« iepriekšējāTurpināt »
these companies have successfully fuzzy ruled-based controller. There also Hitachi Corporation, Systems equated “fuzzy” and “neuro-fuzzy” with has been a concerted effort toward Development Laboratory, "high technology” in the minds of the integrating fuzzy rule-based systems with Kawasaki Japanese consumers, and one now finds neural nets, largely due to the influa multitude of products available bear- ence of Takagi. There are various ways My hosts at Hitachi were Singi ing these labels. For example, while in in which these two reasoning para- Domen, General Manager; Motohisa Tokyo's Akihabara district--the verita- digms may be intertwined. One is to Funabashi, Chief Researcher; Seiji ble consumer-electronics capital of the use a neural net approach to "tuning” Yasunobu, Senior Researcher; and world--I noticed one store that had the membership functions for a fuzzy Akira Maeda, Researcher. I was shown “fuzzy” and “fuzzy-neuro” washing controller. As mentioned earlier, similar three presentations. The first was a machines from Hitachi, Toshiba, and work has been carried out by Hitachi. “self-tuning algorithm for fuzzy memSharp. Other products using fuzzy tech- One demonstration of a system using bership functions using computational nology were mentioned in a previous this approach involved tracking and flow network,” which was mentioned section.
hitting a moving target. Another was in the foregoing section on fuzzy conMy hosts at MEI were Noboru what has been called the Cerebellar trollers, i.e., the technique developed Wakami, Manager, Central Research Model Arithmetic Computer (CMAC), by Maeda for defining the membership Laboratories; Yoshiro Fujiwara, Direc- which uses neural nets to extract fuzzy functions of the terms appearing in the tor, Intelligent Electronics Laboratory, inference rules from databases. Athird inference rules of a fuzzy logic conand Hideyuki Takagi, Senior was the use of neural nets to provide troller. The algorithm is claimed to Researcher, Central Research Labo- exact definitions of the fuzzy-logical work efficiently for systems with up to ratories; together with several other connectives in an information retrieval several thousand fuzzy rules. researchers. At the early part of the system. A fourth used fuzzy reasoning The second presentation was of visit, I was taken on a tour of MEI's to speed up the learning process in Hitachi's expert system shell, ES/ Hall of Science and Technology, where conventional neural nets. A fifth was to KERNEL, which has now sold over I saw displays of its latest products. implement a set of fuzzy rules as a type 2,000 copies. This incorporates fuzzy Those noted as employing fuzzy tech- of neural net wherein the relation logic into a reasoning system that nology were the fuzzy washing between a rule's premises and conclu- employs frames, rules, and meta rules. machine and the Palmcorder video sion is expressed by a connection between The system offers a very convenient camera, discussed earlier.
neural net units. (It is perhaps worth graphics interface for defining memThe Palmcorder is the first video mentioning that similar work is currently bership functions and for specifying camera to offer an image-stabilization in progress at Florida State University, the fuzzy frames and rules. Many expert capability within the consumer price this author's home institution). systems built with this shell are currange. It uses a small microprocessor Another project underway at MEI rently in use, and work toward expandto periodically compare the current video concerns a multi-step fuzzy reasoning ing the shell's reasoning capabilities is image with a prior one, and by means of shell. This amounts to an effort to create
shell. This amounts to an effort to create ongoing. The Plant Operating System fuzzy inference rules adjusts the posi- fuzzy logic controller with a more project at LIFE, in fact, is being directed tion of the current image in response to sophisticated reasoning component of by an employee of this laboratory. the amount by which the two images the kind discussed above. The work is Third were video presentations by differ along certain key (fuzzy) param- being undertaken in consultation with Seiji Yasunobu of the two projects for eters. Matsushita also developed the Professor Motohide Umano of Osaka which he is most well known: the famous first commercially available air condi- University and is still fairly much in the Sendai Railway and an automated crane. tioner (cooler and heater) to use fuzzy idea stage.
Both of these employ the fuzzy control logic. And a combination of fuzzy rule- Matsushita is clearly investing for methods discussed above. based reasoning and neural nets is used the long term in fuzzy technology, to As with Matsushita, Hitachi has for setting the tension in the drives for the extent of training personnel specif- numerous
numerous “neuro-fuzzy" products its video tape recorders.
ically for R&D in this area. Hideyuki available on the consumer market and At the laboratory, I was given a Takagi in particular has been sent for a is investing for the long term in fuzzy demonstration of recent experiments year to study at Professor Zadeh's new
year to study at Professor Zadeh's new technology. Hitachi also is a major with an inverted pendulum controlled institute at UC Berkeley. Matsushita contributor to LIFE. Minoru Yoneda, by a neural net. This shows that neural also has contributed both funding and the director of LIFE's Plant Operating nets can in some respects replace a personnel at LIFE.
System project, is currently on leave from this laboratory. Hitachi's other stabilization of an inverted pendulum.
stabilization of an inverted pendulum. motions and structural responses and interests include industrial applications Noriaki Muranaka and Shigeru Imanishi
Noriaki Muranaka and Shigeru Imanishi (2) the control of motion in civil engiof fuzzy reasoning and financial appli- have worked on ternary logic circuits. neering structures. Some of this work cations for supercomputers.
has been done in collaboration with Osaka University, Osaka James Yao of Texas A&M University. Kansai University, Department
Others working at Kobe University of Industrial Engineering,
There are several people at Osaka are: Shinzo Kitamura, stability analysis Osaka
University working in fuzzy systems. of fuzzy logic controllers; Ayaho
My contact there was Motohide Umano, Miyamoto, fuzzy expert systems for safety From 10 May through 9 July I held who I have known personally for sev- evaluation, maintenance, and rehabila position as Visiting Researcher at eral years and who I met with in Japan itation of concrete bridges; Keizo Kansai University. My hosts were at Kansai University. In addition to the Nagaoka, measurement of student Yoshinori Ezawa, Associate Professor; work discussed above, Umano has response time in computer-aided Ikuo Itoh, Chairman of the Depart- studied fuzzy production systems, an instructional systems; and Naoyuki ment of Industrial Engineering; Sanji implementation of fuzzy reasoning in Tamura, fuzzy Prolog based on intervalNishimura, Dean of the Faculty of Lisp, a form of fuzzy Prolog, and an valued fuzzy sets. Engineering; and Akio Ohnishi, Uni- expert system for damage assessment versity President. Everyone I met was of reinforced concrete bridges. Kyoto University, Department most hospitable. Professor Ezawa per- Another interesting project I became of Precision Mechanics, sonally accompanied me to most of the awareofat Osaka University was in the Kyoto places I visited in the Osaka-Kyoto area Department of Naval Architecture and and was exceedingly helpful in making Ocean Engineering. This was a pro- At Kyoto University I met with many introductions. He is to be thanked posal by Kazuhiko Hasegawa and Tetsuo Sawaragi and Osamu Katai. especially for his kindness.
Hiroshi Yamakawa to use a fuzzy logic Professor Katai is head of a small laboKansai University is a private insti- controller to reduce seasickness on large ratory, which includes Sawaragi as well tution where the faculty have fairly heavy ships. The idea is to mount the ship's as Sawaragi's former doctoral advisor, teaching loads, in spite of which many deck on hydraulic lifters, fastened into Sosuke Iwai. are also productive in research. My the hull, which would then move the Most of the activities at Katai's friend Ezawa is one of those. His work deck in compensation for the rollofthe laboratory are directed toward the more crosses a number of different topics, sea. This same idea could conceivably general issues surrounding knowledgepartly through an ongoing collabora- also be used to stabilize gun turrets and based reasoning systems, and fuzzy logic tion with Motohide Umano of Osaka missile launchers on battleships. is viewed as a tool for addressing some University. His current interests include Also of interest is Toshiomi of the associated problems. One proja new, computationally simpler method Yoshida's use of fuzzy pattern recogni- ect has used a fuzzy logic controller in of representing fuzzy linguistic hedges tion in the control of fermentation conjunction with standard Al techniques (e.g., very, rather, extremely) and fuzzy processes. This is being carried out in for robot obstacle avoidance and motion relational databases. His prior work collaboration with Konstantin planning. In addition, Katai and Iwai also concerned linguistic hedges, as well Konstantinov of Japan's International have studied representations of belief as the general problem of fuzzy infer- Center of Cooperative Research in and plausibility in conjunction with ence. This stemmed from his work as a Biotechnology
problems of nonmonotonic reasoning. doctoral student under Masaharu
Other projects have grown out of Mizumoto (currently at Osaka Electro- Kobe University, School Sawaragi's doctoral studies in intelliCommunications University). of Engineering, Kobe
gent decision support systems. That Others at Kansai University who
work is akin to earlier work in "strucare working in fuzzy systems and closely My contact at Kobe University was tural modeling" of complex societal related topics are as follows. Takafumi Hiroshi Kawamura in the Department systems, wherein the causal interrelaFujisawa works on image processing by of Architecture, who I met on 18 May tions between the system's variables means of fuzzy feature extraction. at the meeting of the Kansai Branch of are displayed in a directed graph, e.g., Toshihiro Fujii, Noriaki Muranaka, and SOFT. Kawamura's work has concerned Movement of Business Into Cities --> Shigeru Imanishi have developed the use of fuzzy logic in two different Increased Business Use of Automobiles simulations of fuzzy min-max circuits problems related to earthquakes: (1) the and Truck --> Increased Traffic and have also experimented with prediction of earthquake ground
prediction of earthquake ground Problems --> Movement of Businesses Back Out of Cities. In Sawaragi's thesis Osaka Electro
Osaka Institute of Technology, such diagrams are called “cognitive Communications University, School of Industrial maps.” More recent work has devel- Department of Management Engineering, Osaka oped much more complicated types of Engineering, Osaka events (problem, request, enablement,
Here my hosts were Kiyoji Asai and revenge, etc.) suitable for modeling I visited Professor Masaharu Junzo Watada. Professor Asai was international political-economic sys- Mizumoto, who is well known for his among the very first Japanese to study tems. These use fuzzy linguistic terms many publications in fuzzy systems. fuzzy sets, having taken an interest in it for characterizing properties of such These date from his theoretical work shortly after Zadeh's first publication events and have entailed developing with the late Professor Tanaka (his on the subject in 1965. He was foundnew methods for extracting fuzzy sets doctoral advisor) on the algebraic ing president of SOFT and is well known from raw data.
properties of fuzzy numbers, published for his many publications. Professor
in 1975. He happens to have been Asai is now one of the leading figures Osaka Prefecture University, Yoshinori Ezawa's doctoral advisor in the international fuzzy systems Department of Industrial while at Osaka University.
community. Engineering, Osaka
Mizumoto currently studies fuzzy Professor Watada also has many
logic controllers and gave a demon- contributions to fuzzy systems research My hosts at Osaka Prefecture Uni- stration of his own version of the inverted and currently serves as president of the versity were Professor Hideo Tanaka pendulum experiment. His approach Kansai Branch of SOFT. His recent together with several members of his differs from that of Yamakawa in that work has concerned diagnostic expert laboratory and one of his former doc- he is able to accomplish stabilization systems using Dempster-Shafer theory. toral students. Professor Tanaka's pri- with an ordinary 25-MHz personal mary research topic has been fuzzy (or computer. This is accomplished by Meiji University, Faculty of possibilistic) linear regression analy- precomputing all the possible infer- Engineering, Kawasaki sis, although he has also touched on
ence results and then storing them in a numerous other topics through collab- table in main memory, for fast look up. I met with Professor Masao oration with his coworkers.
The drawback of this approach is that Mukaidono, another well-known Further research activities include one loses much of the flexibility (robust- member of the international fuzzy the following. Hidetomo Ichihashi has ness) of a system in which the infer- community. His research over the past studied the mathematical properties of ences can be computed in real time by several years has focused primarily on fuzzy numbers, has investigated the use specially dedicated chips. Mizumoto's developing a fuzzy Prolog. Most of the of neural nets for learning fuzzy set device is useful, though, for experi- work to date has dealt with an appromembership functions from small sets menting with different rule sets and priate adaptation of linear resolution of training examples, and has worked determining which types of rules give (the essential component of standard with Tanaka on a type of fuzzy infer- the best results. He tends to believe Prolog). This is still in the theoretical ence based on the Dempster-Shafer that the min-max compositional rule stage, although there have been some theory of evidence. Hisao Ishibuchi has of inference (as computed by prototype implementations. developed an approach to fuzzy data Yamakawa's chip) is not necessarily In addition to his work on fuzzy analysis by means of neural nets and the best method. His experiments show Prolog, Mukaidono has investigated has worked with Tanaka in the area of good results using product-sum rules. the properties of fuzzy interval logic, fuzzy multi-objective programming. He also continues to delve into presumably because it is computationMasahiro Inuiguchi and Yasafumi Kume theoretical problems and recently has ally simpler than full-fledged fuzzy logic. have also worked in fuzzy mathemati- developed a method of linguistic evi- Interval logic may eventually become cal programming. Tanaka's former dence combination (i.e., conjoining rules the basis for an alternative form of student, Koji Izumi (now at Hannan whose conclusions all involve terms fuzzy Prolog. He has also studied the University in Nara), works on the for- related to the same general property) use of neural nets for learning fuzzy mal aspects of fuzzy logic, being con- involving a kind of linguistic interpola- inference rules and has done rather cerned with fuzzy quantifiers, repre- tion. This may be useful in both fuzzy
tion. This may be useful in both fuzzy extensive work on the subject of fuzzy sentations of linguistic hedges as modal controllers and fuzzy expert systems. switching functions. operators, and a kind of fuzzy reasoning based on intuitionistic logic.
Hosei University, Department generalization of the binary logic J-K 1-meter model helicopter, as well as of Instrument and Control flip-flop, which is the basis for most the hovering function for a 3-meter Engineering, Tokyo
conventional digital computers. An model. In both models, each function is
important property of Hirota's fuzzy automated by means of a fuzzy logic At Hosei University I met with Kaoru flip-flop is that it embodies the binary controller composed of a relatively small Hirota and saw video demonstrations flip-flop as a special case. Thus with set of inference rules. For example, of robots developed in his laboratory. this device one should be able to pro- hovering requires 36 rules, broken down One of these is a general-purpose robot gram both binary and fuzzy logicon the into one group of 18 main rules (to that is now being manufactured by same machine. There reportedly has stabilize position and speed) and one Mitsubishi. The first video showed an been discussion at LIFE about starting group of 18 subrules (to stabilize attiearly version ofthis robot, presented at a project to advance this idea.
tude), where each of these is further IFSA-87, playing two-dimensional ping
decomposed into simpler groups of 2 pong with a human opponent. The Tokyo Institute of Technology to 6 rules. Ordinary flight control experiment used a video camera to track (TIT), Departments of
requires another 30 rules, 14 for speed the location of the ball and a collection Intelligence Science and and 16 for attitude, and rotation requires of 30 inference rules to determine the Systems Science, Yokohama 28 rules, 14 for speed and bank angle hitting position. The movement of the
and 14 for attitude. In total 15 input robot was very slow and deliberate but At TIT I met with Shigenobu variables need to be monitored. Nine nonetheless surprisingly accurate con- Kobayashi in the Department of Intel- of these are the current value, rate of sidering that it was being controlled by ligence Science and Anca Ralescu in change, and acceleration of pitch, roll, a 16-bit, 5-MHz NEC personal com- the Department of Systems Science and yaw. Others include altitude and puter. A second demonstration of the The latter is also the home department various speeds. Outputs are instrucrobot showed it doing a moderately of Michio Sugeno, whom I had met on tions to the helicopter's usual controls, good job of traditional Japanese flower 5 August at LIFE.
i.e., stick, pedals, etc. arrangement, using a collection of Because Kobayashi's work is almost In reviewing the outline for Sugeno's between 30 and 50 rules. Third was a exclusively concerned with case-based controller, one cannot help but be tape of the robot playing yo-yo. Here reasoning, it will be discussed in the struck by its relative simplicity. Although a video camera watched the movement section below. Professor Ralescu is an there are in total more than 100 rules, of the yo-yo against a white background, American, who was on a visiting appoint- each lowest-level rule group is logically and the robot moved accordingly. ment in Sugeno's laboratory, and start- independent of all the others and can
A final video showed a more recent ingin mid-August she will be visiting at be treated conceptually as a selfexperiment, involving a fuzzy logic LIFE for a year under a grant from the contained unit. Thus overall design controlled robot that throws darts at a U.S. NSF. One of her interests is a and tuning should be fairly easy. Realfalling object. The more dramatic seg- model of fuzzy reasoning based on time tests of the 1-meter model showed ment showed the robot throwing at an possibility and necessity measures. She that the fuzzy logic controller actually object falling through an array of pegs, has also studied the use of neural nets does better than a trained human, i.e., like in a pinball machine, and scoring a for learning fuzzy set membership it reaches stability in the hovering operahit on virtually every try. There also functions.
tion more quickly. It also did very well have been some attempts to program a Of special interest is Sugeno's cur- on the other functions. Given successrobot to grasp a spherical object mov- rent work, mentioned previously, on a ful completion of tests with the 3-meter ing erratically in two dimensions. The voice-controlled helicopter. The objec- model, they will next seek funding to results here have so far been only tive is to build a helicopter that can be work on a full-size helicopter. marginally satisfying, with the robot flown by simple voice commands, such missing the object more often than as “hover,” “forward,” “back,” “left,” Tokyo Electro-Communications successfully grasping it.
“right,”“up,”“down,” and “land.” One University, Department of Another interesting project, men- potential application would be an Communication Systems, tioned earlier, is design of a fuzzy flip- unmanned helicopter used for sea res- Tokyo flop circuit. This work shows promise cue, controlled by voice commands from of leading to the first “fuzzy memory” a mothership
My last visit was with Nakaji Honda, and conceivably can serve as the basis At the time of our meeting, Sugeno where I learned of a project to control for a general-purpose fuzzy computer. had already successfully automated all the lighting of a room by means of a In essence, Hirota's fuzzy flip-flop is a the necessary basic functions for a hierarchically organized rule-based fuzzy controller. The aim is to produce either CASE-BASED REASONING at the Tokyo Institute of Technology. uniform brightness or special effects in
Kobayashi is well known for his prior certain parts of the room, so as to pro- Case-based reasoning (CBR) is work in Al, having written rather vide a physiologically and psychologi- largely the invention of Christopher extensively on the subject of machine cally comfortable atmosphere. The Riesbeck of Northeastern University learning. His work has explored the problem is complicated by requiring and Roger Schank of Yale. The under- possible theoretical connections that the system be able to control inte- lying intuition is that much of human between explanation-based learning and rior lighting in such a way that it sup- reasoning and decision making is derived CBR, and he and his coworkers had plements and/or cooperates with exter- from remembered personal experiences. undertaken applications in three areas: nal natural light, which is constantly The term "cased-based” reflects a focus (1) an expert system for configuring a changing. While this research focuses on areas like medicine and law, where cigarette rolling machine, (2) producon an application of seemingly limited remembered experiences can be repre- tion process diagnosis, and (3) inforutility, it apparently is directed toward sented as individual cases. The basic mation retrieval. The work on (1) feaa much larger and important applica- idea is to develop a database of cases tures an “interactive" CBR system, which tion, namely, to automate the control that can be accessed for the purposes of foregoes the attempt at full automaof large power distribution systems in a current problem situation. The given tion and provides that the user play a response to changing loads.
situation is analyzed for its relevant role in each phase of the CBR process.
properties, and based on these proper. This is reasonable, as one should not Hiroshima Institute of
ties, any similar cases are retrieved from expect to be able to fully automate Technology, Faculty of
the database and modified or other- CBR for problems much more comEngineering
wise utilized for the present purposes. plex than simple cooking recipes.
This produces a new case that is then The other focus of activity in CBR is From the May 1991 issue of the added to the database for future refer- Osamu Katai's laboratory at Kyoto Journal of the Japan Society for Fuzzy ence. The CBR paradigm is in this University. Here the main proponent Theory and Systems, I learned of some manner also a learning paradigm, build- appears to be Tetsuo Sawaragi, workrecent work by Kazuho Tamano on an ing its knowledge base through ongoing ing in collaboration with Katai and optical fuzzy inference system. In this interactions with the user. This feature Sosuke Iwai. Projects in this area are system, graphs of the membership is attractive, as it offers a way to get currently only in the planning stages, functions for a rule's fuzzy linguistic beyond the knowledge acquisition with proposed application domains terms are etched onto transparent plates, bottleneck inherent in most other Al similar to those undertaken by a light is passed through the plates, paradigms.
Kobayashi, particularly diagnosis and focused by a set of Fresnel convex lenses, An early investigation of the meth- information retrieval. and the amount of light coming out is ods for implementing such a system From my meeting with Seiji measured by a set of photo diodes. was carried out by Schank's doctoral Yasunobu at Hitachi, I learned that These measurements yield the rule's student, Kristian Hammond, where the there have been recent discussions at inferred conclusion, in accordance cases were recipes for Chinese cook- his laboratory about studying CBR, but with Mizumoto's product-sum-gravity ing. Hammond's CHEF served as a that it was not clear to them how or method of fuzzy inference. The infer- vehicle for exploring the possibility of where these principles could be applied. ence is thus virtually instantaneous, fully automating all the phases of case. Having just learned that Hitachi is although multiple conclusions still based reasoning, i.e., case retrieval, beginning a program to develop and need to be combined by an electronic modification, storage, and repair. CBR market a package of software developdefuzzifier.
has subsequently attracted a sizable ment tools, I suggested the possibility This appears to be the very first following in the West, and for each of of using CBR in software engineering, report of such an inferencing device. the last 3 years there has been a work- for both software system design and The results are still very much in the shop on the subject sponsored by the software development project planning. experimental stage, but the possibili- Defense Advanced Research Projects In the former, cases are former designs, ties are intriguing. Future experiments Agency (DARPA).
and in the latter, cases are plans for are planned wherein the etched plates During my stay in Japan, I uncov- implementing those designs. His iniare replaced with programmable light- ered only two places where there is any
ered only two places where there is any tial reaction was that this seemed like a emitting diode (LED) displays. significant activity in CBR. The main reasonable idea. It is likely that Japanese one is Shigenobu Kobayashi's laboratory involvement in CBR will continue to