7. Fuzzy Logic We are in the process of discussing bow automated systems can deal with uncertainty. Tn the lase chapter we discussed a number of methods to do this, among others, probability theory, in particular Bayes’s method. Instead of using erisp numbers, we used probabilities. As we know. in the real world, there is no certainty. Thus i is bettzr to associate the probability of an event (say 2 disease) with a particular symptom (e.g. the result of a test). We saw that while Bayesian methods have their definite merits (they show us how we can deal with uncertainty in mathematically correct ways by including a priori probabilities), they also have their problems (some numbers are not available and hard to define, the real world changes). So, they do not represent the definite solution to deal with uncertainty in the real world. We brifly mentioned certainy factors. But the latter also have their problems when put t0 work in the real world, as the large number of failed export systems projects demonstrates. Many of these projects in fact employed certainty factors (oF something similar). So, it seams that we have not yet found the final solution. Many alternatives have been suggested. A prominent one is fuzzy logic: As we will see, one of the reasons why fuzzy logic is so popular is that it has a highly ‘appealing way to deal with the real world, Rather than tying to define how things “really are", fuzzy logic takes account of the fact that things ia the real world are not either this way or the other way, but most of the relevant properties are in fact gradual ones. Fuzzy logic has been very successful. It has been. and still is, especially popular in Japan, where fuzzy logic has been introduced into all types of consumer products with great determination. Nowadays, “Fuzzy”, in Japanese TFaS{ ‘has become something like a quality seal. In table 7.1 a number of applications of fuzzy logic are given (more applications can be found in Dubois et al. — the latter also includes theoretical papers) Table 7.1; Applications of fuzzy logic in Japan and Korea (fielded products) (1992) Based of Kosko, B. (1992). Fuzzy thinking. The new science of fuzzy logic. New York, NY. Hyperion. Additional examples have been included. Product ‘Company Fuzzy logic rote Air conditioner | Hitachi, Matsushita Prevents overshoot-undershoot Mitsubishi, Sharp temperature oscillation and consumes less on-off power Anti-lock brakes | Nissan Controls brakes in hazardous cases based fon car speed and acceleration and on wheel speed and acceleration ‘Auto engine | NOK/Nissan Controls fuel injection and ignition based ‘on throttle setting, oxygen content, cooling water temperature, RPM, fuel ‘volume, crank angle, knocking, and ‘manifold pressure ‘Auto tansmission| Honda, Nissan, Subaru Selects gear ratio based on engine load, diving style, and road conditions Chemical mixer | Fuji Electric ‘Mixes chemicals based on plant conditions Chapter 7 101Product Company Fuzzy logic role Copy machine Canon Adjusts drum voltage based on picture density, temperature, and humidity Cruise control Isuzu, Nissan, Mitsubishi Adjusts dhrotle setting co set speed based ‘on car speed and acceleration Dishwasher ‘Matsushita Adjusts cleaning cycle and rinse and wash stategies based on the number of dishes ‘and on the type and amount of food ‘encrusted on the dishes Dryer ‘Matsushita Converts load size, fabric type, and flow of hot air to drying times and strategies Elevator control Fujitec, Mitsubishi Electric Toshiba Reduces waiting time based on passenger waffic Factory control | Omron Schedules tasks and assembly line strategies Golf Maruman Golf Selects golf club based on gotfer’s diagnostic system physique and swing Health Omron Over 500 fuzzy rules track and evaluate an ‘management employee's health and fitness Humidifier Casio Adjusts moisture content to room conditions ron mill control | Nippon Steel Mixes inputs and sets temperatures and times Kiln control Mitsubishi Chemical Mixes cement Microwave oven Hitachi, Sanyo, Sharp, Toshiba Seis and tunes power and cooking strategy Palmtop computer Sony Recognizes handwritten Kanji characters Paper industry Cellulose do Caima, Portugal Pulp production Plasma etching | Mitsubishi Electric Sots etch time and strategy Refrigerator Sharp Sets deftosting and cooling times based on usage. A neural network learns the user's usage habits and tunes the fuzzy roles accordingly. Rice cooker ‘Matsushita, Sanyo ‘Sets cooking time and method based on steam, temperature, and rice volume Shower system ‘Matsushita (Panasonic) ‘Suppresses variations in water temperature 102 Chapter 7Product Company Fuzzy logie role Still camera Canon, Minolta Finds subject anywhere in frame, adjusts autofocus Stock trading | Yamaichi Manages portfolio of Japanese stocks based o: ‘macroeconomic and microeconomic data Subway control | LIFE Institute, Yokohama Controls the subway during peak hours, Television Goldstar (Korea), Hitachi, Adjusts screen color and texture for each Samsung (Korea), Sony frame and stabilizes Volume based on viewer's room location ‘Translator Epson Recognizes, translates word in penci Toaster Sony ‘Sets toasting time and heat strategy for each Dread type ‘Vacuum cleaner | Hitachi, Matsushita, Toshiba | Sets motor-suction strategy based on dust ‘quantity and flocr type Video camcorder | Canon, Sanyo ‘Adjusts autofocus and lighting Video camcorder | Matsushita (Panasonic) Cancels handheld jittering and adjusts autofocus Washing machine | Daewoo (Korea), Goldstar Adjusts washing strategy based on sensed dir] (Korea), Hitachi, Matsushita, | level, fabric type, load size, and water level. Samsung (Korea), Sanyo, Some models use neural networks to tne Sharp rules to user's tastes. ‘The goal of this chapter is to introduce fuzzy logic and to show how it can be applied to real-world problems. Moreover, we will put this technology into the broader context of real-world computing, We will proceed as follows. First we will introduce some af the basic concepts and compare them to classical probabilistic concepis. Surprisingly enough we will discover some fundamental differences. Then we introduce the so-called “Kosko Cube", an instrument that greatly helps to visualize fuzzy sets. We then look at how fuzzy rule systems work and how they can be made adaptive. We then briefly look at hard- and software for fuzzy logic applications. Finally we discuss some of the success factors, We end we a note on cognitive ‘uzziness vs. randomness Bart Kosko, one of the champions of fuzzy logic starts his book, “Fuzzy thinking: the new science of fuzzy logic” as follows: sn apple in your hand. Is it an apple? Yes. The object in your hand belongs 10 the clumps of space-time we call the set of apples — all apples anywhere, ever. Now take a bit, chew it, swallow it Let your digestive tract take apart the apple's molecules. Is the object in your hand still an apple? Yes or no? Take another bite. Is Chapter 7 103the new object still an apple? Take another bite, and so on down to void" (Kosko, 1992, p. 4). Initially the appe is clearly an apple. But as ibe mumnber of pieces bitten off increases it sradually looses the property of “apple-ness”. At the end, when the apple has been Completely eaten, itis no longer a member of the class of apples. The basic idea of fuzzy Togi i to associate a number with cach object indicating the degree 10 which it belongs to & articular class of objets Initially. for out apple, this aumber willbe I cr close to L. At the end it will be zero, since the apple ceases to exist. In between it will be slowly decreasing. The function that associates a number with the object i elle th imeioa tn clasial set theory this fonction is ether 1 (ihe object belongs to the set) oF © {the object does not belong tothe se) itis also called thsGharsteiste”Fubction In the last chapter we discussed probabilities as a way to deal with uncerianty in the real ‘world, Probables are also numbers bexveen 0 and, just lke the membership fonctions in fuzzy logic. We might suspect that hey are pretty much the same. But it rns out that this is by no means the ease. Just think for a moment aboot the apple. Assume that you have eaten two thirds of it and it really begins to loose its “apple-ness". So, the value of the rmembership function might be something like 3. We would say the degree to which i is an pple is -3. We could ak: "What is the probability that this is an apple”" Would thet make Sense? The probably always requires that an event takes place or doesnot take place, the ony thing is that itis not nown whether thas ten place Or not, Figure | aE makes more sense to postulate that this is a circle to a certain degree oar 8) the poebty fais ake 8 Anotor example is the citle shown in figue 7.1 Is it a cece of is ie aot a crse? The interesting point sbout this example is Dat everyhing is ven, all firs ace Known, but the uncertainty temains. Let us take eather Teporo we of ear stenen's HOA ORNe SPANO mor Tus Incas probability and fines. There i certain probity tat it wil rain Itt were nly tht then the underlying assumption would be that it can be decrmined with 100% Certainty whetber iti raining or not At we all know, tere are many weater conditions where itis unclear if i is talning or not. heavy fog with small water particle slow driing towards the ground, someting in between fog and rice, or just afew raindrops but no “real” rain, Puzzinss i a kind of inrinsie propery of elements of categories. Probability apple to any even inthe eal Word An event inthe real word Is never certain to happen (although the probability canbe very high as inthe case of dropping a rock. Thus, in the real work we have bot, probaly and fuzziness & Basic concepts “The original ideas were developed by Lotfi Zadeh of University of California Berkeley in 1965 (Zadeh, 1965. We biety inroduee some formal In reer © sROW some interoaing siferences Between lasseal Wester and Easter thinking. The fundamental concept isthe fueay se Pury lope designates a peicular Kind of inference calculus based on fey sets. Fonzy systems employ We says st wih 9 one SRE 5. Te makes sense that itis symmetric with respect © the “middle”, and it captures the right intuition about the relative fuzziness of fuzzy sets. (BRERETA. Cy te cs oft on be he sane we Fn on ‘There have been many suggestions as to how E should be defined. One is to resort to classical information theory and use the definition of entropy (which is often called H). E is defined as: EMH Hex HA = £65 nem) Entropy is measure for the “disorder” or the degree of uncertainty in a system. Iti, in a sense, the opposite of information content. This i why information cantent is sometimes called “negative entropy", Recall from information theory that entropy is measured as follows: (8) rnp, o ‘where pj are the probabilities for an event i to happen, e.g. the probability that it wll rain tomorrow. (8) is defined in analogy to (9). I is the standard way for measuring uncertainty ‘or lack of information Let us look at the intuition behind definition (8). If m,(x,)=0 then the comtribution of that term to Hf is obviously 0. If m,(x,)=1 we have In(m,(x,))=In(1)=0. Thus, if we have & classical crisp set, its entropy and thus the measure of fuzziness E will be 0, and our definition fulfills requirement (i) above. Once again, the technical details of the formulas are not essential for our purposes, What cine he non, We cn seh vale enbexti fnson move vars 1 wil aren, ching» acim fa m;)¢. Ph eminig prope ean esty be verified by the reader, There are other measures that satisfy conditions (i) through (iv), We showed a common one, and we will add another one below This calculus can be extended in many ways. There is a whole literature on the topic. The purpose of this introduction was to provide an intuition of the basic ideas and to show how they can be formalized in a straightforward manner. In what follows, we will present a different way of viewing fuzzy sets, a very elegant one. We will call it the “geometrical” view. It is due to Bart Kosko and is beautifully explained in his textbook on “Neural networks and fuzzy systems” (Kosko, 1992). ‘What is the “geometry” of a fuzzy set? What does that mean? The classical way that we hhave just seen, interprets fuzzy sets as generalized classical (crisp) set, where the ‘membership functions assumes values in the interval [0,1] rather than from the binary set (0.1) only. Normally, we can visualize the membership functions by having an x-axis with the elements of a fuzzy set, and a y-axis with the value of the membership function. There are better ways of visualization, for osko cube". Before starting, just one note on terminology. falue to designate the value of the Remember (from your school days) that the power set is the set of all subsets. Thus, 2* is, Chapter 7 107the set of all subsets of X and X is our universe of discourse, n ‘Wepre Ts tt Yoke ke cake, We wl cal tthe “oso cote”. Wht does a fuzzy set Took like? It is a point in a cube. The set of all fuzzy subsets equals the unit hhypereube 1° =[0,1I'. This notation means that we have an n-dimensional cube with one ‘corner in its origin. A 2-dimensional version is visualized in figure 7.3, 10) ty x=an Figure 7.3 Musratin ofthe geometry of fuzzy ses. The furry subset A is a point in the 2. dimensional unit cube with coordinates (5 2). TRB element of second element, x, f0 degre 2. The cube consists of all possible fuzey sobs of two elements {1.x}. The four comers represent the power set ofthe classical set, consisting of Detements {49}. (fom Kosko, 1992, p. 270) The membership functions of the elements are. gi . 1e power set of the set consisting only of 2 elements {x,23}, 2° is {4x \{xs}{,.%4}}- These four sets correspond respectively to the four bit veetors (0 0), (1 1), (1 0), and (0 1). More precisely, they can be defined by the two-valued membership functions m,:X — (0,1) 13 A3) as one of the 32 of th membership functions m,:X [0,1] with continuous values. The bit-vector representation 135 ate ty Now consider the fuzzy subsets of X. We can view the fuzzy subset for crisp sets can be replaced by fit vector representation ( 3 "Now took atthe midpoint, AI its membership values, o fuzzy values, are 4, From our discussion above we know that this corresponds to maximum fuzziness (requirement (ii). ‘This point has funny properties. It is equal to its own compliment. Moreover, it is equal to its intersection with its compliment, and it is equal to its union with its complement. Or formally, Az ANAS =AUAS= AS ao) Just as before we can define the operations on the membership functio Mung = mnin(m,.m,) -max(m,,m,) an To practice our notation a bit, look at the following example: 108 Chapter 7Ane 20 and AUAS#X, Ins normal, ergp Sts, we always neve AEE os SEER is no: te se tere. A set A ts called propery fuzzy if has non-degenerate overlap and underlay, i In figure 7.4 we have illustrates the four sets 2 2 ‘As A moves towards the center, all four sets become maximally fuzz), if they move towards one of the comers, they converse to the crisp sets. G)=a0) an @=(00) ()=(0) Figure 7-4: Completed 2dimensional Kosko cube, Details, see text. ‘This leads to interesting considerations at ‘osko argues, that classical logie excludes this midpoint Se, a Elimination of this point leads to the paradoxes of which there is 2 lot in classical logic ike the liar from Crete who said that all Cretans are Tiars). Moreover, Kosko argues, that the paradoxes are Cardinalty yg ‘When introduetioa the classical definitions we had defined cardinality M(A) as M(A)= Ym, (2) @) 1 13.38 For our furzy set 4=(23) we ger M(A) 33. We can also interpret the fu22 0 y Gp we get MAS SHS = 5 Jpterpret the fuzzy cardinality measure M(A) geometrically. This is shown in figure 7, M(A) equals the magnitude of the vector drawn from the origin to the fuzzy set A. Buf rather than taking the normal Euclidean mettic (which in two dimension is «[x} +7 ) we use a so-called !/-norm: Chapter 7 109cara Sn2)=[Link],c4)-O82 Ss) m CePA) 0 keno) xe09 3 4 6=(00) {x}=10) Figure 7.5: Illustration ofthe cardinality of a furey st. Measure (12) is sometimes called the “fuzzy Hamming distance”. With this in mind we can have another fook atthe notion of fuzziness and see if we can find a geometric interpretation of our measure of fuzziness E defined earlier. The fuzzy entropy theorem Let us define fuzziness F of fuzzy set as follows: 28M AAae) FOS TAA.) We can now check if criteria () through (iv) are fulfilled () In the comers where we have the erisp (non-fuzzy) sets we have o=0 and therefore E(A)=0 which is what we want as i) Ta the mide, where all coordinates are , we have s=b and therefore E(A)=1. This is the maximum, since at any other point in the square a will always be smaller than b. The ‘mid-point is the only point where a is equal to b. (iii) We always have £(A’)2 E(A) if A is less fuzzy than A’, ie. closer to @ comer. This is obviously true, (iv) Inspection of figure 7.6 shows that E(A) = E(A°) ‘The fuzzy entropy theorem is as follows: MAD) MAGA) ‘Again, using the Kosko cube and geometric reasoning, the theorem is obvious (see figure 7.7). Mis again the '-norm which is the measure for the “size” of a fuzzy set. Thus, the fuzzy entropy theorem states that the fuzziness of a fuzzy set is the ration between the overlap and the underlap between the set and its complement. A lot of overlap of fuzzy set A and its complement means high degree of fuzziness. The denominator is used for scaling purposes to make E vary between 0 and 1. E(A)= ay n0 Chapter 7xeay t=) 1 3 Figure 7.6: Mustration of the fuzay entropy theorem, fu} xean 3 a 8=(00) Ix)=00) Figure 7.7: Geometric proof of the fuzzy entropy theorem, In classical set theory only if every element in A [Link] element of B. The power set 2" contains all of B's subsets. Thus we ean rewrite AC B 1e subset relation for fuzzy sets A and B as follows: as) ‘TWAS(3 0 .7) and 7 9), then A is a fuzzy subset of B, but B is not a fuzzy subset of A. If A=(.5 0.7), then 4 is no longer a subset of B. In other words, fuzzy set A either is, or is not, a fuzzy subset of B. Let us look for a moment at the following definitions of crisp sets A and B. A and B are defined in terms of the characteristic function on x, ¢ X. Chapter 7 nmPierrorrioe to bidadadoriridord for x momen, wach Ai ndash of Bat ven if we ince Simos 2 ae gf Beh eer word, bere something ae = vy call i ‘apni is high im tis example, Sine tere is'ony one foe cone pation ih is ao subsumed by B, the degre of sabetod ie large, Ths slags sts he. M(A) ngs) nd ely few lt, SUBSETHODD i eh Te moe violators aod be nrg ose latent te seek sk ieee a mre A sa SUPERSET of B. What we a heres att some dee one mt a saat te ier bla tne une Wi ome Gogee't mpent ofthe cher For funy ses, avilaton le dened ss flows: it A's conadred to bea subset of Ba Wlaion mea at mC) aye) Ifwe mam up al sialon We gue scious of Summuemapo foraat A ern ¥ maxt0, m6) =m 639) sit ———ea—e 16 SUPERSETHOOD(A, B) TE as Using this definition, SUBSETHOOD (A,B) is defined as F max(0,m,(0)—m, (0) - 5) a 1f mg (x) is al@YSEZOUAB the numerator there will always be a zero. This implies that A is entirely a subset of B (Zadeh’s definition). In other words, set A is in the cube between B and the origin (figure 7.8). ° SAB) te)=€0) x=) 3 2=09) tsa) L 3 Figure 7.8 Ilusraton ofthe notion ofa fuzzy subset. But if we do it like that we have a binary, ie. non-fuzzy definition of subsethood. But we ‘would like to express the degree to which A is a subset of B. We use our intuitions about ‘violations that we developed above to define a measure of subsethood. It is called the fit- violation-strategy. We take into account the number and magnitude of the violations: The greater the number of violations (relative to the size M(A) of A), the less A is a subser of B, fr, equivalently, the more A is a superset of B. ‘When there is no violation, there should be no contribution to the supersethood of A. (remember, we want (0 assess A with respect to B). If m,(s,)~m,(.,) is positive we have a ‘violation. In this case we want 10 take the magnitude of the violation into account. Thus the violation, normalized with respect to the size of a (i. M(A)), we have ne Chapter 7Ymax(0.m,(2,)— m6) SUPERSETHOOD(4, B)= <> cas) ‘And for subsethood (A,B) we have Lmao. 64) m4) MA Formula (19) has the propertios we would like to have, If there are no violations S(A, BI S(A,B)=0 if and only if B is the empty set. ‘What we have discussed so far in terms of basic concepts should be sufficient to provide an intuition of what fuzzy theory is about and the sorts of problems it tries to deal with. As ‘we saw, there are a number of surprising implications, For example, the fact that a set A can at the same time be a subset and a superset of B, in fact — in the real world — it will ‘always be both. We have focused on the notion of fuzzy sets, Now we want to look at fuzzy systems, Le. systems that employ fuzzy logic to deal with fuzzy sets. In the field of knowledge-based systems we discussed rules to write programs, patlern-driven programs, We showed how conclusions can be drawn under uncertainty in the previous chapter. In this chapter will demonstrate how fuzzy logic works. S(A,B) as 7.3 Fuzzy associative memories Following Kosko’s idea fuzzy sets can be viewed as points in hypercubes. stems tu fs) see foray ets 8 fnsy ste 8 ns an oT een ene a the fuy subuetsof he domain space, or np universe Gfdicous, Fels) 7 hose athe fry subse ofthe Fangs spn oF Capt universe of discourse, ¥={y,,...J,)- In what follows we will focus on systems S:I" > 1° that map “balls” of fuzzy sets in 1* to “balls” of fuzzy sets in I” in such a way that close inputs close outputs. Such systems behave as associative memories. They are cal The simplest FAM encodes the FAM rule or association (4,,8,), which associates the p imensional fuzzy set B, with the n-dimensional fuzzy set 4, In general a FAM system FIFI? encodes and processes in parallel a FAM bank of m FAM rules (4,-B).-(AgsB,)- Each input A to the FAM system activates each stored FAM rule to a different degrec. ‘The minimal FAM that stores (4,B.) maps input A to BY, partially activated version of ,. The more A resembles A,, the more B resembles B. This is the essentially property of associative memories. Engineers sometimes call the fuzzy-set association (4.,B) a “rule”, This is abit misleading, since reasoning with sets is not the seme as reasoning with propositions. If we know that we ace talking about numbers and matrices then it is OK to view them as a kind of TF-THEN rule: IF 4, THEN B,. The general FAM system architecture is shown in figure 7.9. The So, F defines a fuzzy-system transformation which maps fuzzy sets in the unit cube J" to fuzzy sets in the unit cube 7”. Ifthe input A is nota fuzzy set, but a given numerical value and the output is also a numetical value (rather than a fuzzy set), this is called a BIOFAM, # binary input ‘output fuzzy associative memory. In practical applications, BIOFAMS are most often used. Chapter 7 13FAM Rule | pLae FAM Rule? ight: Consider a fuzzy association for the flexible control of 2 traffic e fuzzy association is 4 |. The input fuzzy-set value HEAVY. The output fuzzy variable set value LONGER. Another fuzzy association might be sumes the fuzzy R), merical input. nt from a rule in an expert system which, in essence, processes symbols. It may process some numbers such as certainty factors on the side, but the numberSare not essential. In the case of FAMS, the essence is the numerical processing, — ‘What we have t0 do is to define, for example, what we mean by concepts like HEAVY or LONGER. Figure 7.10 shows a set of possible membership functions for traffic density (HEAVY, MEDIUM, and LIGHT). On the x-axis there is a numerical measure of traffic density and on the y-axis there is the degree of membership to the corresponding variable. y 1 mete Mccl®) meals) ° 0 100 150200 FFigure 7.10: The membership functions for HEAVY, MEDIUM, and LIGHT traffic. On the horizontal axis there is 2 measure of trafic density ue Chapter 77.4 Binary input-output fuzzy associative memories: BIOFAMs ‘As mentioned in “numeri vaiet For example. the tl FA een light duration, ie density (0 “To make things abit more concrete, let us look ata standard benchmark for any optimization technique, the inverted pendulum (also called the pole balancing problem). It has been siseusted extensively in the Iiterature on neural networks, genetie algorthms, control theory, and, mia se, fuzzy control. The goal is to adjust a motor to balance an inverted pendulumin rw8 dimensions. The adjustment is to be achieved by a BIOFAM. There-er 1 of steps to be followed in the development of a BIOEAM: (1) Mdentifieation of th ‘These variable are called linguistic because they are derived from (2) Definition of the membership functions, (3) Construction of a set of fuzzy rules (the FAM rules). (4) Combination of all outputs of the “rule set” into a geometrical output function (5) “Defuzzification” in order to get a non-fuzzy output value. ‘The experimental set-up is shown in figure 7.11. 1 otyso Figure 7.11: Basic set-up ofthe pole balancing problem. proximate the angular velocity as the difference-betweenthe current angle ©, and the previous angle @,,: A@, =@,-©,,, ‘The variable that we want to control is the ea aeaeeie is actually not quite*trae- ‘What we do want to control is the force applic. ‘art. But we cannot influence the force directly. Thus, we assume that there is a linea, of atleast monotonous r between motor curren e. We define the following ditection ‘we have to define the domain over which they can vary, ie. the universe of discourse. For all three variables this is the real ine R. Since angles greater than Ives to the interval [-90°,90°] for the (angle) variables we have to define the membership functions. In universe of discourse (ie. for ©, A@, and v) into ice, whether we want to have only a coarse quantization or “whether we request a more fine-grained one. The appropriate quantization depends on the Chapter 7particular problem. Let us choose a coarse one consisting of small, medium, and large. This leads to seven fuzzy-set values, namely Negative Large ‘Negative Medium Negative Small Zero Positive Small 2 Positive Medium eahaae Positive Large “The fuzzy set values have labels, comesponding to the fuzzy-set values they can assume. The tariable © can, for example, take NL asa fuzzy set value. This i 2 er) important sep in the design ofa fuzzy controller, Let us quote KBMMGEPThe expressive power of the FAM approach stots from teas foxzpeiee quamtsamonsr ia Sor' robe we mee system Gimensions, and we describe a nonlinear numerical process with linguistic commonsense terms.” (Rosko, 1992, p. 318) Now we have to choose the membership Functions. The shapes can vary. The set ZE may define a Gaussian curve forthe pendulum angle ©, a tingle for angular velocity 40, and @ Gaussian curve forthe motor curent But ll fuzzy sets center about the numerical vale zero. The most common shapes are triangles or trapezoid shapes because the require very Tite computation. The exact shape does not mater very much. Atypical eboice is shown ia figue 712 oo Figure 7.12: Example of atypical choice of membership functions contiguous fuzzy sets overlap is roughly 25%) This value has been rlap blurs the distinction between the fuzzy set values, le results. Let us choose similer membership ‘The degree little overlap tends to produc tions for the variables A® nd fe HOW have to design the rules. They can be jn various forms. A natural-language like formulation would be: IF the angle is negative medium AND the angular velocity ero ositive medium Using the labels for the fuzzy sets introduced above we can write them in a more compact, way: IF @=NM AND A@=ZE THEN \=PM ‘An even more compact way to write the rules would be eM ZEEM 1 onder to design the complete bank of rules we have to define 49 rules, For each entry inthe rule bank (see figure 7.13) there are seven possible values, thus there are 73 possible rules. Th isin desing te os i ow St on i oft fewer it re at variables ven to them (about zero, negative small, etc.) t t us start at the middle, Ifthe angle © is about zero (ZE) 6 Chapter 7and the angular velocity A is about zero (ZE), then the motor current v should be about zero (ZE) and we have our first rule: (ZEZE:2E). Now suppose that the angle © is zero but the pendulum moves. If the Seca per a tnpn angular velocity is in agate eee aoe © velocity should be. This leads, for example to the following rules: (ENSiPS) (ZENM:PM), etc, The symmettic cases of these rules are obvious. ‘The result of this commonsense analysis is shown in figure 7.13. As we see, only 15 of the 49 possible entries contain rules. In practice itis often the case that we do not need all the But which ones should we take? There are, in essence three ways in which this ean be wich is what we have done. The ee why Ree Prot ene bank (ss et a exec te ap Hc iy atc of timer “notng nut te ms tr the contest" vhonever tts sinatons osc the sm il aw be Sr designed pope asco sboul eve: as) c really required (see below). ae NL NM NS PSPM PL NS NS ze [ot [| | es PS ‘PS PM PL. alzizjalalz|e| 8 a : z “Figuce 7.13: A FAM sulle bank for the inverted pendulum problem, : Now results of wwe have the rules we have to say how to apply them an individual rules. The idea is illustrate in figure 7.14, Chapter 7 nTFigure 7.14: Mstration of the ‘ot applied since the input values @=15 and AO: ‘membership functions. clas In gral he antecedents of al FAM ules iy In figure 7.14 the input values ©=15 and A@=-10 are shown. How do w etermine the output? The rules are all activated in parallel, but to varying degrees. The degree depends on how well the input value matches the membership Function on the left- hhand side (antecedent side). For example, in the top rule shown in figure 7.14 the membership functions NS and PS do not intersect with the input values. The only «wo rules for which this is the case are the bottom two ones shown in the figure Let us look at the first. The membership function for PS is intersected by @=15 at 0.8, the one for ZE is intersected by A@=—10 at 0.5, We can write m$,(15)=0.8 and mif(-10)=0.5 ‘These values are combined by using the minimum (logically speaking the AND) function: sin(m’ (15), m2 (-10)) = min(0.8,0.5)=0.5. Thus, this rule is applied to degree 0.5. An sieramive would be to take the prodst intend of the minimum. Ths would be he (BorrtaoRiproducrmeno} The colon predict teiod leads smaller cas onthe us ‘The fiat male is }0 do not intersect any of their Chapter 7consequent side. Thus, in order to have the same weight, a rule in a correlation-product inference procedure has to match to a greater extent, In practice, both methods— correlation- minimum and correlation-product — are used The ‘Src mens sede es the angular-velocity fuzzy set ZE to degree 0.5 which is propagated tothe consequent side: ‘min (15),mf2(—10)) a min (0) = 0.5.0 m0) Similarly we find for the other rule: smin(am$ (1.5), m2 (~10)) a.m,(v) = 0.2.0.m5,(v) If we now take the maximum over all the consequent sides of the rules (the shaded areas) we ‘get the fuzzy centroid shown at the bottom. This is the geometric function that we ‘were looking for Now we have to ‘value for v, since the motors require one crisp single value for current. (20) “Ton ‘his isthe general forma. Ie can be considerably simplified forthe kinds of trapezoid Shapes tht we have See in oir exatnpe. Moreover, in standard futy logic software, there tr libraries of function forthe various shapes ofthe membership funcons The BIOFAM inference procedure described hee requires exact inputs (0 called binary or dela pulses). The procedures can be generalized tothe cae where the Snput elf ea fizzy Sst whichis the cate of (general) FAM inference system, Adaptive BIOFAM clustering In the last section we have used our common-sense or engineering knowledge to design a fuzzy controller. Earlier we have seen that often, if enough data are available, systems can be enerated automatically. This was illustrated in the Bayesian approach and is also intrinsic in neural network or induction-oriented designs. Similarly, the structure of fuzzy controllers can be adaptively “learned” if there are enough data available. In control applications, humans or automatic controllers generate a continuous stream of— obviously appropriate— input-output data. Adaptive BIOFAM clustering convers this data to weighted FAM rules. In essence, BIOFAM clustering counts synaptic quantization vectors in FAM cells. The clustering procedure samples the nonfuzzy input-output stream Gy.3,9,)-- Am unsupervised clustering procedure distributes the k synaptic quantization vectors. m,...1m, in Xx¥. Learning distributes them to different cells in the rule bank. The key idea is that each cluster equals a rule. This is ilustrated in figures 7.15 and 7.16 Chapter 7 119Figure 7.15: Distibution of input data (x,),) in the input-output product space XY. SX] ML NM ONS) ZEB OPS) MPL, NS 7 ze Te PS 7 PM aT . Figure 7.16: Distribution ofthe quantization vectors, given the dts shown in figure 16. ‘The procedure to perform adaptive BIOFAM clustering can be described as follows: + Identify the variables. - Find the linguistic labels and define the membership functions. = Take an existing, functioning controller, a human or a machine = Choose the number of quantization vectors (e.g. one for each cal) = Generate cases. For each input vector generate the appropriate output action (i.e. the control actions). Register into which cells the output actions fall, given a particular Input vector. - Perform a sensitivity analysis. This may enable you to remove some of the rules, thus ‘making the system smaller and more efficient, - Deal with the cases for which you do not have data on the basis of intuition or prior knowledge. In practical applications a combination of methods will typically be used to generate the FAM rule-bank, for example adaptive clustering together with theoretical analysis (using prior knowledge and intuition) = us look at 2s ‘example, the truck backer-upper. Like the inverted pendulum it is one of the universal benchmarks in the optimization literature. It will be used to illustrate a ‘number of points. Membership functions can vary depending on the degree of precision required. Sensitivity analysis can be used to test the rule set for various properties. The truck Dacker-upper with a trailer requires a three-dimensional FAM rule bank, 120 Chapter 7ow Figure 7.17: The truck backer-upper. Let us simplify! the problem a bit by only specifying shegeeooniinateyatber than they coordinate. The input variables are then the x coordinate angle of the truck @ . The output variable is the steering angle ©. The ranges of the variables are: 0x10 -905@ 5270 3050530 Let us, once again, use our commonsense to determine the fuzzy variables and the membership functions. The following table shows the labels for the variables: Angle @ xposition x steering angle © RB: Right Below LE: Let NB: Negative Big RU: Right Upper LC: Left Center NM: Negative Medium RV: Right Vertical CE: Center NS: Negative Small VE: Vertical RC: Right Center ZE: Zero LV: Left Vertical RIE Right PS: Positive Small LU: Left Upper PM: Positive Medium LB: Left Below PB: Positive Big ‘The membership functions sre shown in figure 7.18 Chapter 7 21mx) Le Le cE RC RI 0 50 100 x me ®) 20 ° 50 10 270 ®, m2) 2 30 2 2 «8 Figure 7.18: The membership functions forthe track backer-uppor. oN iw ic ce oR Rm rp |S] PM | PM | PB] PB ru [RS [Ss | PM | PB | PB rv [NM [NS [Ps [eM | PB ve [NM [NM [Ze [PM | PM Ly [NB [NM [NS [s_ [ow Lu [NB PNB [NM [NS | PS La [NB [NB [NM | NM [NS Figure 7.19: Rule-bank for the truck backer-upper. Sensitivity analysis Generally speaking the goal of sensitivity analysis is to get a better feel for the system's ‘behavior. In particular we are interested in its robustness. For examples, we would like to know what happens if a rule is changed or removed, or if additional rules are added, Remember that one of the goals is to have as few rules as possible in the system—not all the cells in a rule-bank need to be filled. We have to find out which ones are essential, As one might expect, Kosko (1992) demonstrates that the fuzzy controller for the truck backer- ‘upper is very robust (in additional to being better than the neural controller). m2 Chapter 7Fuzzy truck-and-trailer controller ‘The problem of the truck backer-upper can be made increasingly more complex. For example, a trailer can be added, giving the system additional degrees at freedom. Instead of having only the angle of the truck we have two angles, one for the truck and one for the trailer. The consequence is that there are now three antecedents, The FAM rule-bank is no longer @ two-dimensional table, but a three-dimensional cube. If we want to represent the rule-bank two-dimensionally we have to list the 2-d rule-bank forall the values of one variable, say x. For each value of x there is a table. ‘This should suffice to provide a good ides of what fuzzy logic is all about, how the basic mechanisms work, and how it can be applied. We will now discuss a number of important related topics. 7.5 Hardware and software ‘As mentioned initially, fuzzy is considered a “quality seat” for many consumer products. [n order to bring i ito these products, fuzzy logic has to be implemented in cheap ways. In forder to develop fozzy logic controllers, lols are required. There are many software tools 5 Well as specialized hardware available for fuzzy processing, While for consumer electronics speed is often not truly critical there are many applications ia which high precision and hish speed is required (e.g. control of a ear engine). This is why specialized hardware has been developed Recent itaplementations of fuzzy logic in analog and digital hardware are described in TEEE Micro, Augest 1995, special issue on fuzzy hardware. For example, Donalé L. Hung (1995) reports on an implementation of dedieste digital hardware (2 ficld-programmable gate array) capable of performing 33Mio FLIPS (Puzzy Logic Inferences Per Secon). While currently most fuzzy logie processors are digital, there have been attempis to develop fuzzy inferencing systems based on analog fuzzy chips (e.g. Miki and Yamakawa, 1995). Recendy, hardware for combining neural network and fuzzy logic technology has become inctessingly popular: they take account of the fact that the real world is intrinsically fuzzy and they exploit the learning and generalization capabilities of neural networks (see next chapter). On the software side there are a number of software packages a\ailable which typically inchide libraries for standard membership functions (angular sbaped, ox tapezoid), a9 well as interpreters for fuzzy inference. Am example is demonstrated in class. The software enables the user to quickly come up with a system and perform a sensitivity alysis: rules and membership functions ean be changed rapily so thatthe engineer quickly gets a good feel for the system. This kind of prototyping ean form the basis for hardware implementations for « particular product. We conclude by providing a summary of different types of implementations of fuzzy systems. Chapter 7 13Table 7.2: Implementation of fuzzy systems(from Miki and Yamakawa, 1995) Implementation ‘Advantages Disadvantages ‘Software implementation TEasy to implement; Tow Low-speed operation cost Implementing fuzzy logic _| Easy programming = ‘operations in microcode CPU supported by fuzzy High-speed fuzzy Data-handling time limits coprocessor ‘operation the operation speed CPU including fuzzy Additional high-speed — ‘operation at hardware fuzzy operation level (fuzzy CPU, fuzzy CPU core) Exclusive fuzzy inference | Fastest fuzzy operation; High cost engine (fuzzy chip) suitable structure for fuzzy logic architecture 7.6 Fuzzy logic for real-world computing The appeal of fuzzy logi As pointed out initially, fuzzy logic introduces a number of highly interesting concepts in dealing withthe rel vor ES MORAEMERETAnaeesunerinn, we are to automate any sort of activity in the real world, we must s intrinsic uncertainty into aecount—one way or other. Fuzzy thinking contrasts with probabilistic thinking. The latter cals with uncertainty because of lack of information (e.g. lack of knowledge of future events), but once the information is availabe, the uncertainty has been removed. By contrast, fuzzy thinking tries to capture those uncertainties that are simply inherent in anything in the real world. When we say that it wl swith 20% chance, we have two types of uncertainty, a probabilistic one don't know whether it will ain or nod, and one having to do with fuzzy categories (even if we talk about the current weather conditions, we sill may not be sure whether itis raining or not). Note thatthe way we have been using the term “fuzzy thinking” is in the sense of “applying concepts from fuzzy logic”, rather than unclear and nebulous thinking, ‘As pointed out earlier, fuzzy logic is especially popular in Japan. There are various reasons for it, Firs, it seems’that this kind of thinking, a kind of non-Cartesian thinking, comes more natural to Eastern peoples than to Western ones, For Westerners, the principle of the excluded middle is a very natural one. It is hard for us to imagine how things might be different. Fuzzy thinking, in some very important sense, reflects a different way of being in the world. Sccond, though originally developed in the US, it has been brought to its current level of popularity and success by the Japanese, not by the Americans or European, In this sense, i is something genvinely Japanese. The success of fuzzy logic ‘There is no doubt that fuz#GHS BASIGSEN NEB) SveeessTAl. AI the products which are now available in consumer and industial products testify that this is indeed the case. In this rection ¥¢ would like 19 look at some of the success factors for fuzzy logic applications Whi has fuzzy logic been so sucessful? What do we have to take into account when defining, a project? ia Chapter 7acer ert ol choice of application is, of course, crcial for any technology. If we 1ook at ‘here fozzy Tose has been used, its one the one hand consumer products, gd on the other industrial applications, Most of it has been atthe ie ae. though the technology of fuzzy Togie permits in principe, the de sok reasoning “yee cool bls peal mach cn hn SRS? Shustons, and (©) iiilOEinipeon oe reamonsuatere ), the real world is defined by the designer's low-level ontology (sensors, actuators). In consumer applications, it is relatively simple: concentrations of dirt and calcium in a washing machine, time-of-flight signals in ultrasound-based auto-focus cameras, or temperatures and temperature differences in refrigerators. Note also that in these types of applications, the damage entailed if for some reason the device is not function correctly, is relatively minor. In industrial applications, or in wansportation systems, a kind of “over- engineering” will have to take place: redundant sensors and monitoring systems will have to be introduced in order to minimize the occurrence of unanticipated situations. Remember that the big problem are not the situations in which the systems encounters an unforeseen combination of facts, but rather one in which the system does not “realize” that it is confronted with @ novel situation, because the appropriate sensors are not available. In a subway system, for example, unforeseen combinations of sensor data can be dealt with by simply stopping the train (to reduce risk of accident). Redundant measure of speed might include wheel turns per second, rpms of the electrical motors, time between stations, etc. The more dif spect cone ORGS WOH SOROS ever reasoning in humans is grounded in a person's experience with the physical and social environment. This is the basis for all reasoning and judgment. Sine a machine cannot make such experiences, it will not be able to interpret these situations in similar ways. This is, as ‘we mentioned in earlier chapters, one of the main reason why expert systems have not been 25 successful as originally expected: they tried to automate processes at the level of decision making requiring a “social individual’, Fuzzy logic might seem more appropriate, since it takes the fuzziness of natural language and everyday concepts into account. But just like traditional rule-based systems, the—high-level—eoncepts are not grounded. This is why, one might speculate, natural language understanding systems based on fuzzy logic will have much the same problems as traditional ones, Choosing the right level of application is, of course, a prerequisite for success: if this choice hhas been inappropriate, there is no hope. But there is another factor that is equally important: where is there a chance of success for the technology? Let us illustrate this point with the case study on kiln control systems that we outlined in a 3. Introducing fuzzy technology implies that the data required is of sufficient qual Moreover, what is the value of the best control software, if the hardware cannot exploit it? For example, if the controller “suggests” to increase the dosage of fuel by a very small amount, and the physical system only works on the basis of tons, the controller will be Prony oicles, Thus, AURA! be ncodueed, the technology of the application domain has to be right, Where can we expect big successes? Typically in those places where the technology is currently outdated and not well-maintained. This is exactly What the proponents of fuzzy control in large companies have boon doing. Fist, they aviomated factories in developing countries, rather than in countries where the maintenance level and the technology standard is generally high, ike Switzerland, Germany, or Japan. for example, with their concept of which is based on Tuzzy logic, have reported tremendous successes. We can safely wust the figures reported. What the reports do not stress, however, is the fact that in order for “high-level control” to work, the factories had to be modernized in terms of equipment (sensor technology, effectors), and in terms of enforcement of maintenance rules. They deliberately chose factories in Brasil and Pakistan first, Had they sterted in Switzerland (so the expert sid explicitly), they could not have demonstrated impressive suecesses since they are already working at top performance with the current systems. Chapter 7 125In summary, the introduction of fuzzy control often has highly beneficial “side-effects” that 0 not oa p a late to fuzzy technology If fuzzy technology is to work, it has to be done seriously. Real-world applications require experienced engineers for tuning the rule-sets, for choosing the variables and for adjusting the membership functions. The high quality of Japanese engineers is certainly an important success factor. Even though fuzzy logic is easy to understand for any computer science student, thls should not be interpreted as implying that applications are easy to develop, even ‘though the intuitive plausibility of the technology may act as an enhancer. Other factors for So far, we have discussed success factors generally. Now we briefly look at some additional factors that might have contributed to the success of fuzzy logic i We have already mentioned the fit with Easter thinking. Another one concems a different attimde towards automation. This can be seen in consumer products, but also in industry. Automation is considered intrinsically positive, at least this has been the case until very recently. There is a slight shift in manufacturing, where the degree of automation is no longer increased, but rather there is a tendency to reduce it somewhat. But still, the main thrust of development is on automation. This relates to the following problem. For Japanese people it is OK to have a device where you push a button “on” and then you let the machine do the work for you. Westerners like to have many buttons to have more control over the device. American marketing analysts have suggested that one of the success factors might be the éetrminion wih which ih Sp reGUHSe Wig REMI ito consumer products. The stress is on digital technology, rather than fuzzy logic technology. They might have achieved the same level of success, 50 the analysts, with other, more conventional technology (like classical control theory). While this may be truc, it is missing an important psychological point. Whenever a technology is to be successful, vision of going to the moon has been such a vision, and has enormously boosted technological development in the sixties. Building “intelligent machines” (whatever that exactly means), has generated an enormous amount of highly creative research in artificial intelligence and related fields. “Fuzzy” might just be such 2 vision capable of generating enough enthusiasm As a last success factor that benefited the Japanese success, was certainly an excellent marketing strategy. This strategy is, in some sense, related to the vision provided by fuzzy". The Japanese managed to promote “fuzzy” to become a quality standard in a world where everybody is complaining about degradation of quality: quality of service, quality of products, and generally quality of life Fuzzy logic or other approaches ‘There has been a lot of discussion whether fuzzy logic is superior to other approaches like neural networks, or control theory. Reasons given concern the inclusion of expert knowledge, ease of implementation, robustness, and lower computational requirements, As ‘we saw above, good engineers are still required to get an application working. The choice of the variables and the shapes of the membership functions require a lot of expertise and know- how. Fuzzy controllers cannot be designed by boginners only familiar with the principles of fuzzy logic: the designer needs to have experience with controllers. This is sometimes used as ‘an argument against fuzzy logic. But the fact remains that it is easy to incorporate a priori knowledge. On the other hand, because of the complexity of the interpolation process (via the membership functions), it is very difficult to derive proofs for the control laws generated. Obviously, what can be achieved by fuzzy logic approaches, can, one way ot other, also be achieved by other methods. More bluntly speaking: in the end everything maps down 10 machine code. The question seems less whether it could be done by other methods, to0, but rather whether it is done or not. The Japanese have done it with fuzzy logic. And they have bbeen successful. It is clear that success always produces envy and draws criticism, 126 Chapter 77.7 A note on cognitive science One of the goals of artificial imelligence is to make models of intelligence. As we discussed in earlier chapter, the symbol processing model (the PSSH) has only had limited success. Many alternative approaches have been proposed such as neural networks, Bayesian approaches, autonomous agents, and, of course, fuzzy logic. Natural language, for example, is vague. We continuously use expressions such as big, small, intelligent, colorful, etc. They 0 not have a crisp, well-defined meaning. Thus, it seems natural to apply fuzzy logic to natural language processing or to the modeling of psychological concepts in general. Many theorists of fuzzy logic, including Lotfi Zadeh himself, and HJ. Zimmermann in Aachen, Germany, have tried to apply fuzzy logic to natural language understanding. Our hypothesis is that this will not work any better than previous approaches. Let us examine this more closely. Just as in classical approaches, a high-level domain ontology is required. The only thing ‘which changes is that objects only belong to these categories to a certain extent. A fifly year ‘old person belongs to the category of old people only to degree, say 0.6. But the category is there @ priori, as defined by the designer. The grounding (see chapter 4) is lacking just as much as it is in expert systems. As long as high-level categories, fuzzy or crisp, are pre- defined, the system will never be able to develop its own categorization of the real world, and will therefore never have natural language capacities as we know them from humans, ‘This is a cognitive science perspective. It may well be the case, that—pragmatically speaking—an acceptable level of performance can be achieve, for example for automatic translation programs. ‘The same arguments apply, in essence to expert systems work. The minute fuzzy logic is applied to high-level human reasoning and decision making, it is—for the same reasons, i sgrounding—doomed to failure Points to remember - Fuzzy logic has been enormously successful in many applications, especially in consumer products. But many industrial applications have also been reported Fuzzy logic takes the intrinsie uncertainty inherent in the real world into account. Even though all the information is available about a circle drawn on paper, it only belongs to the category of circles to a certain degree, ‘There is @ fundamental difference between probabilistic and fuzzy logic thinking. The fact that T don’t know whether it will rain tomorrow, is due to lack of information, ‘The fact that I can't say whether it is currently (really) raining or not (érizzle, heavy fog), is due to the intrinsic uncertainty. - Fuzzy sets differ from classical crisp sets by their membership functions which vary between 0 and 1, rather than being binary. All the operations like intersection or union ccan be generalized to fuzzy sets = In contrast to crisp set theory, the intersection of a fuzzy set with its compliment is not the empty set. Aristotelian thinking in terms of the excluded middle leads to paradoxes. The “Kosko cube” is an excellent instrament to visualize fuzzy sets: they are interpreted as points in an n-dimensional hypercube. The cotners in this hypercube correspond to the classical crisp sets, - Size and “fuzziness” of a fuzzy set have straightforward geometric imerpretations in the “Kosko cube” view = There are five steps in the development of a fuzzy rule system, namely identification of the fuzzy variables, definition of the membership functions, construction of the FAM rules, combination of the outputs into a geometrical output, and the defuzzification, = Fuzzy inference procedures apply all rules in parallel. Typical inference methods used are correlation-minimum and correlation-product. An OR-function is used on the Chapter 7 12128 consequent side, The most often used defuzzification method calculates, in essence, the center of gravity, Fuzzy systems can be made adaptive. Combinations between fuzzy and neural systems have also become popular. A lot of specialized hardware and software is available for implementing and prototyping fuzzy systems. ‘The choice of the right level of application is a erucial success factor, Fuzzy logic is not a “deus ex machina”: its applications requires good experienced engineers Fuzzy logic is not @ good tool for cognitive science. Chapter 7