Lapas attēli
PDF
ePub

computations to form the rules actually appearing in his controller. The reason for this approach is that the compositional rule involves a matrix operation that cannot normally be performed fast enough on a standard digital computer. In Yamakawa's system, this problem was overcome by designing a chip specifically for this computation. An important feature of Yamakawa's approach, moreover, is the use of analog techniques, rather than digital. This was done because the elementary operations employed in the compositional rule of inference, and for the most part also in the defuzzification operation, are the arithmetic max and min, which can be implemented so as to run much faster on an analog device.

The controller presented at IFSA-87 used seven rule chips and one defuzzifier chip, and it demonstrated balancing response speeds approaching 100 times faster than those heretofore accomplished by a conventional PID controller. This result generated a flurry of commentary, including a few negative responses both from without and within the fuzzy systems community. The latter stemmed from the fact that the controller only maintained vertical, and not horizontal, stability of the inverted pendulum, whereas the classical problem entails both. Moreover, it was shown rather easily that, with that particular system, accomplishing both was impossible. Hence Yamakawa suffered criticism for publishing results that were as yet incomplete.

Less than a year later, however, Yamakawa was able to vindicate himself by producing a system with only four additional rule chips that performed

both vertical and horizontal stabiliza

tion at the same speed as before. Since that time, Yamakawa has demonstrated the robustness of his system for nonlinear control by attaching a small platform to the top of the inverted pendulum, on which is then placed a wine glass filled with liquid, or even a live white mouse. The controller nicely

compensates for the turbulence in the liquid, as well as the totally erratic movements of the mouse. Thus in the latter, a claim could be made for executing control even beyond nonlinearity and into truly random or “chaotic" domains.

Before reporting these results, Yamakawa applied for patents on his chips in Japan, the United States, and several European nations. He then proceeded to trade his patents to several Japanese corporations in return for their subsidizing a laboratory in which he could continue his research. (Japanese university professors are not allowed to make money outside of their duties as a faculty member.) Omron, a major producer of second tier electronic devices, was a major proponent and has subsequently decided to invest heavily in fuzzy control. Omron has been rapidly expanding on Yamakawa's original designs, producing a host of new chips, both analog and digital, and churning out scores of applications. Due to its purchase of Yamakawa's patents, in fact, it has recently become the first Japanese corporation to ever obtain a U.S. patent. As of July 1991, Omron boasted 700 patents for fuzzy logic devices either acquired, pending, or in application. Most of these devices either have appeared, or will appear, in commercial products. Three or four dozen alone are earmarked for use in automobiles, e.g., antilock brakes, automatic transmissions, impact warning and monitoring, windshield washers, and light dimmers. Omron is also incorporating fuzzy control into products for use in industrial and manufacturing processes.

Numerous commercial products using fuzzy technology are currently available in Japan, and a few are now being marketed in the United States and Europe. Canon uses a fuzzy controller in the autofocus mechanism of its new 8-mm movie camera. The Matsushita/Panasonic "Palmcorder," currently being promoted on U.S.

television, uses fuzzy logic for image stabilization. This happens to be the very first video camera to appear with image stabilization capability. Matsushita, Hitachi, Sanyo, and Sharp now have their own “fuzzy washing machine," which automatically adjusts the washing cycle in response to size of load, type of dirt (soil versus grease), amount of dirt, and type of fabric. In Matsushita's machine, the type and amount of dirt are detected by means of light sensors, which also use fuzzy controls. Other products using fuzzy control include vacuum cleaners, air conditioners, electric fans, and hot plates. One senses that the possibility for such applications is virtually endless. Another, now famous, application is a road tunnel ventilation system, also designed by Yasunobu at Hitachi.

A somewhat more ambitious project is the voice-controlled helicopter being developed by Michio Sugeno at the Tokyo Institute of Technology. Here the objective is to develop a helicopter that responds to voice commands like "hover," "forward," "back,” “left," "right," "up," and "down," where each such operation is handled automatically via fuzzy logic. Sugeno has successfully accomplished all functions with a 1-meter model and is now working on a 3-meter model. As of August, he had achieved hovering and was confident that the other operations could be accomplished as well. Hovering is well known to be a very difficult stability problem; beginning helicopter pilots typically train for weeks before being able to do this manually. Hence, successfully automating the hovering operation is in itself a very impressive result. [See the article by D.K. Kahaner and D.G. Schwartz, “Fuzzy Helicopter Flight Control," Scientific Information Bulletin 16(4), 13-15 (1991).]

These few examples illustrate the variety of possible applications for fuzzy logic control. Japanese manufacturers are in fact now opting for fuzzy coneven where conventional trollers

controllers would serve just as well. The reasons are that simple fuzzy logic controllers are much easier to design, require fewer electronic components, and are therefore cheaper to produce.

The problem of how to design more complex controllers, however, has only recently met with what appears to be a practical solution. Typically the most difficult part of designing any fuzzy logic controller lies in selecting the fuzzy sets to use for the meanings of the linguistic terms appearing in the inference rules. As the number of rules grows large, the trial-and-error method of selecting the optimal collection of membership functions becomes less feasible. Somewhat of a breakthrough on this problem appears to have been achieved by Akira Maeda at Hitachi's System Development Laboratory. Maeda's idea is to use a form of neural net with back propagation to learn the needed membership functions from a set of training examples. As a test case, Maeda and his coworkers applied this technique to the development of a controller that had been designed previously by trial-and-error. Using this technique, they were able to accomplish in 1 month what had formerly taken 6 months.

A common opinion among Japanese researchers is that most of the important theoretical work in fuzzy control has now been completed and that the next step is up to the commercial manufacturers, i.e., to start churning out applications. This is reflected, for instance, in the fact that fuzzy logic control was one of the three major areas of focus in the original program at the Laboratory for International Fuzzy Engineering Research (LIFE), whereas 3 years later, we find that the subject is barely mentioned within its current program. What is perhaps more correct, however, is that only an initial stage of theoretical development is now more or less complete, and moving to the next stage will require solving an assortment of substantially more difficult

problems. Therefore, from a purely pragmatic standpoint, it makes sense to focus on reaping the commercial benefits of what has already been done and to leave for the future the more challenging theoretical issues.

The possibilities for future work, leading to far more sophisticated logic-based controllers, are nonetheless very clear. This will amount to moving from simple one-step rule-based systems to systems employing multi-step reasoning--i.e., rule chaining, together with the necessary truth maintenance systems--which are integrated with other knowledge representation, reasoning, and learning schemes (e.g., semantic nets, frames, conceptual graphs, neural nets, and case-based reasoning). Taking the theory to this next stage will accordingly require progress in a number of important subareas before realizing the more advanced levels of automatic control.

OTHER FUZZY

SYSTEMS RESEARCH

As may be seen from the attached bibliography, current fuzzy systems research encompasses a large variety of topics, far too numerous to be covered thoroughly in this short report. Therefore, the following sections will focus rather on the organizations, research institutes, universities, and corporate laboratories with which I became familiar during my visit. Wherever appropriate, this includes a brief survey of the principal researchers and research activities.

The Japan Society for Fuzzy Theory and Research (SOFT)

SOFT was founded and held its first annual meeting in 1984, at which time it had about 20 members. As of 1991, membership amounts to 1,800 individuals and 100 corporations. The current president is Professor Michio Sugeno of the Tokyo Institute of Technology

and the vice president is Hideo Tanaka of Osaka Prefecture University.

The society has a few regional chapters, and on 18 May I attended the meeting of the Kansai Branch in Osaka. There I noted from the talks given that the fuzzy systems group in Japan is not completely focused within that area but also overlaps with the broader realms of AI. Two talks were given, one on neural nets by Kazuyoshi Tsutumi of Ryukoku University and one on casebased reasoning by Tetsuo Sawaragi of Kyoto University.

During 12-14 June 1991, I attended and gave a presentation at the seventh annual SOFT symposium, held in Nagoya. Total attendance was 500, of which approximately 100 were from corporations. There were 165 papers presented, covering the entire spectrum of fuzzy systems research, but with a lessor emphasis on theory than on applications. Most Japanese researchers involved in fuzzy systems research are engineers, so that most presentations were in the engineering disciplines. Recognition of the importance of theoretical work was evident, however, by the invitation of Satoko Titani (Chuba University) to give a plenary talk about her work with Gaishi Takeuchi on formalizations of fuzzy logic.

A plenary talk by Masao Mukaidono titled "Fuzzy' and AI" revealed that the Japanese AI community has followed the lead of its American counterpart in adopting a somewhat hostile attitude toward fuzzy systems research. I was told, however, that the controversy in Japan is not as severe as in the West. Some speculations on why this is the case are taken up in the concluding section of this report.

Laboratory for International Fuzzy Systems Engineering Research (LIFE), Yokohama

LIFE is a 6-year project (1 April 1988 through 31 March 1995) funded by the Japanese Ministry of International

Trade and Industry (MITI) in conjunction with 49 major Japanese corporations. Its stated objectives are (1) to promote research and development (R&D) on applications of fuzzy theory to engineering and (2) to promote domestic and international exchange on the study of fuzzy theory. The managing director is Toshiro Terano, Professor of Control Engineering at Hosei University. Professor Terano is well known as one of the first Japanese researchers in fuzzy systems.

At its inception, LIFE was organized into three "laboratories”: (1) Fuzzy Control (especially for production processes and robots), (2) Fuzzy Intellectual Information Processing (decision support systems, image understanding, expert system shells, diagnosis system for power station, language understanding for robots, and evaluation and understanding of numerical information), and (3) Fuzzy Computer (including system architecture, hardware, and software).

In 1990, however, LIFE's goals and research emphasis underwent some fundamental changes, largely because by that time fuzzy technology had found widespread use in industrial and commercial products. Most significantly, it was decided that the subject of fuzzy control had advanced to sufficient maturity that there was no further need for a laboratory on that topic; it was felt that the necessary theoretical work was largely complete and that it was next up to the corporations to start producing applications. As a result, the projects were reorganized into three groups, having three projects each:

[blocks in formation]
[blocks in formation]

of some corporation and producing a series of 8 or 10 publications. The publications then count in place of a doctoral dissertation. Many of the people at LIFE are currently involved in this process. The advantage to the participating companies is that their employees in this manner gain good research experience, together with a solid grounding of knowledge in fuzzy technology, which is then brought back to the corporate research laboratory.

The senior researchers at LIFE are more experienced corporate employees or university professors who visit LIFE on a part-time basis and serve mainly an advisory role. The institute also encourages both short and extended visits by foreigners. Other important functions include organizing national and international conferences and seminars. LIFE has become the de facto hub of activity and communications regarding fuzzy systems research in Japan.

The Science and Technology Agency (STA)

The STA is roughly the equivalent of the U.S. National Science Foundation (NSF), funding primarily universityrelated research. (STA is also a small contributor to LIFE, but only at 1/20 the amount provided by MITI.) In 1989, STA initiated a program titled “Fuzzy Systems and Their Application to Human and Natural Problems," which began funding fuzzy systems research at the rate of around ¥200 million (or $1.5 million) per year. As such, the actual level of funding provided by STA for fuzzy systems research is much smaller than MITI's contribution to LIFE, but it actually supports a great deal more research activity. This is because STA has no need to provide salaries. Japanese university professors are paid year-round by their institutions and are normally provided at least minimal research facilities. Thus grants from STA cover only funding for

equipment, travel to meetings, businessrelated entertainment, books, and other incidental expenses.

(15) Application of Fuzzy Logic to Social and Management Systems

During the summer of 1991 there (16) Earthquake Forecasting were 18 projects being funded under

this program, and the word at that time was that the program's budget was soon to be enlarged. The list of project titles is as follows:

(1) Fuzzy Logic

(2)

(3)

(4)

inverted pendulum experiment. Construction on a new building, planned to house 40 researchers plus administrative personnel, had been scheduled to begin very soon and was expected to be

(17) Prediction of Air Pollution in completed in 1992. The institute also

Wide Areas

(18) Modeling of Plant Growth

Fuzzy Logic Systems Institute (FLSI), Kyushu Institute of

Algorithm of Fuzzy Reasoning Technology, lizuka, Kyushu

Programming Language and Architecture

Intelligent Control of HighSpeed and Unstable Systems (helicopter)

[blocks in formation]

FLSI was established in March 1990 to conduct experimental research into fuzzy information processing and neuroscience and to promote the wider use of the scientific findings in these domains. The chairman and person primarily responsible for its creation is Professor Takeshi Yamakawa, whose work on fuzzy logic controllers was discussed in a foregoing section.

The initial budget for FLSI was ¥100 million ($750,000) provided by 13 private corporations working in collaboration with the Kyushu Institute of

Recognition of Handwritten Technology (KIT) and Fukuoka PreLetters

[blocks in formation]

fecture. From the prefecture's standpoint, this is part of a long-term effort to establish a new center of technological industry in the Iizuka area, which formerly was a coal mining community and is now economically relatively depressed. The local campus of KIT was itself established for this purpose,

and Yamakawa's move there from his former position at Kumamoto University was largely to enable his participation in the formation of FLSI.

New participants have continued to join, most notably Omron Corporation, which as note earlier has traded its support in part for the rights to Yamakawa's patents. At the time of my visit, the institute consisted of three researchers, an administrative director, and two staff assistants, in a small temporary building near KIT. I was given a demonstration there of the given a demonstration there of the

(13) Fuzzy Association (voice researchers, an administrative direc

recognition)

(14) Evaluation of Reliability of Large-Scale Systems (safety)

produces the new Journal of the Fuzzy Logic Systems Institute, which currently is published only in Japanese.

In addition to Yamakawa's work on the fuzzy inference and defuzzifier chips, he has recently developed a fuzzy neuron chip and demonstrated its utility by means of an application to pattern recognition. A device has been constructed that correctly recognizes handwritten characters with the same degree of accuracy as prior devices, but with much greater speed. The results of these experiments were to be presented at the International Fuzzy Expert Systems (IFES) conference in November 1991.

Other fuzzy systems researchers at KIT are Toyohiko Hirota and Torao Yanaru, together with their student Tomokazu Nakamura. Some of their work concerns the theory of fuzzy inference and its applications in expert systems. In addition, they have been exploring the use of projective geometry to represent various properties of fuzzy systems, a study initiated recently in the United States by Bart Kosko.

Omron Corporation,
Fuzzy Technology Business
Promotion Center, Kyoto

Omron is primarily a second tier corporation, producing components that other manufacturers use in products for the commercial markets. It also makes equipment for large manufacturers. Omron invested early in fuzzy technology and is now the world's leading innovator in the creation of fuzzy logic devices. A brief chronology follows below. Some of this has been extracted from a 9 May 1991 report by Thomas Hagemann of the German National Research Center for Computer

Science, Tokyo. Additional information has been derived from my own visit. My hosts were Masaki Arao and Satoru Isaka.

In 1966, Omron developed the DECIVAC, an analog computer that implemented a type of probabilistic decision making and was to a certain extent intended to address the same variety of problems as was fuzzy sets. It is interesting that this work almost exactly coincided with Zadeh's first publication in this area.

In 1983, when the first fuzzy technology appeared in Japan (control of a drinking water treatment plant by Fuji Electric and the Sendai Railway project by Hitachi), Takeshi Yamakawa visited Omron's Tokyo office in search of financial support for his fuzzy integrated circuit (IC), a hand-made sample of which he had completed in his laboratory at Kumamoto University in October of that year. Omron saw the potential for this new technology, and in October 1984, Yamakawa came to Omron's head office in Kyoto for a lecture, which was also attended by Kazuma Tateishi, the founder of Omron. He showed a deep interest in fuzzy technology and that year took control of Yamakawa's patent ideas and began developing fuzzy hardware under Yamakawa's supervision. They also at this time began R&D on fuzzy expert systems. By 1986, they had produced fuzzy hardware and a medical diagnosis expert system.

The fuzzy boom in Japan began in 1987. As mentioned earlier, this was the year Yamakawa presented his inverted pendulum at the Second Congress of the International Fuzzy Systems Association. The chips used in that demonstration had been built at Omron. Also during this year, Omron built a "fuzzy computer" [a fuzzy arithmetic logic unit (ALU)]. In 1988 this was marketed as the FZ-1000, and a special task force, the Fuzzy Project Team, was established within Omron.

A prototype fuzzy chip designed by Yamakawa was manufactured (now the FZ-5000), and the hybrid (Fuzzy+PID) temperature controller E5AF was developed. Omron took part in establishing LIFE, participated in the Fuzzy Committee of the STA, and received a grant of ¥600 million ($4.5 million) from the Japan Research and Development Corporation (another STA organization).

In 1989, 10 new fuzzy products were announced by Omron, and 60 fuzzy demonstrations appeared at the Omron Festival, an idea contest held regularly within Omron. Professor Zadeh became a senior advisor to Omron, and the Fuzzy Project Team was transformed into the Fuzzy Technology Business Promotion Center, while the team leader was dispatched for 2 years as head of one of the original three research laboratories at LIFE.

By 1990, Omron's total number of patent applications for fuzzy logic devices reached almost 600. These included a fuzzy human body sensor, a fuzzy expert system for machine diagnosis (together with the machinery manufacturer Komatsu), a digital fuzzy chip (the FP3000), a fuzzy inference board incorporating the digital fuzzy chip (the FB-30AT), and a new tuning method for fuzzy controllers.

As of my visit in July 1991, the total number of patent applications had exceeded 700 and included a camera that can follow moving objects, a robot with sufficient sensitivity to lift cakes of tofu, a color combination recognizer, a bottle cap recognizer, and a temperature controller for a chemical reaction plant. As mentioned earlier, plans were being laid for developing approximately 40 fuzzy logic devices specifically for use in automobiles.

Omron currently employs more that 30 engineers working in fuzzy systems R&D, with applications divided into 5 problem types:

(1) tracking problems (noisy, timevariant systems), e.g., temperature control, tension control, position control, chemical plant control

(2) tuning problems (conflicting constraints), e.g., gain tuning, crane control

(3) human factors (feelings, intuitions), e.g., cruise control, engine diagnostic systems, steering control

(4) interpolation (multi-inputs, multilevel processing), e.g., automotive air conditioner, washing machine, gas/liquid flow regulator, manufacturing device control, label identification

(5) classification (complex pattern recognition), e.g., handwriting recognition

The 31 July issue of the Daily Yomiuri newspaper announced an agreement between Omron and NEC❝to combine NEC's semiconductor technology with Omron's fuzzy logic expertise to develop fuzzy logic support systems and microprocessors." I was told that Omron expects about 30% of its total business (¥350 billion, or $2.5 billion, in sales per year) to be fuzzy related by 1995. Other manufacturers are, of course, also quickly moving into this area, and Oki Electric Industry Company has recently announced a competing fuzzy inference chip.

Matsushita Electric Industrial (MEI) Company, Central Research Laboratories, Osaka

Matsushita, also known as National/ Panasonic, is one of the world's leading producers of consumer products employing fuzzy technology. Others include Fuji Electric, Fuji Film, Hitachi, Mitsubishi Electric, Nissan Auto, Sharp, and Toshiba. Through their advertising,

« iepriekšējāTurpināt »