Lapas attēli
PDF
ePub

these companies have successfully equated “fuzzy" and "neuro-fuzzy" with "high technology" in the minds of the Japanese consumers, and one now finds a multitude of products available bearing these labels. For example, while in Tokyo's Akihabara district--the veritable consumer-electronics capital of the world--I noticed one store that had “fuzzy" and "fuzzy-neuro" washing machines from Hitachi, Toshiba, and Sharp. Other products using fuzzy technology were mentioned in a previous section.

My hosts at MEI were Noboru Wakami, Manager, Central Research Laboratories; Yoshiro Fujiwara, Director, Intelligent Electronics Laboratory, and Hideyuki Takagi, Senior Researcher, Central Research Laboratories; together with several other researchers. At the early part of the visit, I was taken on a tour of MEI's Hall of Science and Technology, where I saw displays of its latest products. Those noted as employing fuzzy technology were the fuzzy washing machine and the Palmcorder video camera, discussed earlier.

The Palmcorder is the first video camera to offer an image-stabilization capability within the consumer price range. It uses a small microprocessor to periodically compare the current video image with a prior one, and by means of fuzzy inference rules adjusts the position of the current image in response to the amount by which the two images differ along certain key (fuzzy) parameters. Matsushita also developed the first commercially available air conditioner (cooler and heater) to use fuzzy logic. And a combination of fuzzy rulebased reasoning and neural nets is used for setting the tension in the drives for its video tape recorders.

At the laboratory, I was given a demonstration of recent experiments with an inverted pendulum controlled by a neural net. This shows that neural nets can in some respects replace a

fuzzy ruled-based controller. There also has been a concerted effort toward integrating fuzzy rule-based systems with neural nets, largely due to the influence of Takagi. There are various ways in which these two reasoning paradigms may be intertwined. One is to use a neural net approach to "tuning" the membership functions for a fuzzy controller. As mentioned earlier, similar work has been carried out by Hitachi. One demonstration of a system using this approach involved tracking and hitting a moving target. Another was what has been called the Cerebellar Model Arithmetic Computer (CMAC), which uses neural nets to extract fuzzy inference rules from databases. Athird was the use of neural nets to provide exact definitions of the fuzzy-logical connectives in an information retrieval system. A fourth used fuzzy reasoning to speed up the learning process in conventional neural nets. A fifth was to implement a set of fuzzy rules as a type of neural net wherein the relation between a rule's premises and conclusion is expressed by a connection between neural net units. (It is perhaps worth mentioning that similar work is currently in progress at Florida State University, this author's home institution).

Another project underway at MEI concerns a multi-step fuzzy reasoning shell. This amounts to an effort to create a fuzzy logic controller with a more sophisticated reasoning component of the kind discussed above. The work is being undertaken in consultation with Professor Motohide Umano of Osaka University and is still fairly much in the idea stage.

Matsushita is clearly investing for the long term in fuzzy technology, to the extent of training personnel specifically for R&D in this area. Hideyuki Takagi in particular has been sent for a year to study at Professor Zadeh's new institute at UC Berkeley. Matsushita also has contributed both funding and personnel at LIFE.

Hitachi Corporation, Systems Development Laboratory, Kawasaki

My hosts at Hitachi were Singi Domen, General Manager; Motohisa Funabashi, Chief Researcher; Seiji Yasunobu, Senior Researcher; and Akira Maeda, Researcher. I was shown three presentations. The first was a "self-tuning algorithm for fuzzy membership functions using computational flow network," which was mentioned in the foregoing section on fuzzy controllers, i.e., the technique developed by Maeda for defining the membership functions of the terms appearing in the inference rules of a fuzzy logic controller. The algorithm is claimed to work efficiently for systems with up to several thousand fuzzy rules.

The second presentation was of Hitachi's expert system shell, ES/ KERNEL, which has now sold over 2,000 copies. This incorporates fuzzy logic into a reasoning system that employs frames, rules, and meta rules. The system offers a very convenient graphics interface for defining membership functions and for specifying the fuzzy frames and rules. Many expert systems built with this shell are currently in use, and work toward expanding the shell's reasoning capabilities is ongoing. The Plant Operating System project at LIFE, in fact, is being directed by an employee of this laboratory.

Third were video presentations by Seiji Yasunobu of the two projects for which he is most well known: the famous Sendai Railway and an automated crane. Both of these employ the fuzzy control methods discussed above.

As with Matsushita, Hitachi has numerous "neuro-fuzzy" products available on the consumer market and is investing for the long term in fuzzy technology. Hitachi also is a major contributor to LIFE. Minoru Yoneda, the director of LIFE's Plant Operating System project, is currently on leave

from this laboratory. Hitachi's other interests include industrial applications of fuzzy reasoning and financial applications for supercomputers.

Kansai University, Department of Industrial Engineering, Osaka

From 10 May through 9 July I held a position as Visiting Researcher at Kansai University. My hosts were Yoshinori Ezawa, Associate Professor; Ikuo Itoh, Chairman of the Department of Industrial Engineering; Sanji Nishimura, Dean of the Faculty of Engineering; and Akio Ohnishi, University President. Everyone I met was most hospitable. Professor Ezawa personally accompanied me to most of the places I visited in the Osaka-Kyoto area and was exceedingly helpful in making many introductions. He is to be thanked especially for his kindness.

Kansai University is a private institution where the faculty have fairly heavy teaching loads, in spite of which many are also productive in research. My friend Ezawa is one of those. His work crosses a number of different topics, partly through an ongoing collaboration with Motohide Umano of Osaka University. His current interests include a new, computationally simpler method of representing fuzzy linguistic hedges (e.g., very, rather, extremely) and fuzzy relational databases. His prior work also concerned linguistic hedges, as well as the general problem of fuzzy inference. This stemmed from his work as a doctoral student under Masaharu Mizumoto (currently at Osaka ElectroCommunications University).

Others at Kansai University who are working in fuzzy systems and closely related topics are as follows. Takafumi Fujisawa works on image processing by means of fuzzy feature extraction. Toshihiro Fujii, Noriaki Muranaka, and Shigeru Imanishi have developed simulations of fuzzy min-max circuits and have also experimented with

stabilization of an inverted pendulum. Noriaki Muranaka and Shigeru Imanishi have worked on ternary logic circuits.

Osaka University, Osaka

There are several people at Osaka University working in fuzzy systems. My contact there was Motohide Umano, who I have known personally for several years and who I met with in Japan at Kansai University. In addition to the work discussed above, Umano has studied fuzzy production systems, an implementation of fuzzy reasoning in Lisp, a form of fuzzy Prolog, and an expert system for damage assessment of reinforced concrete bridges.

Another interesting project I became aware of at Osaka University was in the Department of Naval Architecture and Ocean Engineering. This was a proposal by Kazuhiko Hasegawa and Hiroshi Yamakawa to use a fuzzy logic controller to reduce seasickness on large ships. The idea is to mount the ship's deck on hydraulic lifters, fastened into the hull, which would then move the deck in compensation for the roll of the sea. This same idea could conceivably also be used to stabilize gun turrets and missile launchers on battleships.

Also of interest is Toshiomi Yoshida's use of fuzzy pattern recognition in the control of fermentation processes. This is being carried out in collaboration with Konstantin Konstantinov of Japan's International Center of Cooperative Research in Biotechnology.

Kobe University, School of Engineering, Kobe

My contact at Kobe University was Hiroshi Kawamura in the Department of Architecture, who I met on 18 May at the meeting of the Kansai Branch of SOFT. Kawamura's work has concerned the use of fuzzy logic in two different problems related to earthquakes: (1) the prediction of earthquake ground

motions and structural responses and (2) the control of motion in civil engineering structures. Some of this work has been done in collaboration with James Yao of Texas A&M University.

Others working at Kobe University are: Shinzo Kitamura, stability analysis of fuzzy logic controllers; Ayaho Miyamoto, fuzzy expert systems for safety evaluation, maintenance, and rehabilitation of concrete bridges; Keizo Nagaoka, measurement of student response time in computer-aided instructional systems; and Naoyuki Tamura, fuzzy Prolog based on intervalvalued fuzzy sets.

Kyoto University, Department of Precision Mechanics, Kyoto

At Kyoto University I met with Tetsuo Sawaragi and Osamu Katai. Professor Katai is head of a small laboratory, which includes Sawaragi as well as Sawaragi's former doctoral advisor, Sosuke Iwai.

Most of the activities at Katai's laboratory are directed toward the more general issues surrounding knowledgebased reasoning systems, and fuzzy logic is viewed as a tool for addressing some of the associated problems. One project has used a fuzzy logic controller in conjunction with standard AI techniques for robot obstacle avoidance and motion planning. In addition, Katai and Iwai have studied representations of belief and plausibility in conjunction with problems of nonmonotonic reasoning.

Other projects have grown out of Sawaragi's doctoral studies in intelligent decision support systems. That work is akin to earlier work in "structural modeling" of complex societal systems, wherein the causal interrelations between the system's variables are displayed in a directed graph, e.g., Movement of Business Into Cities --> Increased Business Use of Automobiles and Truck --> Increased Traffic Problems --> Movement of Businesses

Back Out of Cities. In Sawaragi's thesis such diagrams are called "cognitive maps." More recent work has developed much more complicated types of events (problem, request, enablement, revenge, etc.) suitable for modeling international political-economic systems. These use fuzzy linguistic terms for characterizing properties of such events and have entailed developing new methods for extracting fuzzy sets from raw data.

Osaka Prefecture University, Department of Industrial Engineering, Osaka

My hosts at Osaka Prefecture University were Professor Hideo Tanaka together with several members of his laboratory and one of his former doctoral students. Professor Tanaka's primary research topic has been fuzzy (or possibilistic) linear regression analysis, although he has also touched on numerous other topics through collaboration with his coworkers.

Further research activities include the following. Hidetomo Ichihashi has studied the mathematical properties of fuzzy numbers, has investigated the use of neural nets for learning fuzzy set membership functions from small sets of training examples, and has worked with Tanaka on a type of fuzzy inference based on the Dempster-Shafer theory of evidence. Hisao Ishibuchi has developed an approach to fuzzy data analysis by means of neural nets and has worked with Tanaka in the area of fuzzy multi-objective programming. Masahiro Inuiguchi and Yasafumi Kume have also worked in fuzzy mathematical programming. Tanaka's former student, Koji Izumi (now at Hannan University in Nara), works on the formal aspects of fuzzy logic, being concerned with fuzzy quantifiers, representations of linguistic hedges as modal operators, and a kind of fuzzy reasoning based on intuitionistic logic.

Osaka ElectroCommunications University, Department of Management Engineering, Osaka

I visited Professor Masaharu Mizumoto, who is well known for his many publications in fuzzy systems. These date from his theoretical work with the late Professor Tanaka (his doctoral advisor) on the algebraic properties of fuzzy numbers, published in 1975. He happens to have been Yoshinori Ezawa's doctoral advisor while at Osaka University.

Mizumoto currently studies fuzzy logic controllers and gave a demonstration of his own version of the inverted

pendulum experiment. His approach pendulum experiment. His approach differs from that of Yamakawa in that he is able to accomplish stabilization with an ordinary 25-MHz personal computer. This is accomplished by precomputing all the possible inference results and then storing them in a table in main memory, for fast look up. The drawback of this approach is that one loses much of the flexibility (robustness) of a system in which the inferness) of a system in which the inferences can be computed in real time by ences can be computed in real time by specially dedicated chips. Mizumoto's device is useful, though, for experimenting with different rule sets and determining which types of rules give the best results. He tends to believe that the min-max compositional rule of inference (as computed by Yamakawa's chip) is not necessarily the best method. His experiments show good results using product-sum rules.

He also continues to delve into theoretical problems and recently has developed a method of linguistic evidence combination (i.e., conjoining rules whose conclusions all involve terms related to the same general property) involving a kind of linguistic interpolation. This may be useful in both fuzzy controllers and fuzzy expert systems.

Osaka Institute of Technology, School of Industrial Engineering, Osaka

Here my hosts were Kiyoji Asai and Junzo Watada. Professor Asai was among the very first Japanese to study fuzzy sets, having taken an interest in it shortly after Zadeh's first publication on the subject in 1965. He was founding president of SOFT and is well known for his many publications. Professor Asai is now one of the leading figures in the international fuzzy systems community.

Professor Watada also has many contributions to fuzzy systems research and currently serves as president of the Kansai Branch of SOFT. His recent work has concerned diagnostic expert systems using Dempster-Shafer theory.

Meiji University, Faculty of Engineering, Kawasaki

I met with Professor Masao Mukaidono, another well-known member of the international fuzzy community. His research over the past several years has focused primarily on developing a fuzzy Prolog. Most of the work to date has dealt with an appropriate adaptation of linear resolution (the essential component of standard Prolog). This is still in the theoretical stage, although there have been some prototype implementations.

In addition to his work on fuzzy Prolog, Mukaidono has investigated the properties of fuzzy interval logic, presumably because it is computationally simpler than full-fledged fuzzy logic. Interval logic may eventually become the basis for an alternative form of fuzzy Prolog. He has also studied the use of neural nets for learning fuzzy inference rules and has done rather extensive work on the subject of fuzzy switching functions.

Hosei University, Department of Instrument and Control Engineering, Tokyo

At Hosei University I met with Kaoru Hirota and saw video demonstrations of robots developed in his laboratory. One of these is a general-purpose robot that is now being manufactured by Mitsubishi. The first video showed an early version of this robot, presented at IFSA-87, playing two-dimensional pingpong with a human opponent. The experiment used a video camera to track the location of the ball and a collection of 30 inference rules to determine the hitting position. The movement of the robot was very slow and deliberate but nonetheless surprisingly accurate considering that it was being controlled by a 16-bit, 5-MHz NEC personal computer. A second demonstration of the robot showed it doing a moderately good job of traditional Japanese flower arrangement, using a collection of between 30 and 50 rules. Third was a tape of the robot playing yo-yo. Here a video camera watched the movement of the yo-yo against a white background, and the robot moved accordingly.

A final video showed a more recent experiment, involving a fuzzy logic controlled robot that throws darts at a falling object. The more dramatic segment showed the robot throwing at an object falling through an array of pegs, like in a pinball machine, and scoring a hit on virtually every try. There also have been some attempts to program a robot to grasp a spherical object moving erratically in two dimensions. The results here have so far been only marginally satisfying, with the robot missing the object more often than successfully grasping it.

generalization of the binary logic J-K flip-flop, which is the basis for most conventional digital computers. An important property of Hirota's fuzzy flip-flop is that it embodies the binary flip-flop as a special case. Thus with this device one should be able to program both binary and fuzzy logic on the same machine. There reportedly has been discussion at LIFE about starting a project to advance this idea.

Tokyo Institute of Technology (TIT), Departments of Intelligence Science and Systems Science, Yokohama

At TIT I met with Shigenobu Kobayashi in the Department of IntelKobayashi in the Department of Intelligence Science and Anca Ralescu in the Department of Systems Science. The latter is also the home department of Michio Sugeno, whom I had met on 5 August at LIFE.

Because Kobayashi's work is almost exclusively concerned with case-based reasoning, it will be discussed in the section below. Professor Ralescu is an American, who was on a visiting appointment in Sugeno's laboratory, and starting in mid-August she will be visiting at LIFE for a year under a grant from the U.S. NSF. One of her interests is a model of fuzzy reasoning based on possibility and necessity measures. She possibility and necessity measures. She has also studied the use of neural nets for learning fuzzy set membership functions.

Of special interest is Sugeno's current work, mentioned previously, on a voice-controlled helicopter. The objective is to build a helicopter that can be tive is to build a helicopter that can be flown by simple voice commands, such as "hover," "forward," "back," "left," "right," "up," "down," and "land." One potential application would be an unmanned helicopter used for sea rescue, controlled by voice commands from a mothership.

Another interesting project, mentioned earlier, is design of a fuzzy flipflop circuit. This work shows promise of leading to the first "fuzzy memory" and conceivably can serve as the basis At the time of our meeting, Sugeno for a general-purpose fuzzy computer. had already successfully automated all In essence, Hirota's fuzzy flip-flop is a the necessary basic functions for a

1-meter model helicopter, as well as the hovering function for a 3-meter model. In both models, each function is automated by means of a fuzzy logic controller composed of a relatively small set of inference rules. For example, hovering requires 36 rules, broken down into one group of 18 main rules (to stabilize position and speed) and one group of 18 subrules (to stabilize attitude), where each of these is further decomposed into simpler groups of 2 to 6 rules. Ordinary flight control requires another 30 rules, 14 for speed and 16 for attitude, and rotation requires 28 rules, 14 for speed and bank angle and 14 for attitude. In total 15 input variables need to be monitored. Nine of these are the current value, rate of change, and acceleration of pitch, roll, and yaw. Others include altitude and various speeds. Outputs are instructions to the helicopter's usual controls, i.e., stick, pedals, etc.

In reviewing the outline for Sugeno's controller, one cannot help but be struck by its relative simplicity. Although there are in total more than 100 rules, each lowest-level rule group is logically independent of all the others and can be treated conceptually as a selfcontained unit. Thus overall design and tuning should be fairly easy. Realtime tests of the 1-meter model showed that the fuzzy logic controller actually does better than a trained human, i.e., it reaches stability in the hovering operation more quickly. It also did very well on the other functions. Given successful completion of tests with the 3-meter model, they will next seek funding to work on a full-size helicopter.

Tokyo Electro-Communications University, Department of Communication Systems, Tokyo

My last visit was with Nakaji Honda, where I learned of a project to control the lighting of a room by means of a hierarchically organized rule-based fuzzy

controller. The aim is to produce either CASE-BASED REASONING

uniform brightness or special effects in certain parts of the room, so as to provide a physiologically and psychologically comfortable atmosphere. The problem is complicated by requiring that the system be able to control interior lighting in such a way that it supplements and/or cooperates with external natural light, which is constantly changing. While this research focuses on an application of seemingly limited utility, it apparently is directed toward a much larger and important application, namely, to automate the control of large power distribution systems in response to changing loads.

Hiroshima Institute of Technology, Faculty of Engineering

From the May 1991 issue of the Journal of the Japan Society for Fuzzy Theory and Systems, I learned of some recent work by Kazuho Tamano on an optical fuzzy inference system. In this system, graphs of the membership functions for a rule's fuzzy linguistic terms are etched onto transparent plates, a light is passed through the plates, focused by a set of Fresnel convex lenses, and the amount of light coming out is measured by a set of photo diodes. These measurements yield the rule's inferred conclusion, in accordance with Mizumoto's product-sum-gravity method of fuzzy inference. The inference is thus virtually instantaneous, although multiple conclusions still need to be combined by an electronic defuzzifier.

This appears to be the very first report of such an inferencing device. The results are still very much in the experimental stage, but the possibilities are intriguing. Future experiments are planned wherein the etched plates are replaced with programmable lightemitting diode (LED) displays.

Case-based reasoning (CBR) is largely the invention of Christopher largely the invention of Christopher Riesbeck of Northeastern University and Roger Schank of Yale. The underlying intuition is that much of human reasoning and decision making is derived from remembered personal experiences. The term "cased-based" reflects a focus on areas like medicine and law, where remembered experiences can be represented as individual cases. The basic idea is to develop a database of cases that can be accessed for the purposes of a current problem situation. The given situation is analyzed for its relevant properties, and based on these properties, any similar cases are retrieved from the database and modified or other

wise utilized for the present purposes. This produces a new case that is then added to the database for future reference. The CBR paradigm is in this manner also a learning paradigm, building its knowledge base through ongoing interactions with the user. This feature is attractive, as it offers a way to get beyond the knowledge acquisition bottleneck inherent in most other AI paradigms.

An early investigation of the methods for implementing such a system was carried out by Schank's doctoral student, Kristian Hammond, where the cases were recipes for Chinese cooking. Hammond's CHEF served as a vehicle for exploring the possibility of fully automating all the phases of casebased reasoning, i.e., case retrieval, modification, storage, and repair. CBR has subsequently attracted a sizable following in the West, and for each of the last 3 years there has been a workshop on the subject sponsored by the Defense Advanced Research Projects Agency (DARPA).

During my stay in Japan, I uncovered only two places where there is any significant activity in CBR. The main one is Shigenobu Kobayashi's laboratory

at the Tokyo Institute of Technology. Kobayashi is well known for his prior work in AI, having written rather extensively on the subject of machine learning. His work has explored the possible theoretical connections between explanation-based learning and CBR, and he and his coworkers had undertaken applications in three areas: (1) an expert system for configuring a cigarette rolling machine, (2) production process diagnosis, and (3) information retrieval. The work on (1) features an "interactive" CBR system, which foregoes the attempt at full automation and provides that the user play a role in each phase of the CBR process. This is reasonable, as one should not expect to be able to fully automate CBR for problems much more complex than simple cooking recipes.

The other focus of activity in CBR is Osamu Katai's laboratory at Kyoto University. Here the main proponent appears to be Tetsuo Sawaragi, working in collaboration with Katai and Sosuke Iwai. Projects in this area are currently only in the planning stages, with proposed application domains similar to those undertaken by Kobayashi, particularly diagnosis and information retrieval.

From my meeting with Seiji Yasunobu at Hitachi, I learned that there have been recent discussions at his laboratory about studying CBR, but that it was not clear to them how or where these principles could be applied. Having just learned that Hitachi is beginning a program to develop and market a package of software development tools, I suggested the possibility of using CBR in software engineering, for both software system design and software development project planning. In the former, cases are former designs, and in the latter, cases are plans for implementing those designs. His ini

tial reaction was that this seemed like a reasonable idea. It is likely that Japanese involvement in CBR will continue to grow.

« iepriekšējāTurpināt »