Fuzzy Reference Gain-Scheduling Approach as Intelligent Agents: FRGS Agent

Transcription

1 Fuzzy Referece Gai-Schedulig Approach as Itelliget Agets: FRGS Aget J. E. ARAUJO * eresto@lit.ipe.br K. H. KIENITZ # kieitz@ita.br S. A. SANDRI sadra@lac.ipe.br J. D. S. da SILVA demisio@lac.ipe.br * Itegratio ad Testig Laboratory LIT Computer Sci. ad Applied Math. Associated Lab LAC Istituto Nacioal de Pesquisas Espaciais INPE Av. Astroautas, São José dos Campos (SP) BRAZIL # Departmet of Electroics Egieerig Istituto Tecológico de Aeroáutica ITA Praça Marechal Eduardo Gomes, 5 São José dos Campos (SP) BRAZIL Abstract Goal drive Itelliget Agets ad Fuzzy Referece Gai-Schedulig (FRGS) approach are described i this paper as iterchageable cocepts that are able to deal with dyamic complex problems. It is advocated that the FRGS approach may be viewed as a autoomous goal-based aget, that is, a fuzzy logic based aget able to autoomously adapt itself to evirometal chages itroduced by exteral iputs. The cocept of fuzzy systems ad itelliget aget are employed i decisio-makig problems to lead to a certai degree of autoomy i decisio support system. Although the FRGS method was origially proposed for cotrol applicatio, this approach was exteded to decisio-makig tasks due to its ability of emulatig huma reasoig. This ew aget approach uses the exteral iput iformatio also deomiated referece (goal) as the drive mechaism to determie the behavior of the system i order to achieve the desired objectives (goal). Thus, the FRGS approach ca be modeled i the framework of a adaptive ad goal (also cotext or eviromet) drive aget. I. INTRODUCTION The cocept of itelliget aget has bee employed alog with fuzzy sets i decisio-makig problems to lead to a certai degree of autoomy i decisio support systems [2][5][9]. Relevat applicatio areas for this cocept are plaig, schedulig, ad cotrol problems, where oe ca collect data, moitor processes ad make local itelliget decisios, as well as recommedatios [7][8]. The fuzzy theory proposed by Zadeh to deal with fuzzy sets ad afterwards with fuzzy logic has bee provig its effectiveess ad showig its value i may fields of research. Although the future of fuzzy approach keeps flatterig, it is also ot easy to be predicted. Nevertheless, it could be realized that i the preset time a ew frotier for this research area is comig up i which fuzzy theory is beig associated with itelliget agets to build ew sorts of autoomous systems [][6][][3]. Itelliget agets ca be thought as decisio-makig uits that are capable of gettig iformatio (perceptio or measuremet), reasoig about what must be doe (judgmet ad coclusio) ad actig upo the exteral eviromet to reach a desired task (cotrol fuctios). Although there are differet iterpretatios about what a aget is, there is a cosesus that they have to possess special skills, such as, beig autoomous (i.e., idepedet), reactive (i.e., respodig to evets), pro-active (i.e., iitiatig actios of their ow volitio), ad social (i.e., commuicative). Sometimes stroger features are added (beliefs, desired, itetios) givig itetio otios for agets. Despite the fact that a aget exhibits the iterestig ability to represet itelliget behavior, oe issue that comes up is how the agets represet their itelligece sice itelliget systems typically require some form of kowledge represetatio ad the kowledge has to come from somewhere. A approach for kowledge based systems is the rule-based system for represetig ad processig kowledge i terms of rules, that is, a kowledge base i the form of IF-THEN rules, a database, ad a iferece mechaism for reasoig [9]. A special form of kowledge base system, able of reasoig by usig rules i a eviromet with ucertaity, is the fuzzy systems. This approach icorporates the expertise of humas ad allows the approximatio of huma reasoig. Sice a fuzzy system emulates huma reasoig, it ca be applied i plaig, schedulig, ad cotrol problems. This approach makes use of a set of membership fuctios ad a set of rules for the purpose of represetig the kowledge. As metioed, a aget must be able to perceive the eviromet, make decisio, represet sesed data, acquire kowledge, ifer rules, ad, fially, iteract with the eviromet. Whe a aget iteracts with the eviromet it may act directly upo the world as ay cotrol system or idirectly iflueces aother mechaism to act upo the world. This defiitio of a aget is so far almost idetical with that of cotrol systems as ay cotroller must sese a system to be cotrolled, represet the data, maipulate the iformatio ad determie a actio to modify part of a eviromet or uiverse. Thus, i this paper, cotrol systems ad agets are cosidered to be iterchageable cocepts. Additioally, the Fuzzy Referece Gai-Schedulig (FRGS) approach ad goal-based Itelliget Agets (IA) are show also as iterchageable cocepts, which are able to deal with decisio makig i complex problems. I so doig that the

2 former oe will supply a framework for adaptive behavior for the latter oe. The FRGS approach is formally preseted i [4] ad it was applied as a efficiet set of autoomous actios to supervise ad maitai safe a idustrial system [3]. Although it has bee preseted as a sythesis to cotrol complex problems it ca also be employed i decisiomakig tasks [2]. The FRGS methodology itroduces adaptive decisio ad cotrol performace ito autoomous systems whe there are chages i the objective or goal (referece) to be achieved. Exteded to itelliget agets, this approach allows adaptive behavior accordig to goals, itetios, desires, or beliefs. I so doig, the FRGS approach may be modeled as a particular class of goal drive agets that employ fuzzy logic rule-based systems. II. FUZZY CONTROLLER, FUZZY DECISION AND INTELLIGENT AGENT Like a cotrol system, a aget may be uderstood as if it perceives its eviromet through sesors ad act upo that eviromet through effectors (actuators). Nevertheless, while a itelliget aget is the oe that employs reasoig to fid out the appropriate actio to solve a problem, it is ot every cotrol approach that uses reasoig to fid out the actio to cotrol a system. The cotrol approach that icorporates reasoig is the fuzzy cotroller. I a glimpse, fuzzy cotrol theory ad itelliget agets ca be uderstood as presetig similar behavior sice they have the same properties (TABLE ). Aother correspodece betwee agets ad fuzzy cotrollers is that the former oe is itself a static oliear mappig betwee its iputs ad outputs. It maps iput to output through codig (fuzzificatio) iferece mechaism ad decodig (defuzzificatio) processes. Like fuzzy cotrol systems, itelliget agets are programs that map a possible percept sequece ito appropriate actios accordig to a objective to be achieved. Fuzzy cotrol systems are based o the fuzzy theory, which uses the cocept of fuzzy sets associate with the vague ad imprecise iformatio, the compositioal rule of iferece that is used to reaso (make iferece) ad to make decisios associated with the fuzzy logic. The iferece mechaism is based o a set of rules of the kid IF <coditio> THEN <coclusio> that cotais iformatio about the relatioships betwee the iputs ad outputs. Oe advatage of this approach is that fuzzy cotrollers represet ad maipulate iformatio i the same way huma beigs reaso whe dealig with ucertai or vague iformatio. Further, if a learig mechaism is adopted, fuzzy cotrol systems permit the adjustmet of the membership fuctios related to liguistic variables as well as the adjustmet of the set of coditioal-actio rules. This characteristic allows them to lear from a set of iput ad output ad behave o its ow. Artificial eural-etwork, geetic algorithm, simulated aealig are some examples of learig mechaisms applied to tue fuzzy cotrol systems [][4]. Thus, beig able of emulatig the huma reasoig process fuzzy cotrollers may also be cosidered to be idepedet as presetig a certai degree of autoomy. I [9] ad [2], it is show that by usig the compositioal rule of iferece, it is possible to treat the fuzzy cotrol approach ad the decisio-makig process similarly. Fuzzy decisio-makig ad fuzzy cotrol together ca also be foud i the literature [][7][2][22]. I additio, fuzzy decisio-makig ad fuzzy cotrol process ca also be viewed ito the cotext of aget theory i order to make decisios. Fuzzy decisio applied together with fuzzy cotrol ad agets are preseted i [8][2]. I order to carry o the selectio process of decidig which the appropriate actio is, a aget must satisfy a TABLE SIMILARITIES BETWEEN FUZZY CONTROL AND INTELLIGENT AGENT Characteristics sese the eviromet to decide about what should be the appropriate actio to chage de eviromet; act upo the eviromet by usig effectors (actuator); map the dates from iput to output; decide upo equivalet requiremets (objective, costraits etc.) are ratioal sice they decide about what to do by usig reasoig i this case a set of IF-THEN rules; are reactive sice they perceive the eviromet, decide, ad respod to chages that occur i it; are pro-active because fuzzy cotrol systems are able to decide based o referece (goaldirected); are autoomous, whe associated with ay mechaism of optimizatio to adjust its parameters ad keep cotrol over their actios ad iteral state, idepedetly of the huma beig; preset social ability i the case of beig iserted ito distributed artificial itelligece or distributed cotrol systems or whe i cooperatio with humas, or other agets i order to achieve their tasks. Fuzzy Cotrol Itelliget aget FRGS Cotrol X X

3 performace idex (measure) or goal. I a fuzzy eviromet, the decisio-makig process uses the cocept of fuzzy goal, fuzzy restrictio ad fuzzy decisio i such a way that the goal ad restrictio are fuzzy sets i the same space of alterative [5]. This sort of decisio-makig process may be studied from differet perspectives; however, all of them have i commo the advatage that fuzzy goals may be derived from a give performace idex. Itelliget agets ad fuzzy theory applied together to make decisio may be foud i the literature [][][2][5][7][9]. I the cotext exposed here, all those characteristics make the fuzzy cotrol approach, as well as the decisio-makig process, ad itelliget aget iterchageable cocepts that are able to deal with decisio-makig. Sice the reasoig employed by fuzzy systems are based o a set of IF-THEN rules (also called situatio-actio rules), the structure of this aget ca be classified as a reflexive aget [6]. Aother characteristic preset i both aget ad fuzzy cotrol system is the ability to represet the iteral state. The iteral state represetatio is guarateed because fuzzy cotrol systems ca deal with preset ad past iformatio about the eviromet to be maipulated (for istace, error ad error variatio). Thus, from this perspective, a fuzzy cotroller is also a reflexive aget with iteral state. a b Perfil Idividual Decisão c 3 Fução Distribuição de Possiblidade Percepção ou Medida a 2b Perfil Idividual Decisão 2c 3 Fução Distribuição de Possiblidade Percepção ou Medida d 2d III. FRGS AGENTS: A NEW FRAMEWORK FOR INTELLIGENT AGENT The drawback of traditioal fuzzy decisio approach is that there is o adaptatio, whe there are chages i the objective (goal) to be achieved. Objective here may also be iterpreted as desires, itetios, pla (goals), or beliefs, i the belief-desire-itetio (BDI) aget paradigm [7] of the class of procedural reasoig system (PRS) aget architecture. Oe way to suppress this disadvatage ad itroduce adaptive performace ito the system is to use the Fuzzy Referece Gai Schedulig (FRGS) approach applied both i cotrol systems as well as i decisio-makig processes. That is possible because of the iheret ability of FRGS systems to chage their membership fuctios (parameters) accordig to differet operatioal coditios represeted by the goals, or ay exogeous parameter. This approach may chage its coditio or coclusio as well as the rules i accordace with chages i exteral iputs. f e 2e 2f A. FRGS Cotrol Approach As it was metioed durig the itroductio, the fuzzy referece gai schedulig approach has formally bee preseted i [3] ad [4] as a adaptive cotrol approach. The adaptive mechaism is carried out by the selectio of several operatig poits determied by exteral referece state trajectories that modify the support ad core of the membership fuctios of fuzzy cotrollers (Fig. ad Fig. 2). Fig. Example Fig. 2 - Example 2 Membership fuctios ad fired area of cotrol actio. I so doig, the cotrol surfaces will chage as required by operatioal coditios, determied by the referece. Differet from the scalig factor method, the FRGS method permits the mixig of costat ad adaptive fuzzy sets (liguistic terms) o-lie. It permits parameters to chage homogeeously, as a scalig factor, or allows some parameters to modify

4 idepedetly or eve stay costat. This adaptive mechaism may be used i the codig (fuzzificatio) or i the decodig (defuzzificatio) processes, as well as i the IF- THEN rules, ad with differet ifereces schemes (e.g. Mamdai, Sugeo, Tsukamoto etc.). Like ay other fuzzy cotroller, it icorporates the expertise of huma beigs acquired i past experieces to figure out a approach to cotrol/supervise those processes that, for istace, are ot liear ad whose dyamics chage with time accordig to operatioal coditios, ad/or preset time-delay. Due to its ability to deal with huma reasoig, FRGS ca be exteded to decisio-makig task [2]. B. FRGS Decisio Approach The use of fuzzy sets i decisio-makig systems has become a field of great iterest, sice Zadeh ad Bellma itroduced the fudametal elemets about decisio-makig uder fuzziess [5]. A decisio i a fuzzy eviromet was defied as the itersectio betwee fuzzy costraits ad a fuzzy goal. Thus, if there is a fuzzy goal, G, ad a fuzzy costrait, C, i a space of alteratives X, the G ad C combie to form a decisio, D: D = G I C. () Whe FRGS is exteded to fuzzy decisio-makig, the kowledge about the goal (costrait) liked with the exteral iput is icluded ito the decisio-makig tasks. Thus, if the CONSTRAINT DECISION (R(.)) GOAL(K,R) GOAL(K,R2) GOAL(K,R3) Fig. 3 - Adaptive fuzzy decisio-makig (differet fuctios for goals) CONSTRAINT DECISION (R(.)) GOAL(K,R(.)) GOAL(K,R2(.)) GOAL(K,R3(.)) Fig. 4 - Adaptive fuzzy decisio-makig (same fuctio, differet parameters). goal chages, decisios eed also to be modified from these exteral sources, as it follows: [ ( r) ] I C[ K( r) ] D = G K. (2) The kowledge, K, built ito the set of rules, is altered, as there is chages i the referece, r. I this way, the referece is part of the goal (costrait) ad is associated with the exteral sources preset durig the decisio-makig process. Thus, i the same way that the referece composes the goal (costrait) that modifies membership fuctios i FRGS approach, the exteral factor will also alter membership fuctios i decisio-makig process. Fig. 3 ad Fig. 4 depicted the ability of the FRGS approach to adapt the membership fuctios ad thus decidig i a adaptive maer, whe a exteral fact (goal) chages. C. FRGS Approach as Itelliget Agets: FRGS Aget Sice the FRGS approach is itself a fuzzy cotrol, it presets a equivalet behavior of itelliget aget as it has bee adopted i this paper. However, because of the advatage of the FRGS cotrol or decisio-makig process to adapt as fuctio of the goal, it gives the ability to a aget to chage its behavior i a adaptive maer. This approach gives a framework for a adaptive aget. Thus, if the FRGS approach is employed, fuzzy systems ca ot oly be classified as reflexive agets with iteral state, but also they are a kid of adaptive goal-based aget Further, the FRGS approach is govered by the goal almost i the same way that the classical goal-based aget. Therefore, because the FRGS method chages its behavior (reasoig) to reach a desired task i accordace with the goal, this approach may also be sorted i the class of goalbased aget [6]. The oly differece betwee them is that while i the latter oe the goal gover directly the actio to be used to chage the eviromet, i the procedure proposed here the goal acts first upo the coditio rules ad membership fuctios, ad the upo the eviromet. Thus, the FRGS approach is a kid of adaptive goal-based reflexive aget. It is a framework for adaptive referece-drive aget or, simply, a FRGS aget. The FRGS aget is depicted i Fig. 5. The figure was built i accordace with the reflex aget ad goal-based aget preseted i [6]. This ew framework for itelliget agets icorporates the expertise of humas; allows the imitatio of huma reasoig by usig approximate reasoig; ad adapts its behavior accordig to chages i the goal, as well as desire, itetio, or belief if they are faced as exogeous iputs or parameters. The FRGS aget supplies flexibility i decisio ad cotrol actio because it makes available adaptive actios to accomplish the aget s objective, eve if it chages with time.

5 Reflex Aget State Dyamic Effect Coditio Rules Sesor Curret state Impact of Decisio Actio effector E v i r o m e t Goal-based Aget State Dyamic Effect Goal Sesor Curret state Impact of Decisio Actio effector E v i r o m e t Sesor Goal FRGS Aget State Dyamic Effect Coditio Rules & Membership Fucitos Curret state Impact of Decisio Actio effector E v i r o m e t Fig. 5 - FRGS Aget This flexibility allows the FRGS techique to reaso about the problem by cosiderig the future actios to be performed. D. FRGS Aget: A represetative example Whe applied i cotrol or decisio-makig problems the FRGS aget seems to allow emulatig the adaptive huma thikig. I this case, a itelliget aget would be desiged to substitute a huma i a cotrol or decisio-makig process maily if it is dealig with complex eviromet (systems). The coditioal syllogism embedded i the huma reasoig described by the method proposed here may be exemplified as it follows: If error = Small If error = Medium If error = Large THEN u = Small; THEN u = Medium; THEN u = Large, This set of IF-THEN rules represets the kowledge or expertise acquired by ay idividual. This represetatio is depicted i both Fig. ad Fig. 2. Although very similar i appearace, graphics (a) ad (2a) reveal a tiy differece itroduced by small chages i the core ad support of the membership fuctios. As proposed i this paper, these chages are obtaied whe exogeous iputs or parameters supply additioal iformatio related to ew goals to be achieved (for istace, r). Graphics (b) ad (2b) represet a perceptio of the eviromet (or measuremet), by whichever device, related to iformatio which will defie the actual behavior of the system ad that will affect the aget s fial decisio. The emulatio of the huma thikig performed by the aget is (3) accomplished i sequece by the graphics (c) ad (2c), (d) ad (2d), ad (e) ad (2e) for each example give. They respectively correspod to the cylidrical extesio applied to the perceptio (measuremet) over the kowledge base, cojuctio priciple, ad projectio priciple. Fially, the areas i the graphics (f) reflect the resultat projectio of the compositioal rule of iferece cocered with the fial decisio of the itelliget aget. Those areas are differet sice there are modificatios of the membership fuctios that embody the kowledge base. Because the membership fuctios chage cosiderig the goal i this case, the fial decisio that determies the behavior of the aget over the eviromet is a adaptive goal-based decisio. I so doig, this sort of decisio fits the structure proposed i this paper for a adaptive goal-based aget called here as FRGS aget. IV. CONCLUSION A emergig adaptive ad goal-drive based o fuzzy referece gai-schedulig (FRGS) approach is preseted i this paper. This kid of goal-based reflexive aget seems to be able to mimic the paradigms ad mechaisms related to adaptive huma decisios. The FRGS methodology itroduces adaptive cotrol performace ito autoomous systems, whe there are chages i the objective (referece) to be achieved. The use of this cocept i fuzzy decisio-makig process demostrated its potetial to cope with complex problem. Exteded to itelliget agets this methodology becomes a FRGS aget that permits adaptive behavior accordig to goals, itetios, desires, or beliefs. It is show that the FRGS aget icorporates the mai

6 characteristics foud i traditioal itelliget agets. The proposed architecture for the aget is pro-active, autoomous, ad reactive ad, besides, it may preset social ability. Fially, the FRGS aget is a itelliget system able to emulate the huma behavior by adaptig the decisio-makig process accordig to the goal to be achieved i eviromets that preset ucertaity, i the presece of vagueess, ad/or whe there is imprecise iformatio. ACKNOWLEDGEMENT J.E.A.F. ackowledges support from the Brazilia research fudig agecy (CNPq) uder grat /97 ad the Brazilia Natioal Space Research Istitute. REFERENCES [] R.A. Aliev, B. Fazlollahi, R.M., Vahidov, Soft Computig based multi-aget marketig decisio support system, Joural of Itelliget Fuzzy Systems, v.9,.-2, pp.-9, 2. [2] J.E. Araujo Filho, K.H. Kieitz, Adaptive referecedrive decisio-makig process, Proc Iter. Cof. o Fuzzy Systems (FUZZ-IEEE),v., pp , St. Louis, 23. [3] J.E. Araujo Filho, S.A. Sadri, E.E.N. Macau, "A New Class of Adaptive Fuzzy Cotrol System applied i Idustrial Thermal Vacuum Process", Proc 8th IEEE Iter. Cof. o Emergig Techologies ad Factory Automatio (ETFA-IEEE), Frace, 2, v., pp [4] J.E. Araujo Filho, S.A. Sadri, E.E.N. Macau, "Fuzzy Gai Schedulig Cotrol Systems", I: Proc.: The 9th Iter. Meetig of the North America Fuzzy Iformatio Processig Society (NAFIPS), Atlata, EUA, July, 2, pp [5] R.E. Bellma, L.A. Zadeh, Decisio Makig i fuzzy eviromet, Maagemet Sciece, 97, v. 7, pp. B4-B264. [6] A. Boarii, F. Basso, Learig to compose fuzzy behavior for autoomous agets, Iter. Joural of Applied Reasoig, 997, v. 7, pp [7] M.E. Bratma, D.J. Israel, M.E. Pollack, Plas ad resource-bouded practical reasoig, Computatioal Itelligece, 998, v.4, pp [8] C.-H. Chag, Y. Che, Autoomous Itelliget Aget ad its Potetial Applicatios, Computers Id. Egg., 996, v.3,.-2, pp [9] C.W. de Silva, Itelliget Cotrol: Fuzzy Logic Applicatios, CRC Press Ic., 995. [] C.W. de Silva, Itelliget cotrol of robotic systems with applicatio i idustrial processes, Robotics ad Autoomous Systems, v.2,.3, pp , 997. [] M.S. El-Nasr, J.Ye, R.R. Ioerger, FLAME Fuzzy logic adaptive model of emotios, Automatic Aget ad Multiaget, v.3,.3, pp , 2. [2] B. Fazlollahi, R.M. Vahidov, R.A. Aliev, Multi-aget Distributed Itelliget System based o Fuzzy-Decisio Makig, Iter. Joural of Itelliget System, 2, v.5,.9, pp [3] A.F. Gomez-Skarmeta, H. Martiez Barbera, Fuzzy logic based itelliget agets for reactive avigatio i autoomous systems, I Proc. Fifth Iter. Cof. o Fuzzy Theory ad Techology, Raleigh, 997, pp [4] S. Guillaume, Desigig fuzzy iferece systems form data: A iterpretability-orieted review, IEEE Tras. Fuzzy Systems, 2, v.9,.3, pp [5] P.McCauley-Bell, Itelliget aget characterizatio ad ucertaity maagemet with fuzzy set theory: a tool to support early supplier itegratio, Joural of Itelliget Maufacturig, v.,.2, pp.35-47, 999. [6] S. Russel, P. Norvig, Artificial Itelligece- A Moder Approach, New Jersey: Pretice Hall Ed., 995. [7] M. Tarrazo, L.Gutierrez, Ecoomic expectatios, fuzzy sets ad fiacial plaig, Europea Joural of Operatioal Research, v.26,., pp.89-5, 2. [8] A. Tuma, H.J. Muller, Usig Fuzzy-directed Agets for Ecological Productio Cotrol, Itelliget Automatio ad Soft Computig, 2, v.6,.3, pp [9] R.R. Yager, Pealizig Strategic Preferece Maipulatio i Multi-aget Decisio Makig, IEEE Tras. Fuzzy Systems, 2, v.9,.3, pp [2] R.R. Yager, Fuzzy Sets ad Approximate Reasoig i Decisio ad Cotrol, Proc. IEEE, v.9,.3, 2. [2] W.R. Zhag, Nestig, safety, layerig ad autoomy: A reorgaizable multiaget cerebellar architecture for itelliget cotrol with applicatio i legged robot locomotio ad gymastics, IEEE Tras. Systems, Ma ad Cyberetics Part C, 998, v.28,.3, pp [22] Zimmerma, H.-J., Usig fuzzy sets i operatioal research, EJOR, 987, v.3, pp