rick@cs.arizona.edu (Rick Schlichting) (01/23/91)
[Dr. David Kahaner is a numerical analyst visiting Japan for two-years under the auspices of the Office of Naval Research-Asia (ONR/Asia). The following is the professional opinion of David Kahaner and in no way has the blessing of the US Government or any agency of it. All information is dated and of limited life time. This disclaimer should be noted on ANY attribution.] [Copies of previous reports written by Kahaner can be obtained from host cs.arizona.edu using anonymous FTP.] To: Distribution From: David K. Kahaner ONR Asia [kahaner@xroads.cc.u-tokyo.ac.jp] Re: 2nd Fuzzy Logic Systems Inst Hardware Practice Seminar 15 Dec 1990. 21 Jan 1991 ABSTRACT. A hardware practice seminar (15 Dec 1990) on building real time control devices using commercially available fuzzy chips is described by Dr. Thomas Hagemann, GMD Tokyo. INTRODUCTION. Prof Takeshi Yamakawa Dept of Computer Science and Control Engineering Kyushu Institute of Technology Iizuka, Fukuoka 820 JAPAN Tel: 81-948-28-5551, x401, Fax: 81-948-28-5582. is one of the key Japanese scientists in the area of fuzzy logic. He is the Chairman of the Fuzzy Logic Systems Institute (FLSI), (see my report "fuzzy", 24 August 1990) and has designed and built various fuzzy chips. FLSI, founded March 1990, has been sponsoring a series of workshops on aspects of fuzzy logic. The current workshop was attended by Dr. Thomas Hagemann German National Research Center for Computer Science (GMD) Deutsches Kulturzentrum 7-5-56 Akasaka Minato-ku, Tokyo 107 Japan Tel: +81 3 3586-7104, Fax: +81 3 3586-7187 Email: HAGEMANN@TYO.GMD.DBP.DE For additional information on GMD see my report "gmd", 6 Dec 1990. Hagemann wrote the summary given below of the one day workshop and has agreed to allow its general distribution. 2nd FLSI Fuzzy Hardware Practice Seminar 15.12.1990, 9:30-16:45, Tokyo Lecturer: Prof. Takeshi Yamakawa Abstract The seminar was organized by the Fuzzy Logic Systems Institute (FLSI), to show how to build real time control devices through programming ones own special highspeed fuzzy hardware using commercially available fuzzy chips. Material handed out at the seminar consisted of copies of viewgraphs and papers of Prof. Yamakawa on fuzzy hardware devices dating from the years 1987 to 1989, as well as two rule chips and one defuzzifier chip (with instruction manuals) designed by Yamakawa and produced by Omron Tateishi. It follows a summary of the seminar, some comments and a description of the FLSI. 1. Summary of the Seminar The seminar was planned for an audience of 100 people, participating were about 60, incl. one woman and one foreigner. Identical seminars were held on different days in Osaka and Fukuoka. Attendance fee was 30.000 Yen (ca. 160 ECU) for the ordinary audience and 5.000 or 10.000 Yen for FLSI-members; lunch, coffee and tea was included. Considering the value of the 3 fuzzy chips this was a very reasonable price. The seminar was the second in a serious of seminars organized by FLSI, the first was about trade mark and patent right issues for fuzzy engineers. Others are to follow about fuzzy business management, stability of fuzzy systems, neuro computing and chaos, fuzzy inference and again, because of the strong demand, on patent and intellectual property right issues. The lecturer, Prof. Yamakawa, is a very agile person in his forties; during the seminar he wore a travel bag and a mobile telephone at his belt, using the latter one in the lunch break. He seems to be quite devoted to research and education about fuzzy systems and is a good and experienced teacher. The seminar had four parts. First, the basics of fuzzy system theory were explained, stressing the need to incorporate concepts of vagueness into the description of complex systems, where purely symbolic methods were not sufficient. In the second part a concrete example of a control system was given, where using fuzzy methods turned out to be easier than traditional ones. The approach to the original problem, to balance an inverted pendulum, is cited rather often recently and belongs to the repertory of many talks about fuzzy theory. The usual claim, however, controlling an inverted pendulum by classical methods is nearly impossible and hence fuzzy control is necessary, seems to be an exaggeration, since there is at least one system I know about (at Fraunhofer Society in Karlsruhe, Germany), which dates back some five years. Anyway, Yamakawa extended the problem into that of balancing a full glass of wine standing on a rod, which itself stood on a movable cartridge, and asked his students to build a fuzzy control system for that. Some 20 students in their second and third university year, with no experience in classical control theory, were given a printed board with 3 rule chips an 1 defuzzifier chip, an equivalent of 24 school hours and a budget of 5.000 Yen (ca. 26 ECU). The system was completed just the day before the seminar. The fuzzy rules were of the kind "If the rod is leaning to one side VERY MUCH, move the cartridge FAST to the other side" (capitalized words indicating fuzzy variables). It took the students only a lot of experimentation with the rules and the underlying membership functions, but no knowledge of control theory. The third part was on fuzzy logic control and its realization as electronic circuits. Each fuzzy variable of a rule is modeled as a membership function. Within the rule chips of Yamakawa/Omron Tateishi, a membership function is a piecewise linear function with values between zero and one, represented approximately by seven discret values between 0 and 5 Volt, ranging on an abscissa of 25 points, usually between -5 and +5 Volt. The graph of a typical membership function is triangular or trapezoidal with five parameters of variation: abscissa of highest point and angle of both rising and falling slope and maximum of the function. Infering a conclusion from a set of facts by using fuzzy rules is done in the following way, assuming that the premise of a rule consists of logically ANDed predicates: The membership function of each predicate of a rule is evaluated at the point given by the fact. Next, the minimum of the obtained values is taken. The membership function representing the conclusion of the rule is then truncated at the height of that minimum. This procedure is done for each rule and the truncated membership functions are combined to one function by taking the maximum of these functions. The resulting membership function represents the conclusion of the given facts. This is done by fuzzy rule chips, one for each rule, and a MAX-circuit. Now, the concluded membership function must be turned into one single value, the "essence" of the concept represented by that function, by using the defuzzifier chip, which computes the center of gravity of the (graph of the) function. Inference speed of the rule chip is 1 microsecond, that of the defuzzifier chip 5 microseconds. The final part of the seminar should have been an introduction to the usage of the rule and defuzzifier chips, but had to be deleted because of the advanced time. The instruction manuals being handed out together with the chips are said to be self-explaining for people with experience in electronic circuits (which I don't have). All in all, the seminar was very interesting and instructive. No advanced frontline research results were taught, the material distributed was about 2 or 3 years old, but the seminar helped to "defuzzify" the fuzzy-hype of certain magazine reports, and was one measure of educating the public about this new technology. 2. Comments on the seminar The participants of the seminar were all in their thirties to fifties, apparently not coming from an academic environment alone. This shows (again; cf. the intensive media coverage and product advertisements about "fuzzy") the awareness of the general public in Japan of the concept of fuzziness. Whether this is also due to cultural influences - in Japan vagueness comes not necessarily with uneasiness - I won't discuss here. If, on the other hand, this awareness is only superficial and without real understanding, there might be some danger in applying fuzzy theory to real world problems. The seminar at least made also clear some of the weak points of the present fuzzy technology, some examples of which follow. If an engineer uses the Yamakawa/Omron-chips or possibly other chips, she will have to deal with constraints of that hardware, and might even be unaware of that. One constraint is certainly that the rule chip can handle only special forms of membership functions: trapezoidal functions with 5 determining parameters, discretly approximated on 25 equidistant abscissa and 7 ordinate points. Being not an expert I wonder whether this might be a severe restriction. Another constraint built into the hardware is the way the defuzzifier interprets the inferred membership function by calculating its center of gravity. First, there are other methods of defuzzification like center of support, center of maximal area or half area median. Second, computation of the center of gravity usually involves division of a weighted area by the whole area of the functions graph, and to avoid complicated division circuits in the chip, Yamakawa chose to modify the original membership function to make its area equal to 1 by shifting it along the ordinate (and truncating at zero), which is accomplished in the chip by a feedback circuit; but this process, which is much easier to implement in silicon than a division, generally changes the center of gravity, since some area is cut away, the effect of which is unknown (to me). Third, by defuzzifying a membership function into a single value, here the center of gravity, quite different functions are interpreted identically, like a function pair representing MEDIUM SIZE and EITHER SMALL OR BIG SIZE, or even negated symmetric functions for MEDIUM SIZE and NOT MEDIUM SIZE. Another point of caution in the usage of the chips implementing fuzzy rules is the choice of membership functions for a rule and the distribution of the functions along the abscissa. If, for example, there is a fact for which no rule fires (i.e. no membership function has a non- zero value at the given point), the whole system might react in an unpredictable and chaotic way. This chaotic behaviour seems not to be well understood and might cause calamity. Different from classical general purpose computers, which can be programmed with some effort ad libitum, this kind of hardware has built in various constraints, and a lot of care is needed when using it. Although these constraints have not been addressed explicitly, sometimes only declared as necessary for easier implementation in silicon, the seminar helped to show what's about those fuzzy chips one hears about so often recently, and that there is still open area for research. This need for further research and a certain concern about the recent fuzzy boom in Japan, where "fuzzy" is a fashinable catch word building an image of an easy to use technology, well suited to the Japanese, solving many if not all problems left over from classical technologies, were reasons for Prof. Yamakawa to initiate the formation of the Fuzzy Logic Systems Institute. 3. The Fuzzy Logic Systems Institute (FLSI) Oaza Yokota 820-1 Iizuka-shi, Fukuoka-ken Japan 820 Tel: ++81-948-24-2771 Fax: ++81-948-24-3002 Since I did not visit FLSI, I can only give a rough translation of the written material handed out at the seminar. FLSI was founded in March 1990 as a public non-profit organization (zaidan hojin) with the permission of both the Ministry of Education, Science and Culture (MESC) and the Ministry of International Trade and Industry (MITI), endowed with a fund of 100 Mio. Yen (about 530 kECU). Chairman is Takeshi Yamakawa, who is also Professor of the Faculty of Information Engineering of Kyushu Institute of Technology. He holds 50 patents related to fuzzy technology, some still pending, 8 of these outside Japan. Since the release of his fuzzy chips he is a very busy person, giving speeches and advise about "fuzzy things" on about 100 days a year. The purpose of founding FLSI was "to carry out experimental research about fuzzy information processing and neuro science and to popularize these fields within society in order to overcome the weak points of the traditional decision theoretical knowledge information processing." Research topics are (1) Basic Research (2) Application to Engineering (3) Application to Medicine (4) Application to Human and Social Sciences like Economics, Commerce, Politics, Education and Legislation. FLSI conducts the following activities in the field of fuzzy systems and neuro science: (1) contract research (2) cooperative research (3) exchange of researchers (4) research guidance for university graduate students (5) IC production training and experimental design of fuzzy and neuro chips (6) holding seminars (7) publishing texts, tutorial books and technology information periodicals (8) application, maintenance, preservation and administration of patents (9) consulting, technological advice and technology transfer (10) performing and supporting international and national conferences, symposia and workshops (like the Int'l Conf. on Fuzzy Logic and Neural Networks, July 22-24 1990, s. D.K. Kahaner's report from 24.8.90). To accomplish its aims, FLSI cooperates with the following overseas organizations in: (1) Spain Universidad de Granada Centre d'Estudis Avancats de Blanes Instituto de Automatica Industrial (2) France Neural & Fuzzy Systems Institute (3) USA NASA Johnson Space Center and Ames Research Center (4) China Beijing Shifan University Sichuan University (5) Brazil State University of Campinas. Organizations supporting the aims of FLSI are able to become "supporting members" by paying an annual fee of (one or more units of) 500.000 Yen (ca. 2600 ECU), obtaining various privileges concerning the above mentioned activities according to the number of units paid. -------------END OF REPORT---------------------------------------------