Sie sind auf Seite 1von 5
171 ¥ Lean A os, "Anyi of eral meow edandany Neal Computation, v2, 226238 1980, Is) C2 Me Sod and Nog. Maiaiy fu ran ua ‘eons ind monocar propria.” im Pr IJ Conf ot earl Net, San Diego" CA, Jone 1980, 9-83-96 ils HotinsrAm St ad Space “Lis to ne itpranee of fnew eo tone hearing Ine Pw 200 TEBE Fal Toe Comping Sp 990. 1. Pare, "Laag topic” Tech Rep Te Cer for Com tna tens Magee SMT Cor 1S. Pratik and 1 Koen “A say off wlerace popes of ical aera net"Toch Rep. Eke and Copy” Dep. Un oT Mascot Ame, 191 $F Pergo and FG "Netorks for sprximaion aed lung Pow BEE. vo. Pp. 148-1097, Sep SD DE, Rormihan and. McClean, Ek, Paral! Disb Pro evi tips ihe Mc of Coon Ca 114) BE Ramet, “Lenning and genera,” amr f pene tess ape la Pr IEEE ICRN San Dio, CA, 9 lis) BLE Soe and Med. Caney, “Potion of praned aay us) io, we ws swan Pr END, Seale, WAL 109, pp. I-82. R'E" Sige and M.. Caner "Fou seni and nodal relevance ‘sianshpe nye pereprons Tech Rep ECEIS. 0002, Dep. Fiend Comput Eng Unt of New Harpe, Dram, NL Sha 190. 107) CH Seguin and RD, Clay, "at tlre ie atic geal fetta in Pr ONY, Sn Deg, CA Joe 198, pp. F 708-18, FL Wats, earn lpr er conetons tos Api [aint eth falar apmuton ta Pro IEE lt. Con Nowra ner San Diego, CA, 1087.0 p. 610-627 1" ero yond egsnan Kw ei or pect and aa Ditto Kova sazneee PLD, Saceraton, Hanada tis, Fuzzy Systems as Universal Approximators PP Bar Kosko Absrct—An adv frzy sytem cam uiformly appeosinat any ‘al comin fanton on compact domain any degree of acura. ‘kn addlive wey sytem approsinatr the func by covering Ws {raph with fsy pate nthe npe-otpu late Space and nero Pte that nertp. The ary system computes a condtoal expectation [EY [|i ew he ey set ado Set ach fey role eines 2 aaey patch and connects commonsense Knowedpe with ate space fect. Neural o statistical cutering stems can approimate the ‘own funy patches rom Waning data. hese apie Tory ses Spprosnatefncton att eves At the lel ee! he mera stn prosiats ad anesthe fesy rales. At the global eve he eles ot pete apprximete the fonction. 1. Fuzzy APPROXIMATION ASA FUZZY COVERING A furzy system approximates fanetion by covering its graph with fuzzy patches and averaging patches that overlap. The appreximation Mans resi 6, 192: eed August 10 199. Earer versions oie won appear in Paceding of 192 IEEE Inertia! Confer ‘n Pacy Spits (P2299) ane Proceedings of the CAN 92 he tr whe: Dene of a Ensign DMeCineck Avon, EE 400, Lov Angi, CA 9003-968 USA, nal improves asthe fuzzy patches grow in number and shrink in sie Fig. shows how fury patches in the fapatoumput product space X'ic ¥ cover the rea function f : X —~ Yn Fig. 1a) a few large patches approximate f. In Fig. 10) several smaller patches tester approximate f. The approximation improves 38 we add more strll patches bt sage and complenty costs inrease. This bref ‘onsbuton gives the algebra duals ofthe fzzy approximation 1 farzy system isa set of ier fuzzy fules that maps inputs 10 covtpts, Fach fury rule defines furry patch in the input-ouput Site space ofthe function, Fig. 2 shows the fuzzy re “if X is Negative Smal, then is Positive Small” asthe eatesin product NS x PS of fury" [12] oF mulalued sets NS and PS. A 2D plot would show the fuzzy patch 'S x PS asa barn tke sucre that rises up fom its rectangular base. Each input belongs to some epic to each inp ery se. So each input es all the fuzzy rules to some degre. Exper state the fuzzy rules or neal or statistical ‘stem lars them from sample dats. Exper andalgrthms can give ‘ferent sets of fuzzy rules and so pve eiflerent approximation of| the fection. ‘Next, we show that azz system can approximate any continuous real function defied oa a compact (closed and bounded in R") Aomain and show tha even a bivalent expert sytem can uniformly ppeoximate a bounded measurable function. The fuzzy ystems have 2 Teeorwatd architecture that resembles the feedforward multilayer eur syslems wsed approximate function [4]. The uniform Spprosimaton of continuous functions allows us co replace eich Continuous fuzzy set witha Brite discretization o¢ 2 point in @ unit hyperebe [7] of “ray cube” of high dimension omik and Whit (6) and thes have used the Stone—Weierstass ‘theorem of functional analysis (9) 10 show uniform convergence of earl networks, The Stone-Wesrstas theorem states hat if CL) Fy the sup-aorn space of coninvous functions on a compact and Hausdorf Nand A.C C(N) isa closed algebea and if is sel ‘jin and separates pins and contains the constant functions, then EV ZCAX, Ths ives the result but not much insight ato Bow to ‘build or lear real ystems. Radial basis nets sum Gaussian fanetons An! also siformly approximate continuous funetons on compact sets [3]. Additive furry systems [7] with Gaussian fuzzy sets (11), Sefine radial basis nets and 30 aso act as uniform approximates. “The theorem below also proves so direy since ie olds forall dive systems. The consrutive proof below shows BOW t use ‘eur systems to Team rles and how to Ft the rules or patches ‘hange with time to wack & nonstationary Fupctin, 1, Apprive Fuzzy Systems Ips fre the if-pa A, ofall furry res EN = Ay, then Y=" and give scaled sam 1 a output, Earlier uray systems combined the outpa Tur sets with prrese maximum in acord vith the so-called “extension peinciple™ [2], Additive furzy systems Sum the cuppts [7] 28 in Fig. 3 p= Son, 7 For now we take the adaptation weights wa unity: a Different limit theorems show how the diferent Combination schemes tehave, Sum combination often tends to pve © symmeuie timed disribtion a the output fuzzy set. We gives the global ‘utp centri in) asthe simple convex sum of set errs (6). one 9 ors0s00 © 1996 IEEE IEEE TRANSACTIONS 4 COMPUTERS, VOL NO. 1, NOVEMBER 198 x Fe (0 Far tage fazzy pats cover par of th gp ft know fnton f= X x ® “+. Fener paces eres compotion and erese ‘rca ce.) More smaller fy pts fer caver fot teste Semanal cot Each fy le ee tein pret ace N'Y ina lage bt fate nme of fary res or pose mcs Gove he ph wiley cewacy. Opal ales ene sem af I X=NS, then Y=PS. Fg. 2. Funy least space pth or crsan past offre en. The pet gah VS PS Mand her te Wary tle Xe Negus ‘Small ten Ys Pnve Shall" Neve tapes and anges ene fry rid ace Combining the scaled or “ied” consequent ezzy sts By. By in Fig. 3 with pairwise maximum gives the enselope of the fuzzy Sets and tends toads the uniform distribution (5) In general the Borel-Cantlt Lemma of probability theory implies tht the “ex tension peiniple™[2} of fuezy theory extemizes fuzzy membership ‘values iat binary endpoint valves: Himsup, et! Aa? Ao at 4 2 which holds with probability ome [5] fr nondegenerate pairwise Independent sequences of, random variables tat take values in (2.3) The symbol “A” stands for pairwise minimum, Soin general the extension principe does not “exten” a fuzzy set at all. I dafucites it imo a inary distibuon. In practice m= 1 = 6 in fuzzy systems. As the number of nonzero fry rle outputs grows, the envelope lealy grows to the constant valve 1, the uniform Aisribtion. Globally the oupuc grows toa rectangular ple “Max combination ignores overap in the fuzzy sets B. Sum combination adds overlp to the peakedness of B. When the input changes sighly, the adcive ouput B changes slighly. The mas ‘combined outpst may ignore smal input changes since for lage se of rules most change occurs inthe overlap regions ofthe fy sts [fay ekg }-— 82 Lexical” | Fe.3,_Aadive fy system whic. Real number defines i ‘ech oD dea pe that uses tough ie FAM se ab eh Thay osme dere (most eo dee). Te stem as the ced nip fry sts ge B New ower apie ses cn nail the ‘ie mpi (ort sets 4, dB, ) with sample dun Th ced 5B piven the ep sums yy b's cen sum fom oto Bi, Here the overlap problem arses sce the centro tends to stay the same for smal changes in input. Bute ceroid smoothly tacks changes in the fuzzy-ct sum (1) a8 (S) shows, ‘Cento defzztcaton compounds the max problem. An it: rable mabivalved set funcion mg: Y"-— [I] gives the Eenueidl out y, o¢ Fess Ei uma tvhdy Te ma widy ‘We can seplace the ner in (3) with smal rete sums indeed only bythe number of fizzy ses that quantize the fuzzy variables U6) See) below. This eliminates both the need 0 approximate the ‘emoid and its computational buen. Digital VLSI harcwae can then implement (3) or an analog VLST chip can implement with 3 simple fllower-agregator circuit (8). In the limit asthe umber of combined ouput set grows, cenoialdefurzifistion may tend to coineide with mode defuzzticaton since the cenoid and mode ccie fra symmetie unimodal dsb, » o Im, Fuzzy FUNCTION APPROXIMATION Sum inference reduces tothe weighted um nS own, ® Fg. 4. Sum vs. max combinin of male cup es. The max envelope (2) ends toward a rstngle a the sumber of apezoids pows, The se vckpe (ytd tmacds Gauan oye nid cre. The ‘Sof te um emelpe i (wast al at Rewer he cents of ‘tet and tr tnpezoid The spproxiation theorem depends on hist (Cen of ax encopes cat fal ote ese bounds hese a i the degre 1 which input x, belongs to fuzzy set Ay in the ule o patch A, x 2. Additive fzzy systems can approximate ‘function J = X— ¥ by shrinking these patches im size and Increasing them ia number (7), For the proof we ean assume no Teaming and equicredible es: wy =o" = iq = 1 ‘Equation (4) gives a simple approximation theorem. Suppose 4: X-—+ Y is measurable and bonded. Thea we can view B a a simple fonction ifthe fray sets, ze nofuzry ses 7], We replace triangles, trapezoids, and other fuzzy sets with rectangles. Simple functions x = X'—> ¥ map "ino initely many valves of YA simple fancton equals finite sum of weighed indicator functions. Choote thet vals! a the weighs an et the noafuzzy sets or rectangles partion X. Then (10] there i simple function {-epalln close 10 f on X. Boundedness of f ensures a uniform ‘pprosimation. This shows that a larg enough AI exper system can Spproximate any bounded measurable Function and reminds us that fuzzy rules role im the Divalent case to exper-sytems rues. It oes not show that an additive fuzy system with multivalued sets ‘converges wnifrmly to f- SO in practice the result gives no eror| bound. The problem aries because tangles, wapezoids, and other fuzzy sets need not converge uniformly to rectangles. ‘Uniform convergence hols if we work with conti instead of measurably. The theorem below requires that f =X — ¥ is ‘Continuous sn that Xie compact closed and bounded) n°. The theorem shows that in principle an adtve Furry system with Bite furzy roles can approximate any continuous funcion to any degree of sccoracy. It nels the Gaussian result in [1] a8 a special ease Theorem: An addve fuzzy system F uiflorly approximates {22 VEX is compact ant f 18 continuous Proof, Pick any small constant = > 0. We must show that F(2)~ fla] <= forall» © X.X isa compact subset of FP (2) i the centroid output (3) or (6) ofthe ative fury sytem Fin CContiuity off on compact X gives uniform continuity. So there is fixed stance 8 such tha foal and si X, |f(2)=f43I] 2/4 it fe 2] c #, We can consnict st of open cubes Mf o"- Mx that cover and that have ordered overlap in heir n coordinates So that each cube comer lies a th point of its neighbors M,.Pck Symmetric output fury sels 2, centered on f(c,) So the centroid of B, i fle) Pick w & X. Then by consrcton w ies in at most 2" overlapping open cubes Mf. Pick any win the same set of cubes. Wu € 3, and w € My, thea for all v © M/A Ma : lu ~ ol <6 and [r= wl < 4 Uniform continuity implies that [f(u) ~ f(~)) < Lea) ~ fo) + [fCe) ~ Few) < £/2. 80 foe cube centers, and calle) He)] < 2 ick # € X. Then 2 too les in at most 2° open cubes sith seers ¢) and [ftc) ~ flz)| < ¢/2. Along the kth coordinate ofthe range space J” the kh component of the additive sytem ‘centroid F(x) lies asin (6) on or Between the th components of| the centroids ofthe By se. So, since [f(c))~ f(ce)] < 2 for ll, Sej)lFC2)~fles) < 2/2, Then|FL2)~H2)] $ Fle) fle It fle) = Ha] < /2+ ef o “The proof may require symmetric ouput funy sets By if corlation minimum encodes [6] the rules A, x B,. Coreaton Product encoding does not require symmety since 0,D) bas the fame centroid a8 By fa; > 0 “The proof shows that we can replace the fury sets A, and B, with ite discretizations or Bt vectors (a, >a) and (H+) “The discrete version of B, must have a centid at or closet the centioid of B,. So we can aways work with large-dimensiona unit Iypereubes and view fuzzy rules or patches as matrix mappings (or {faczy associative memories (7) between hypereubes o a pont in ‘ven larger hypercubes ‘The proof fils for max-combined sets B. The proof traps the centoidal output C(B) of yn (3) Between the centroids C(B:) and Clq) HDs) < CUB) S--- < CIB) : (By) C(B,) cw) a o Yee) o for volume or ar A(B,) = fx maleid2 and for conver area cece cy" ne Wang [11] renames these tems “fuzzy basis Tintions” The proot works for any combined ouput set B= (6(Bi."»Bm) Sch that C(Bs) <6 < C(Br). In general the tmax combination VJ", By does not obey O(B1) < CIV, B) < (Bn). Tis inequity holds ithe val case when sum = max. Since ¢ + y = min(xy) + max(ryy)2 + y = max(ry) if 2 ory = Off the combined sets Bh», By. are disjoint. The proof also works for noncentoidal defunziess D(B) that obey CB) < D(B) < CBr), In general the supremum oF mas membership defuzziter doesnot obey ths inequlty 1. Fi2zy S¥sTEMS As ConpmoNAl. EXPECTATIONS How “fu2y" isthe approximation theorern? How “fuzzy” is = fuzzy system? Pure fuzines stems from the overlap of thing or set A and its opposite a" is farzy iff ANAS # O [7] Fuzzy systems need not be fy inthis pur Sense. The finite area A(B) fof the output et means tht A(B) normalizes B wo give Bas te ondional probably deity py | 2): sar an o =n(Y |X =2) ® ‘We can view the furzy sets Ay and B, as random sets oa lost co ywo point conditional probability densities. The set degre m (2) tepals PAIX 2) the probability of event A given that random ‘rarable X takes onthe domain or index value + So we can view the Bt vale mpar(2) asthe probability that X is Positive Medium PN ir Then X = PAT mans the random variable X takes on the enue random set Postive Mesivm asa random‘ value. Then far input each rue fies with some condition probably. The system emits each ouput y with some conditional probability. The ‘ouput equals the oe average or condtonal mean f the output is the maximum probability vale, the system computes a maxim a poseiod oe MAP estimate "The centroid (3) gives the same result as (8) and implies thatthe fury system ourpat P(r) equals 4 realization of the conditional expectation: Fu) = BI |X o * T ye ae z (On this view a fuzzy system is probability stem. computes the random vaeisble EY | X] or E[Y | X= NEN M,---, PL) “The summed putes give a model fee estimate of f:X —> ¥. ‘This conditional mean is the mean-squared optima among all nonlinear estimates off that depend onthe input "set" or destes NE.NAf---.PL. The approximation theorem says that additive faery systems form an bande around f. They form an =-bundle of conditional means around f V. ADAPave Buzzy SYSTEMS ‘An adopive fuzzy system is a fuzzy system that changes with time. The sets or rules change in shape of number. A leaning 5stem changes the fuzzy ule Weights Wye, a6 it samples input utp data (291). 92) I peace (7), we deshold the Weights, 100 or I the weigh w, equals or exceds the esol, ‘try = 1 anda the fuzzy ule the fury system or “knowledge ‘base fw; falls below the threshold, x, ~ 0 and ignore the jth fuzzy rule aed do no ince icin he fuzey system, ‘Adaptive farzy systems estimate fiery rules from sample dats, “Tis reduces o patch or clser estimation in X > Y called product space claserng (7, Neaal or statistical clustering algorithms convert the sample data (2) im claster estimates. An expt or physical process ives the sample data. Clistring algorithms sare forthe Implicit fuzzy rules tht the expert or physial process "sed" 10 senerate the data ‘Vector quantization estimates clusters. A fixed set of quantization ‘vectors m;qo-»,my tacks the dstbuion of sumple data. In neural systems each quantization vector m, defines afanin synaptic vector to a neuron that competes in a winner-take all network. The neural stem leans or adapts if and only i the quanizatinlynaptic vector ‘my, moves in the inpuc-ouput state space X > Y. In effet each ‘quantization vector m, estimates Tocal cluster in X > Yad, in optimal meansquareé-error lsming, converges tothe chstr's xntroid exponentially quickly [6 Gobally the quantization vectors estimate the unknown jon probability density pry) that gives ise tothe observed data pars (r.y,)-Thev quantization vectors estimate the probability of any region C as ne /r he number of quaization ‘vector in C divided bythe total ue of quantization vector ‘Chsters of quantization vectors estimate fuzry pales, At any time nthe lering process each fuzzy patch -, % B, contains nn quamizaton vectors. This gives ise 1 an alspive histogram oF Trequency disiibuion of quantization vector in the rs overapping fuzzy patches reel In practice we may count a cell as sficieny ‘ecupied (1) = 1) if conthine any quanzation vectors. For Prout spac catering wth vector quantization (Sl dt show teehee sampled (b) Lager dats shows ow he quiz veces then "= By" grows ate A,B, pls coat a ana eters OVS in general sample data greatly cutnumbers the fxed quantization sectors. In the esteme cate, we can count each sample as a ‘quanraton vector {11} This unbounded eases does ot lier noise or ompres the sample da. Fi, S shows product space elustring wih ‘sector quantization after leaming has slowed or stopped. The small dots in Fig, Sta) ae the observed samples (,.9,). The large det in Fig 5b) are the quntization vectors m,, If we set a tresbold fof two quantization vectors pr cll then product space clustering yields 10 fuzzy rules or patches. Other learing Schemes can change the clserreions by arouping the covariance ellipsoids [1] of the ‘quantization vectors and has can change the shape of rules and st. The above uniform approximation theorem implies that Anite ‘umber of quantization vectors my,-s-vmy can Team any sampled comtinaous function ifthe learning stem samples enoogh fonction Samples (1, )) and if the quantization vectors converge 10 local ceo. Several types of competiive ling ensre this ‘convergene [6 In gener, this requires leaning with a prohibitively large number of quandzation vectors. Inthe small-sample case shows that we can sink the sppeopite fazy paces and increase their number as we simple moce dats and increase the number of | aapive quantization vectrs VI. Conetusion {A fay system or spproximator seduces to graph cover with local averaging. That i6 ot unique. An additive fuzzy system With ‘Gaussian sets redues to 4 rail basis network [3] and tht (006 tut one of many graph coverings. The “zziness” or malivalence of sets comes ito play when patches or ouput sts overlap, Noafuz7y Sets can also weigh or average the overlap. A fuzzy system is unique in hati Ges vague words tke "sll and “medium” othe math of éuves andi vectors (pins in nit cubes). So ties natural language snd commonsense rues t state-space geometry. But the “azz Sets are equivalent to random sels of lei of two point conditional probabilities, The fury systems ive a mode-re estimate of some Sinknown conditional expectation [|X]. The approximation power of fuzzy systems lies more in their model freedom than in thei fuzzy interpretation. Nonfuzzy sets and rule patches also lead {0 model fee universal approxima. ACKNOWLEDGMENT ‘The author thanks De. F. Watkins for insights on the analysis and application of fuzzy systems Rerenences 10) 1.4. Dickenan and B. Kosta, “Fuzzy Fanon positon wit ‘apes eligi leuing” Proc. Worl Canes om Newal New: (INNS WCRI’), vl Jay 199, pp. 8-17. Dubois aH Prd, Push Se and Stems Theory an Appin tna Orde FL Academic Pres, 980 {51 Eman. Kel an. Rows, "Layered eal aeons |e) ita Since, a H Whit, “Maliayee efor nt ‘ea we aie posits” Neral New 2, 99-36 15] B Kiso, Fuzzy knowledge combtaton.” Jnl. Sst fp 29820. oR 11 ee Sroka compeive leaning” IEEE Trans, Neral Ne: Gard, nos pp S22 a0 Set 101 (0) Nera Meoworks and Fay Sims: A Dynamical Stems Ap apie ai tigen. pws Nee Hh Meal Analog VS on New Sr Reg MA: An ae and Comper Anais, Seon, New Yok McGraw iim, ots Nang tM. Mend, "Fz bs fenton univer aposima- ionsanéoogna lagers arming’ EEE To Neat Nee Sol Sono 5 SDT sI4 Sep 982 (12) 1A Ze Banyo Cor, 8 pp 388-353, 1965 ‘Optimal Centralized Algorithms for Store-and-Forward Deadlock Avoidance J. Blazewier DP. Hovet J. Brzeidski, G. Gambos, ‘and M. Taleo Aburect—ta this brie contrib, protien of deadlock avoidance ta stead forward ceiver wi at fet two balers per tae Conard fred a wel sda outing For bat ese paloma limes entra each avoidance slots are propsed an shown {ee optinal in users of pose bate lean, When the mane tt butlers ie equal to oe for exch node the problem Ie knew fo be Ne-complet, thay nitty t admit» prema ine algo The ‘resented ress fay be ls Interesting for other appieations, some Imasivly paral computer sss eng one of he examples, Inde Trt—Stre-and forward wet works, dele sven, cen telined approwch, afer ullanten, comply als 1. twropuctios ‘The concep of stov-and-orward packe-swiching is commonly used in computer and telecommunication netwoeks a6 well sin some 7 Bltencz ad J. resi te with he siya nora Pie ‘ike Porat Paro 3, 60953 Porat Poland '. P. Bot is wiht igre 0 Scena el Inooazione, ‘Univers Roma “a Sap Via Sal 13, 198 Roma ay ‘6 Gao with ibe Univer &t Roma “Tr Vergo ea ‘Tmo wih he Diaries dno © Sitemitc, Univer si Ror “La Sapient’ Vn Salas 113, 0018 Roma ay massively pale systoms sich a hypercubes oF wanspuer based Inachines. One ofthe most important ses arising i these systems fs deaiock avoidance. To solve this problem several interesting Aistibuted algorithms have been aleady proposed [2}113] (15) [UG (20), (21. 281, (26), [27], bo ll now this area offers & reat ‘opportunity for improvement ((2)). Distibuted algothms can be ‘ther easily implemented in existing networks. However, its also ‘lear tht none of them is optimal ith respet to buffer utilization tnd to dhe numberof safe sates allowed. Thies Because they are 00 restive, since they cannot use the fll knowledge ofthe network ne which would be avaiable in centralized approach, Thus the Tater, optimal, can be useful 26 a valuable benchmark forall cxising trite algorithns. Moreoer, this optimal approach my tleo form a bae for constuction of new, more efficient, distributed ‘lzccithms, Sich an adsptation of centralized algoithms wo distributed ‘context as Been commonly ed im practic, and many interesting Solutions have been abained in his way (Gee 62 [1)(19}. 22D. ‘Atmore theoretical ean for studying centralized deadlock av0id- ance algorithms follows fom the fact tht a. sore-and-forward ‘network may be considered a. special case ofa centralized computer sytem, in which resources ae required and then released in & ‘respected order fr each pocess (mesage) in he sytem 14] Such {station fs not been investigated yet and the complexity of the ‘corresponding deadlock avoidance problem remued open, except fa networks containing nodes with de buffer. This ast problem has been proved fo be NP-complete {1 “Thea ofthe present paper follows from the shove lscusson and say be tte a ring an optimal (rom the point of view of btfer ization) and ecient (plyeomil time) cemalized algritn for oresind forward deadlock avoidance. In Section Tl, 2 model ofa Store and-forward network is presented, Section TIT presents cen ttlized store-and-forward dealock avoidance algocthm, assuming ‘ynamic routing and more than one buffer pe Bode In Secon IV, fixed outing is considered and coresposing revuls e pesented 1. THe Moet AA soreanforwant transmission network can be defined 38 suatiople SF = {V-E. BN. B), whore V= (osyoorsty) isa Se Of nodes, Eis a se of (bidirectional) links, BN isa Set of bes tnd BV os Ns 4 fonction such dat 2", Bios) =| BN | ‘which associates with each node the numberof bles it comtins. ‘5 sssumed tht al buffers have the same sizes and that each ber an be allocated to one message at atime Denote as Gap = (VE) the graph underlying network SF ‘Messages have tobe transmitted between contiguous nodes along owes Re fre pas (0,..0--4,) i Gap Let us denote as {sed rowing te case in which the route of a mesage is known in ‘vance let us moreover denotes domamic routing the case in Which for cach message, only te next ode along rou Is sven a any time. Message transmission is performed according tothe Following sssumpdions 1) Message originates in node on) the source of) if there exis at lest one fre ber in = am), te a baller no ieady assigned to any message. Such Buffer can be assigned With each message m comin in poe there is associated node mex) contiguoss to, representing the next node (0 ‘nhc as tbe transite.

Das könnte Ihnen auch gefallen