Tel: +81-44-777-1111 Fax: +81-44-754-2594 Haruki Tabuchi, Deputy Manager Development Dept IV Software Division, Computer Systems Group Fujitsu Ltd. Toray Kensetsu Bldg 4843-1, Bunkyo-cho 1-chome Mishima-shi, Shizuoka 411, Japan Tel: +81-559-88-1601 NEC On 14 October 1991 I also visited NEC Laboratories, where my hosts were Katsuhiro Nakamura (senior manager) and Eiji Okamoto (research manager). Okamoto explained his work on encryption, which appears to be well thoughtout and serious (Ref 3). All encryption is by software on the grounds of expense. Separate algorithms are used for data and key encryption. The data encryption algorithm is proprietary, designed for high performance, capable of 100 KB/ s throughput. The key encryption algorithm is a modification of RSA, designed to avoid a directory of public keys, described in the paper. Dr. Eiji Okamoto Tel: +81-44-856-2141 Fax: +81-44-856-2235 E-mail: okamoto@ibl.cl.nec.co.jp Katsuhiro Nakamura Senior Manager On 15 October 1991 I visited Sony Computer Science Laboratory, much smaller than the Fujitsu or NEC laboratories. It has been described in a previous Scientific Information Bulletin article [D.K. Kahaner, "Sony Computer Science Laboratory," 15(4), 61-65 (1990)]. My host was Nobuhisa Fujinami, a researcher. They are working on the Muse operating system, a research system being developed during a 5-year project due to be completed in 1993. They are currently at the stage of implementation of a prototype and preparing for the implementation of version 1.0. An outstanding issue is that of cryptographic communication and authentication. At this stage they are still defining the issues in this area and are a long way from solving them. Security is not at the top of their list at present, because issues such as object identification schemes and dynamic configuration are seen as preceding it. I had the impression that they would reach the end of this project without necessarily having reached a final solution to system security. Nobuhisa Fujinami Tel: +81-3-3448-4380 Fax: +81-3-3448-4273 REFERENCES 1. R. Yatsuboshi, Y. Fujisaki, and R. Akiyama, "Data and voice encryption," Fujitsu Scientific and Technical Journal 22(4) (September 1986). 2. R. Akiyama, N. Torii, and T. Hasebe, "ID-based key management system using discrete exponentiation calculation," paper presented at the 1990 International Symposium on Information Theory and Its Applications (ISITA'90), Hawaii, 27-30 November 1990. 3. E. Okamoto and K. Tanaka, "Key distribution system based on identification information," IEEE Journal on Selected Areas in Communications 7(4), 481-85 (May 1989). Jonathan Moffett graduated from Trinity College, Cambridge, in 1961 with a degree in mathematics. After a background as systems consultant on large commercial systems he has been a computer security specialist since 1978. He was computer security advisor for Esso UK and Esso Europe Inc. and has acted as consultant on computer security for banks and other commercial organizations. Dr. Moffett completed his Ph.D. in computer science in 1990 at the Department of Computing at Imperial College, University of London. He is currently a research associate at Imperial College, working in the area of distributed system management and concentrating on security management. Dr. Moffett is a member of the British Computer Society, a Fellow of the Association of Certified Accountants, and a Chartered Engineer. His research interests include security and control in distributed systems, formal representation of organizational structures, and decision making in conditions of uncertainty. JAPANESE ADVANCES IN FUZZY SYSTEMS AND CASE-BASED REASONING This article presents a survey and assessment of fuzzy systems and case- REPORT SUMMARY During this past summer (1991), I spent 2 months on an appointment as Visiting Researcher at Kansai University, Osaka, Japan, and 5 weeks at the Laboratory for International Fuzzy Engineering Research (LIFE) in Yokohama. Part of the expenses for the time in Osaka, and all expenses for the visit at LIFE, were covered by the Office of Naval Research (ONR). While there I met with most of the key researchers in both fuzzy systems and case-based reasoning. This involved trips to numerous universities and research laboratories at Matsushita/Panasonic, Omron, and Hitachi Corporations. In addition, I spent 3 days at the Fuzzy Logic Systems Institute (FLSI), Iizuka, and I attended the annual meeting of the Japan Society for Fuzzy Theory and Research (SOFT-91) in Nagoya. The following report elaborates what I learned as a result of those activities. There is no doubt that Japan has pulled far ahead of the United States and other Western nations in implementing the theory of fuzzy sets and systems. Of primary significance is development of the fuzzy logic controller, wherein the traditional proportion integral derivative (PID) controller is replaced with a fuzzy rule-based expert system. The main advantage of the fuzzy logic controller is its ability to handle real-time nonlinear control problems. The results achieved so far clearly by Daniel G. Schwartz portend a minor revolution in the field of automated control. Implementations of the underlying ideas have been realized in both software and hardware, using both digital and analog devices, using both digital and analog devices, and there already are several dozen commercial products that utilize these control techniques. Very recently, researchers at the Tokyo Institute of researchers at the Tokyo Institute of Technology have succeeded in fully automating the hovering operation in a model helicopter, a task well known for its inherently difficult stability problems. U.S. patents have been obtained by Japanese corporations for various very large scale integration (VLSI) fuzzy inference chips, and Omron Corporation alone boasts over 700 patents either obtained, pending, or in application. The potential variety of applications for these ideas seems virtually unlimited. Fuzzy technology is at the same time being developed in other directions. Hitachi Corporation markets a fuzzy expert system shell that has now sold over 2,000 copies, and research directed at expanding such systems' reasoning capabilities is ongoing in numerous universities, corporations, and research institutes. In recent months, even the more traditional Japanese artificial intelligence (AI) community has begun to recognize the potential value of fuzzy logic as a model of natural human reasoning. Also in progress is development of a fuzzy flip-flop circuit, which shows promise of providing the basis for a general-purpose fuzzy computer. The fuzzy flip-flop can include the conventional binary flip-flop as a special case. Other recent developments include some experiments with an optical fuzzy inference device, which performs the fuzzy inference operation at a much higher speed than electronic VLSI circuits. Overall, Japan's advances in these areas are impressive and suggest many fruitful lines for future research. Case-based reasoning (CBR) is a relatively new idea that currently has only a small cadre of Japanese proponents. The researchers in this area share the view of their American counterparts that CBR holds promise of alleviating some of the knowledge acquisition problems shared by other knowledge representation schemes. As an approach to simulating natural human reasoning from remembered experiences, CBR is implicitly also a learning paradigm. Japanese scientists are currently exploring applications of CBR in such areas as machinery configuration, industrial process diagnosis, information retrieval, and complex system simulation. There is also the possibility of using CBR techniques in computeraided software engineering, for both the design and development phases of complex software systems. Research in these areas is still in the formative stages, however, and further study will be needed before these ideas manifest in concrete applications. INTRODUCTION Today in Japan we are witnessing what is likely to become a textbook example of how a theory is turned into applications. What is especially compelling about this particular example, moreover, is that the applications themselves confirm the intuitions held 25 years earlier by the theory's originator. In 1965, when Lotfi Zadeh at UC Berkeley's Department of Electrical Engineering published his first paper on fuzzy sets, he was implicitly advancing the thesis that one of the reasons humans are better at control than currently existing machines is that they are able to make effective decisions on the basis of imprecise linguistic information. Hence it should be possible to improve the performance of electromechanical controllers by modeling the way in which humans reason with this type of information. The theory developed slowly at first, but by the early 1970s it had attracted a small international following. This included a number of Westerners, mostly mathematicians, and a small number of Japanese engineers. In those days, the interest was spurred primarily by intellectual curiosity, although even then there was a pervasive belief in the theory's ultimate applicability. During this time, investigations focused mainly on the mathematical properties of fuzzy sets and closely related notions, and numerous variants of fuzzy logic were explored. several foundational challenges posed Although most of the work at this Because of its importance in the current state of fuzzy systems research, all of the next section is dedicated to the subject of fuzzy logic controllers. The remaining sections survey other developments in a report on activities at various societies, research institutes, universities, and corporate laboratories; cover current activity in case-based reasoning; and offer a few summary reflections, including conjectures on why the Japanese have outstripped their Western counterparts in this domain, Western counterparts in this domain, together with some suggestions on what the United States might do to catch up. By the late 1970s, interest in fuzzy systems had grown rather explosively, attracting many researchers from around the world and spawning bibliographies with citations numbering in the FUZZY CONTROL thousands. Still, most of the work was theoretical. The main topics included fuzzy knowledge representation and reasoning schemes, the philosophical ramifications of fuzzy logic and fuzzy set theory, fuzzifications of various branches of classical mathematics, and The first paper describing a fuzzy logic controller was published by E.H. Mamdani and S. Assilan of Queen Mary College, England, in 1975. For their study, they chose the example of a simple steam engine. The controller for this engine has four input variables-pressure error, speed error, change in pressure error, and change in speed error--and two output variables--heat change and throttle change. The essential idea was strikingly simple. In the conventional PID controller, the system being controlled is modeled analytically by a set of differential equations whose solution tells what adjustments should be made to the system's control parameters for each type of system behavior. The proposed fuzzy logic controller, on the other hand, was based on a logical model that directly represents the thinking processes that a human operator might go through while controlling the system manually. Such a logical model is expressed as a set of inference rules of the form "if behavior variable B (input to the controller) is observed to be in the state X, then change control parameter C (output from the controller) by an amount Y" (or perhaps to state Y). The model earns the designation “fuzzy" by virtue of its specifying the amounts X and Y linguistically, using terms like "positive big," "positive medium," "positive small," "no change," "negative small," etc., where each such term is represented as a fuzzy subset of the associated measurement domain. [Given a measurement domain D, a fuzzy subset A of D is represented as a function mu_A which maps D to the unit interval [0,1], whereupon the real number mu_A(x) is the degree of membership of x in D. In case the fuzzy set A is being used as the meaning of a linguistic term tau, then this degree of membership is further interpreted as a degree of "compatibility" of x with tau. For example, suppose D is a range of pressures measured in pounds per square inch. Then some fuzzy subset PB of D might be given as the meaning for the term "positive big," in which case, for each pressure x, mu_PB(x) becomes the degree of compatibility of x with the designation "positive big".] Many approaches have been developed for expressing such inference rules mathematically, for choosing an appropriate rule set, for defining the fuzzy sets that are to serve as meanings for the linguistic terms, for combining results when more than one rule might apply, and so on. In Mamdani and Assilan's experiment, the inference rules were expressed in the manner proposed by Zadeh's works, the rules were selected in accordance with Zadeh's "compositional rule of inference," and the fuzzy results from multiple rules were combined using a kind of generalized averaging technique, after which one could easily extract a precise (nonfuzzy) value as the controller's final output. The averaging technique was applied only after the entire set of inference rules had been "tuned," however, so as to eliminate contradictory results. That is, the rule set was manipulated so that it would not be possible for one rule to conclude, for example, that the throttle should be moved in the positive direction, while another would conclude that should it be moved in the negative direction. This tuning process was accomplished by simply checking all possible output combinations and dealing in an appropriate manner with those that were potentially contradictory. The task was not exceptionally difficult, since the controller employed only a small number of rules--15 for the heater and 9 for the throttle. This experiment, together with a few closely related experiments conducted by others, clearly demonstrated that this was an effective means of automated control. Indeed the logical models have a definite advantage over the traditional analytical models in that (1) they work well even when the relation between the controller's input and output variables is nonlinear, and (2) they are much more robust with respect to changes in the controlled system's parameters, e.g., the desired engine speed. It is generally held that classical PID controllers cannot be designed for the case of nonlinear control and that, even for linear control, they typically must be designed anew whenever one resets the basic system parameters. Following this work, Mamdani tried unsuccessfully to secure funding for his research from the British granting agencies. Unable to obtain any support, he ultimately abandoned this line of investigation to pursue other opportunities. His work did not go unnoticed in Japan, however, and approximately 10 years later, Hitachi Corporation announced the Sendai Railway as the first fully automated subway system employing a fuzzy logic controller. Hitachi had for many years been in the business of designing subway control systems, particularly safety mechanisms, and so this next step was a natural evolution of its existing product lines. The new system controlled all aspects of accelerating to speed and braking for corners or stopping at the next platform, so that the only human operator served essentially as a conductor, watching out for passengers' safety while getting on or off the train. Implementation of the Sendai fuzzy controller was largely due to Seiji Yasunobu of Hitachi's Systems Development Laboratory. Although this work was derived from that of Mamdani, Yasunobu's approach differed in two important respects. First, while he retained the use of fuzzy sets for defining the meanings of the needed linguistic terms, his rules were crafted through a more ad hoc trial-and-error methodology, rather than using the compositional inference technique. This evidently was because rote application of the latter did not produce the desired relations between rule premises and conclusions. An ad hoc approach was feasible since the logical part of the system used a mere nine inference rules-five for accelerating and/or maintaining constant speed and four for braking. Second, he introduced a new method for combining the results of multiple rules and extracting a precise output. This approach, now known as the "centroid method," has become the standard solution for this problem. Through simulations, Yasunobu demonstrated that the fuzzy logic controller was superior to the conventional PID controller along several key parameters, including accuracy in stopping at the platform, rider comfort (jerkiness of acceleration and braking), and fuel economy. He proposed his ideas to Hitachi in 1983, published his simulation results in 1985, and the Sendai Subway opened in 1987. It has been performing satisfactorily ever since. Also in 1987 another event occurred which, together with the Sendai Railway, served as the catalyst for an explosion of interest in the subject of fuzzy control. This was Takeshi Yamakawa's demonstration of his inverted pendulum experiment at the Second Congress of the International Fuzzy Systems Association (IFSA-87), held in Tokyo. The inverted pendulum is a classic control problem, amounting to balancing a vertical pole that is attached to a belt by a hinge, so that the pole can fall to the right or the left. The idea is to monitor the angular position and speed of the pole and move the belt to the right or left accordingly, so as to maintain the pole in an upright position. The problem becomes more difficult as the pole becomes shorter and/or is reduced in total mass, since the required response times decrease in proportion to the square of the amount by which either height or mass is reduced. Yamakawa's controller featured two types of VLSI chips of his own design. One was a fuzzy rule chip, which directly implements Zadeh's compositional rule of inference, and the other was a defuzzifier chip, which calculates the centroid of a collection of fuzzy membership functions. This may be contrasted with Mamdani's approach, wherein he precomputed the results of the compositional inference rule for a limited set of possible inputs and then used these |