (MCTS), minorization-maximization (MM), and deep convolutional neuralnetworks(DCNNs)havedemonstratedgreatsuccessinGo[2,14,18,19].In December of 2014, two teams applied deep convolutional neural networks to Go independently [20,21]. Among many ofDCNN’sapplications, it hasseensuccess in image and video recognition. When applied to Go, DCNN is able to recognize move patterns at a significantly lower error rate than MM. For this reason, most state-of-the-art computer Go programs use MCTS combined with either MM or DCNN [2]. ForevaluatingthehumanperformanceonGogames,humanscouldbeadvanced toahigherrankbasedonthenumberofwinninggamesviaaformalhumanagainst humancompetition[1].However,theinvitedhumanGoplayer’sstrengthmightbe affectedbysomefactors,suchastheon-the-spotenvironment,physicalandmental situations of the day, and game settings, so the Go player’s rank may be with an uncertain possibility. Additionally, one player’s strength may gradually decrease becauseofgettingolderorseldomplayingwithastrongerhuman[1].Hence,these uncertainfactorscausethedifficultiesanduncertaintyinevaluatingtherankofone human Go player. In [1], we used T2 FSs to model the requirements of a person specification that is reflective of all the experts’ opinions and this can be used to provide a good evaluation for the rank of the Go players. A T2 FS-based adaptive linguistic assessment system was proposed to evaluate one human Go player’s performance with a semantic analysis such that the proposed system is helpful to increase the human Go player’s enthusiasm for playing with the computer Go program [1], especially for children. In [2], we helped the readership better understand how the development of computer Go programs has arrived at this milestone of winning against one of the top human players, and how IEEE Computational Intelligence Society (CIS) has been involved in this process. This huge achievement in AI is based largely on CI methods,includingDCNNs,supervisedlearningfromexpertgames,reinforcement learning, the use of the value network and policy network, and MCTS. In [3], we constructed a DyNamic DarkForest (DyNaDF) Cloud Platform for game of Go, including a demonstration game platform, a machine recommendation platform, and an FML assessment engine. We used the first-stage prediction results of DarkForest Go engine [18, 19] and the second-stage inferred results of the FML assessment engine [22], we further introduced the third-stage FML-based decision support engine to predict the winner of the game and chose seven games from Master’s 60 games in Jan. 2017 [3, 23] to evaluate the performance. The fourth-stagerobotenginereportsreal-timesituationtoplayers.Thischapterfurther combinesFMLandparticleswarmoptimization(PSO),calledPFML[24],tolearn the domain knowledge of Master’s 60 games [23] by referring to the book pub- lishedinTaiwan[25].Afterlearning,weusetheproposedapproachesin[3]toinfer the game results of the Future of Go Summit in Wuzhen in May 2017. From the experimental results, we can get much higher accuracy than before learning. 4 C.-S.Leeetal. The remainder of this chapter is organized as follows: Sect. 2 introduces the research performance from the proposed T2 FS-based MoGoTW system [1, 2] to theFML-basedDyNaDFopensystem[3,22].Section 3isdedicatedtothehuman and machine co-learning part based on T2 FS and FML. Section 4 shows some experimental results. Finally, conclusions are given in Sect. 5. 2 From T2 FS-Based MoGoTW Linguistic Assessment System to FML-Based DyNamic DarkForest Open Platform This section introduces the research performance from the proposed T2 FS-based MoGoTWlinguisticassessmentsystem[1]totheconstructedFML-basedDyNaDF open platform [3]. 2.1 T2 FS-Based MoGoTW System for Adaptive Linguistic Assessment Over the past years, there were many Go competitions between humans and computer Go programs held in Taiwan or in the world [2, 26]. However, playing with the computer Go program may be boring because the computer Go program cannot express its feelings, especially in one lopsided game [27, 28]. If the com- puter Go program is able to adaptively assess its opponent’s strength and provide one real-time feedback mechanism for Go players,it will behelpfulfor humansto increase their interest in playing with the computer Go program and to find their relevant opponents and/or relevant handicap. Upper Confidence Bounds for Trees (UCT) isthe mostpopularalgorithm intheMCTSfamily[15,29, 30].MoGoTW, developed based on MoGo 4.86 Sessions plus the Taiwan (TW) modifications, plays its move at the board according to the result of the best-move selection mechanism [1]. The strength of MoGoTW is increased when it loses and is decreased when it wins based on item response theory (IRT) [31]. In [1], we proposed the T2 FS-based adaptive linguistic assessment system to evaluate human Go player’s performance whose structure is shown in Fig. 1. TheMCTSSimulationNumber(SN)isadjustedtomeetthestrengthoftheopponent duringoneround andtheiroperationsaredescribedasfollows:(1)Adjustmentin per-moveSN:SNisincreasedbymultiplyingbyV whentheWinningRate(WR)of 1 thecomputerprogramislessthanWR .(2)Adjustmentinper-gameSN:Whenone 1 game ends and the computer program wins the game, the computer program weakensitsstrengthbydividingSNintoV fornextgame.(3)Theinvolvedhuman 2 Go players compete K games against MoGoTW for one round. During the com- petition,MoGoTWadjustsitsstrengthtomatchwiththestrengthofthehumanGo FromT2FS-BasedMoGoTWSystem… 5 Game Results Repository Domain Expert Domain Expert Game Results T2FS Construction Repository Mechanism KB/RB Adaptive Go-Ranking Repository Assessment Ontology PSO Model Human vs. MoGoTW Estimation Mechanism T2FS-based Genetic Learning Adaptive UCT-based Mechanism Go-Ranking Mechanism Bradley-Terry Model Estimation Mechanism T2FS-based Fuzzy Inference Mechanism MoGoTW Players Human-Performance Mapping Mechanism Semantic Analysis Mechanism Personal Profile Repository Players Rank Repository Fig.1 T2FS-basedadaptivelinguisticassessmentsystem[1] player by increasing or decreasing MCTS’s simulation number after playing one move and one game. (4) If human wins the first game, then MoGoTW increases MCTS’ssimulationnumbertostrengthenitsownstrengthatthestartofthesecond game tocompetewith thehuman. Inotherwords,themore consecutivegames are wonbythehuman,thestrongerthehuman.GameWeightdenotesthestrengthofall thegamesplayedbythishuman.ThehigherGameWeight,thestrongerthehuman. WinningRate denotes the winning rate of the human after playing games with MoGoTW. Based on this concept, T2 FSs GamgeWeightLow, GamgeWeightMedium, GamgeWeightHigh WinninggRateLow, WinninggRateMedium, and WinninggRateHigh are constructed. 2.2 FML-Based DyNamic DarkForest (DyNaDF) Open Platform Figure 2 shows the structure of the DyNaDF open Platform whose brief descrip- tions are given as follows [3]: (1) The DyNaDF open platform for game of Go applicationiscomposedofaplaying-GoplatformlocatedatNationalUniversityof Tainan (NUTN)/Taiwan and National Center for High-Performance Computing (NCHC)/Taiwan, a DarkForest Go engine located at Osaka Prefecture University 6 C.-S.Leeetal. Tokyo Tainan Osaka OPU/Japan TMU/Japan NUTN/NCHC HumanGo Players Darkforest Go Engine PALRO Playing-Go Platform FMLAssessment Engine Go eBook Fig.2 StructureofDyNamicDarkForestopenplatformforGo[3] (OPU)/Japan, and the robot PALRO from Tokyo Metropolitan University (TMU)/ Japan. (2) Human Go players surf on the DyNaDF platform located at NUTN/ NCHC to play with DarkForest Go engine located in OPU. (3) The FML assess- ment engine infers the current game situation based on the prediction information from DarkForest and stores the results into the database. (4) The robot PALRO receives the game situation via the Internet and reports to the human Go players. Human can learn more information about game’s comments via Go eBook [3]. Figure 3 shows the screenshot of the game between Ke Jie (9P) as Black and Master as White on Dec. 30, 2016 provided by the FML-based DyNaDF Open Platform [3]. 3 Human and Machine Co-learning Based on T2 FS and FML This section introduces the human and machine co-learning based on T2 FS and FML.Section 3.1describestheadaptivehumanperformanceevaluationongameof Go. The FML-based prediction agent for DyNaDF open platform is described in Sect. 3.2.