descriptions of research activities, and RCAST Journal (once a year), which contains outlines and summaries of research, lists of publications, and current research status. General questions about RCAST can be addressed to

Prof. Setsuo Ohsuga, Director
Research Center for Advanced
Science and Technology,
The University of Tokyo
4-6-1 Komaba

Mejuro-ku, Tokyo 154 JAPAN
Tel: +81 3-3481-4411

Interestingly, Prof. Ohsuga also holds the chair of Knowledge processing within the Advanced systems department, the focus of the current report (this dual appointment speaks worlds about the administrative structure of this research organization).

I have reported on RCAST activities in the past, see for example ["tachi.lab", 9 March 1992]. Professor Tachi is in the Advanced Devices department, and under the chair of Intelligent Sensing Devices.

Ohsuga is a former President of the Japan Society for Artificial Intelligence and has been engaged in research in AI since 1970. His lab is jointly run with Koichi Hori who joined in 1988 from the National Institute of Japanese Literature (Email: HORI@OHSUGA.RCAST. U-TOKYO.AC.JP). Currently, a professor from France and 19 students are visiting; about one third from outside Japan. Although the working language is Japanese, English is widely spoken and most of the technical reports are available in English. Computer user-interfaces are mostly in Japanese however.

On the general topic of expert systems, Ohsuga pointed out that these are still fairly small scale in Japan, but the systems are being rapidly implemented by industry. He

is personally interested in large systems, for example the large knowledge base that is needed by the Electronic Dictionary Research project [see "edr.92", 27 March 1992]; however, he feels that such a knowledge base can't be flat, i.e., it needs to be built as a multilevel structure. Parenthetically, the question of ICOT was brought up. Ohsuga felt that their hardware was okay, but that there were no substantial applications developed yet. He also echoed the Japanese sentiment that ICOT had been a very good advertisement for AI, and in fact had moved students away from traditional areas, to the point now that the best students in Japan seem to be moving into AI subfields. He also offered that ICOT had propelled the growth of the AI in Japanese industry.

OHSUGA LAB

Ohsuga has proposed a knowledge representation language (called multilayered logic) based on an extension of predicate logic by expanding its syntax. (Ohsuga believes that knowledge representation is the central issue in AI. He also feels that multilayered logic satisfies most requirements for knowledge representation; it is very convenient for defining structured objects, and is also convenient to describe and to link to databases and existing procedures.) The Lab has also implemented a knowledge base system (KAUS; Knowledge Acquisition and Utilization System) based on this language.

KAUS runs under Unix. It uses a special procedural atom, 'Exec', that allows cooperation between knowledge processing and conventional computing in a simple way. Exec is a second order predicate that accepts character strings as arguments. It interprets the first argument as a program name and the remaining arguments as arguments of

that program. By using the fork and join mechanism in Unix, it forks a new process and runs the program, eventually returning either 'yes' or 'no' according to the end code of the program. Thus many Unix programs can mesh nicely with KAUS, including various database systems.

Ohsuga and his group have developed a variety of applications using KAUS, primarily in engineering domains, including mechanical design, feedback control system analysis, airplane wing design, chemical compound design. Some details and references are described below. What is important about this work, in my opinion, is the very close link between advanced knowledge processing concepts and traditional engineering issues. Most of the students working in Ohsuga's lab are trained in computer science, not engineering. Nevertheless, the lab as a whole has a strong bias toward practical problems as seen by the applications they are studying. RCAST's philosophy emphasizes openness. When it is coupled to the quality of the work that is occurring there, collaborative activities with Western scientists are bound to be fruitful.

A special interest at this lab is the intelligent CAD systems, i.e., the development of knowledge-based CAD system. In one feedback control system application, the system offers the designer a diagram language to define an object model of a control system with block diagrams and mathematical expressions. Routine tasks such as deriving transfer functions and calculating time responses, are mostly done automatically. The system is distributed by using various computing facilities on a network. Such feedback control system design involves two kinds of man-machine cooperative activities one to build the control system model in the computer, and the other to determine various physical

parameters based on the model (time responses, frequency responses, root locus, relative stabilities, etc.).

We were also told about the prototype aircraft wing design system developed by using KAUS. Ohsuga pointed out the difficulties associated with such designs, and that it may require as many as ten years for complete design and manufacture, and that it may require not only numerical and large scale database processing but also knowledge processing and model processing. The prototype system has two numerical packages, one for aerodynamic and the other for structural analysis, but other packages can obviously be added. A database of NACA wing sections is stored as standard wing sections and can be referred to by designers, or they can add their own wing sections to the database. A wing model consists of an aerodynamic and structural model. In aerodynamic design, designers first make an initial model that specifies the shape of the wing, then they analyze and modify this, repeatedly, until goal characteristics such as lift, drag, and moment are obtained. In structural design, designers first make an initial model that specifies the wing framework constructed by beams and ribs, then they analyze and modify it until the wing has enough strength. In case it isn't possible to obtain a strong enough structural model, the aerodynamic model must be modified. Since both models are integrated into this system, consistency can be maintained. Associated with this project is a package, MODIFY (Modeling tool for Integrated Finite element analysis), developed in the lab. This is an object oriented FEM system, which uses parts objects, and the concept of degree of importance to help with mesh generation.

One of the most interesting papers on this subject appeared in Computer Aided Design, Vol 21, No. 5, June 1989, pp. 315-337, Toward Intelligent CAD Systems, S. Ohsuga. I recommend it to everyone that is studying this topic as a readable and thought provoking summary.

Another related system is Chemilog. This is a logic programming ilog. This is a logic programming language and system to support the development of chemical knowledge information processing systems. It is an extension of Prolog and is viewed as a constraint logic programming language. In Chemilog, graphs representing chemical structures are treated as basic components and isomorphic graphs and are identified. Constraints such as substructure recognition and replacement of substructures make it possible for users to write programs that handle chemical structures in a visual form. There are two parts, an inference engine and a chemical structure database. Ohsuga emphasizes that while this project is in the domain of chemistry the method can also be applied in other engineering domains where graph representation is used. CHAUS is an extension of Chemilog using KAUS. This was a national project that enlisted the aid of chemists throughout Japan who prepared knowledge sheets describing what functionality appears when certain kinds of structural changes in chemical compound are made.

Ohsuga has recently been thinking about applying KAUS to the problem of software design. He claims that an important step in software design is conceptual modeling (rather than formal specifications) and that knowledge processing technology can be fully used in the process of conceptual modeling. Moreover he claims that conceptual

models can directly be converted to procedural programs. He is building a system using the meta-level mechanism of KAUS to explore these ideas.

Prof. Hori is involved in the articulation problem. This is the process of cutting a set of symbols out of the nebulous mental world. "I have a concept but I can't verbalize it." It is necessary to articulate this mental world in order to acquire knowledge, model information, etc. Prof. Hori and coworkers have built a system, AA1 for aiding this articulation. Experts often put fragments of ideas consisting of several words and hand-drawn figures on paper, and frequent erasure and rewriting are added when thinking about new ideas. Hori claims that AA1 assists in this stage, and they have experimented with it in the conceptual design of automobiles. Currently they are trying to investigate, from the cognitive science point of view, why the system works so well.

Mainly, Hori's work concerns the bottom-up symbolization process in the human mind. Ohsuga's work is concerned with solving a problem symbolically modeled. Currently, they are trying to connect these two activities. Ohsuga emphasized that he also has interests in human interface, natural language understanding, analogical reasoning, abductive reasoning, nonmonotonic reasoning, learning systems, hypothesis generation, etc., with special emphasis on the world of engineering.

Recent Papers

S. Ohsuga, "How Can Knowledge Based System Solve Large Scale Problems-Model Based Decomposition and Problem Solving," to appear

in Knowledge Based Systems, 5(3), 1992.

K. Hori and S. Ohsuga, "Word Space Processor for Assisting the Articulation of the Mental World," in Information Modelling and Knowledge Bases, Ohsuga, Kangassalo, Jaakkola, Hori, Yonezaki, eds. (IOS Press, Amsterdam 1992).

Other Papers

S. Ohsuga, "Framework of Knowledge-Based Systems-Multiple metalevel architecture for representing problems and problem-solving processes," in Knowledge Based Systems, 3(4), 204-214 (1990).

T. Akutsu, E. Suzuki, S. Ohsuga, "Logic-based Approach to Expert Systems in Chemistry," in Knowledge Based Systems, 4(2), 103-116 (1991).

During the JTEC visit to RCAST, Prof. Ohsuga also provided us with a package of preprints from graduate students working in his lab. The titles and authors of these papers are listed below, but for specific information it would be best to contact Ohsuga or Hori directly.

Hiroyuki Yamauchi, "KAUS: Knowledge Acquisition and Utilization System"

Chunye Li, "Problem Model Design/Transformation-Based Program Development"

Ning Zhong, "Knowledge Discover and Management in Integrated Use of Knowledge-Bases and Databases"

Kousuke Yoshizawa, "Soft-Defined Machine" Satoshi Kobayashi, "A Method for Acquiring Problem Decomposition Strategy from Traces"

Einoshin Suzuki, "Framework for Connecting Several Knowledge-Based Systems Under a Distributed Environment"

Akira Utsumi, "Primitive Based Representation of Adjectives and Figurative Language Understanding"

Nigel Ward, "Structured Connectionist Studies in Natural Language Processing"

Ionnis Zannos, "Interactive Music Generation System"

Jari Vaario, "Study on the Development of Neural Networks in the Context of Artificial Life"

Shinichiro Yano, "An Approach to Aid Large-Scale Design Problem by Computer"

Y. Minagawa, "The Study to Aid Finite Element Method's Preprocessing with Knowledge Engineering"

Akira Tajima, "Supporting the Development of Intelligent CAD System on KAUS"

Masanori Sugimoto, "A Method to Assist the Acquisition and Expression of Subjective Concepts and Its Application to Design Problems"

Yasunuki Sumi, "A Study of Computer Aided Thinking-Mapping Text-Objects in Metric Spaces"

David K. Kahaner joined the staff of the Office of Naval Research Asian Office as a specialist in scientific computing in November 1989. He obtained his Ph.D. in applied mathematics from Stevens Institute of Technology in 1968. From 1978 until 1989 Dr. Kahaner was a group leader in the Center for Computing and Applied Mathematics at the National Institute of Standards and Technology, formerly the National Bureau of Standards. He was reponsible for scientific software development on both large and small computers. From 1968 until 1979 he was in the Computing Division at Los Alamos National Laboratory. Kahaner is the author of two books and more than 50 research papers. He also edits a column on scientific applications of computers for the Society of Industrial and Applied Mathematics. His major research interests are in the development of algorithms and associated software. His programs for solution of differential equations, evaluation of integrals, random numbers, and others are used worldwide in many scientific computing laboratories. Dr Kahaner's electronic mail address is: kahaner@xroads.cc.u

tokyo.ac.jp.

Dr.

David K. Kahaner

In March 1992, a JTEC (Japanese Technical Evaluation Center) team, headed by Prof. Edward Feigenbaum, Stanford University [Email: FEIGENBAUM@SUMEX

visited

AIM.STANFORD.EDU], Japan to assess and report on Japanese activities in the areas of expert systems (ES). An oral summary of their report was presented in June 1992 at the NSF in Washington, DC, and the written report will be available soon. The JTEC team visited a large number of laboratories at Universities, government, and industrial R&D centers. I was able to accompany this group on many of their visits. The following report is an overview of one particular research organization, Hitachi, and their specific programs. I have organized this to first catalog some specific examples of expert systems that have been developed by and for Hitachi customers. Thus these illustrate some expert systems that are in use here in Japan. Next we describe major research projects underway that related to the topics of AI and ES and connections to other research around the world.

The main software tool provided by Hitachi is a version of ES/KERNEL. (Separately named versions exist for workstation, mainframes, PCs, and for On-line processing. There is also a new workstation version ES/KERNEL 2 that runs under

X-Windows and Motif, which is claimed to be very user friendly.) More than 4,000 copies of ES/KERNEL have been sold in Japan. Source was written in C. Users can make use of various reasoning strategies, including the following:

• Production system: Reasoning proceeds through repetitive sequences of search for selection and execution of applicable rules.

• Object oriented: Reasoning proceeds by repetitive sequences of passing messages to a method within a frame, and method execution. • Assumption-based: Reasoning starts with a number of

possible solutions and solutions that do not meet the specified conditions are discarded. This can be used in the production system as a frame.

• Multilayered: This is a

communication mechanism that executes reasoning through cooperation between different expert systems.

NEL 2 these are implemented in a hierarchical structure via an object-oriented representation. Private memos represent unstructured knowledge, suitable for the description of transient facts that may be obtained in the course of reasoning.

ES/KERNEL 2 also comes with a graphical knowledge editor as well as a knowledge tailor, which allows knowledge to be moved from one representation to another. In addition, there are two domain shells, ES/Promote/W-Diag, for diagnosis applications and based on ES/KERNEL, and ES/Promote/W-Plan, for planning and scheduling applications that are based on Whispa. (I was told that the next generation of a domain scheduling shell will be based on Tosca, and will incorporate extensive user feedback.) Other packages include a knowledge acquisition tool ES/Tool/W-RI rule induction from examples, a shell for real-time process control, another for plant operation, and a user interface building tool UIBT that has some multimedia capability. There is also ES/Guide for Expert System building support based on the spiral and waterfall model. Unit sales of these other tools are a small percentage of the ES/KERNEL sales. Some fuzzy reasoning has been implemented in this package.

Hitachi has implemented Expert Systems (mostly via ES/KERNEL)

in a large variety of application environments, including the following, which are briefly described. Hitachi points out that "even though the range of application of expert systems is very wide, only advanced firms are engaging in their development, and then in limited areas." Thus, like the United States, smaller companies have not yet embraced this technology. The first list below contains material given to us by Hitachi in very abbreviated form. The second lists the titles and organization of expert systems in use that were provided to us in much more detail (omitted here). Thus, we see that major application areas are in high volume large transaction processing, where the actual reasoning is shallow. Finally, Hitachi pointed out that currently the most important application areas for Expert Systems are for scheduling. Less important, but still significant, are for diagnostics. This is also the case in the United States.

Manufacturing

Blast furnace power generation support-User enters information on the status of turbines or other equipment, and system displays diagnostic messages and results of inferencing.

Melting (steel production) planning support-Optimum stock of raw materials and most economical melting plan is created according to required monthly ingot production volume.

Drainage system constructionUser enters size of field and quality of soil, and the system produces design drawings and estimate of expense and time required to complete job.

Resin production planning support User enters sales data and furnace operating status, and system produces integrated production plan from sales to detailed production process. Knowledge base formed

from production conditions and raw material composition.

Boiler breakdown diagnosisUser (repairman) enters details of breakdown, and system produces breakdown cause interactively.

Construction

Shield tunnel planning supportUser enters construction period, diameter of bore, route, soil quality, conditions, and the system detervolume of groundwater, and other mines ideal shape of cutter, number of cutters, and positioning of shield boring machine. This system required about 60 person months to complete and a cost of about US$500K. It contains about 9K lines of C, 100 frames, 850 rules, relates to about 250 parts on a shield machine, and accesses a relational database of about 15K items.

Foundation improvement method selection support-User enters the status of subsidence and the structure to be built, and the system produces optimum foundation improvement method.

Government

Oil transformer breakdown diagnosis-User enters type and volume of gas. (Gas contained in a sample of insulation from transformer located in office or building is extracted and analyzed.) Inference is performed to pinpoint cause of trouble and next inspection date is determined.

Landslide diagnosis-Power generation facility sites are checked to see if they are prone to landslides.

Soil nematode diagnosis-Diagnoses whether or not crop damage is caused by nematodes. If so, determines nematode type and displays method to eliminate the problem.

Building inspection-Supports building inspection in areas of design, structure and facilities, to check whether or not they conform to permits and applicable codes.

Elderly welfare consultationUser enters personal data that then judges status and procedures. System also outputs nursing home entrance eligibility.