Site: NEC R&D Center
1-1 Miyazaki 4-chome
Kanagawa 216, Japan
Date Visited: March 24, 1992
Dr. Masahiro Yamamoto
Dr. Masanobu Watanabe
NEC's net sales, as reported in its 1991 brochure, were $26.2 billion. The company has 118,000 employees. The new corporate philosophy emphasizes computers & communications (C&C). C&C operations accounted for 47 percent of net sales in 1991. In July 1991 the company reorganized into 10 operating groups. We met with members of the C&C Systems Research Laboratories, which is a component of the Research and Development Group. The R&D group accounts for one percent of annual sales, and employs 1,600 people. (Corporate-wide R&D accounts for 10 percent of sales and employs about 10,000 people.)
The C&C Systems Research Laboratories encompasses many individual laboratories devoted to networks, systems, terminals, and basic technologies. Each of these individual laboratories develops ES applications and/or tools.
We discussed the answers to the questionnaire, after which we toured the laboratory and were given demonstrations of five systems: genetic information processing, parallel object-oriented language A'Um-90, Adaptive Model-Based Diagnostic System, Software Synthesis Expert System, and Corporate-wide Case-Based System.
NEC has developed about 1,000 ES applications, of which 10 percent are in routine operation. The major task areas are diagnosis, scheduling, design, and software development aids. The company's biggest success has been the crew scheduling system, COSMOS, developed with Japan Air Lines (Onodera and Mori 1991). This is a large ES (45,000 frames, 7,000 rules, 300,000 lines of code), built on the EXCORE tool. It runs on ten engineering workstations (EWS4800, which has up to 320 megabytes of memory) and four LISP machines (LIME), and is integrated with the host data processing system. The system was turned over to JAL in 1990, where it is now routinely used by JAL schedulers. Before the system was installed, one month of scheduling required two or more days of effort (with overtime) by 24 people. A monthly schedule can now be generated in 15 days by 19 people. Other applications that were reported include a software debugging advisor; FUSION, an LSI logic design tool; and a system called SOFTEX for synthesizing C/C++ programs from specifications (represented as state transition tables and/or flowcharts). SOFTEX is a 300-rule system built with EXCORE, developed by professional programmers and domain experts. The system is still about six months to one year from routine use. In order to make the system appeal to programmers, it was necessary to incorporate in SOFTEX the functionality to enable customization, so that the generated program fits each programmer's style. Their primary general purpose ES building tool is EXCORE, although many systems also use C, LISP, and sometimes even COBOL. EXCORE itself exists in both LISP (EXCORE/CL) and C versions (EXCORE/KWB). NEC has developed a methodology (EXMETHOD) for building ESs, based on the use of EXCORE, that provides guidelines for knowledge representation and acquisition. Training programs are in place that run in duration from five days to one month.
NEC admits to some unsuccessful projects, and attributes the lack of success to a number of reasons, including the knowledge acquisition bottleneck, the difficulty of integration with existing systems, and knowledge base maintenance. These problems tend to get detected late rather than early.
With respect to building large KBSs, the limiting factors are: speed of inheritance; KB management and the amount of memory implied by that; and the need for faster hardware (500 MIPS range). Future ESs at NEC are expected to employ technologies such as model-based diagnosis, case-based reasoning for scheduling and for software synthesis, and combining ES methods with algorithmic methods (model-based reasoning is an example).
As examples of research in large knowledge bases, NEC demonstrated systems containing extensive data on DNA and protein sequences, and on protein secondary structure. The systems are used for searching for common patterns in the sequence information. NEC has a well-documented body of research in model-based diagnosis, genetic information processing, learning (using a technique called Minimum Description Length), learning theory, inductive learning, knowledge acquisition for classification problems and for consultation systems, natural language understanding, and case-based reasoning. The company believes that the important aspects of second generation ES tools are the incorporation of machine learning, model-based reasoning, cooperative problem solving, and facilities for extracting knowledge from large databases.
NEC also has developed a language, called PRIME (PRimary Inference Mechanism with Environment), for describing a semantic model of a domain. The model can be used for deep level inference. The inference process can then be compiled into a shallow knowledge base.
NEC has built two task-specific shells on top of EXCORE: DIAGBOX, for diagnosis; and PLANBOX for planning. These shells were released in January 1992. EXCORE itself has sold about 1000 copies. Its price is about 500,000 yen. The system runs on the NEC workstation, on a mainframe, and, just recently, on a personal computer.
NEC uses system engineers, not knowledge engineers, to develop applications. NEC feels that it is important that KBS technology is "harmonized" with conventional software technology.
The sale of expert systems is not rising as rapidly as had been expected, and in fact is flat now during the current recession in Japan.
NEC's systems for kanji character recognition and for speaker independent machine translation were also mentioned. The latter was demonstrated at Telecom 91, and translates among four languages: Japanese, English, French and Spanish. NEC's machine translation system has been discussed in a previous JTEC report (Rich 1992).
The model-based diagnosis system mentioned above has been used for diagnosing a packet-switched network. The knowledge base contains structural and behavioral models of the device to be diagnosed. Given input/output pairs, the system identifies the faulty component. A minimum entropy technique is used to select the most effective test to perform as the diagnosis proceeds. A single fault is assumed. The Minimum Description Length (MDL) technique is employed here, and in the genetic processing applications mentioned above. MDL is a means to finding an optimum trade-off between overfitting (which implies very long descriptions to cover all examples) and underfitting (which causes a high error rate). The use of MDL seems to be popular at NEC.
Regarding sales, scheduling is increasing; diagnostics are dipping slightly now, but our hosts believe that area will soon resume modest growth. Five years ago their customers needed extensive help in developing ESs but now they don't. Internally, they feel they have many design tasks in NEC that will benefit from the use of ES; they appear to be using it vigorously internally.
Regarding ICOT's influence, today it is a small percentage of the AI'ers in Japan, but this was not so in the beginning. NEC did not build a PIM but is building a parallel processor made up of microprocessors; their architecture is similar to CM-5.
The heart of our visit was the demonstration of five applications. Our hosts offered little or no explanation of the problem solving approach apart from the demonstrations, which leaves a number of questions unanswered. Interestingly, the demonstrations tended to employ the older expert system techniques -- no fuzzy logic or neural nets. Yet the applications were quite advanced with respect to the general state of the art.
Regarding SOFTEX, given that engineers in the telecommunications industry think in terms of state transitions, the input specification can be considered to be high level. However, the transformations to procedural code are fairly straightforward because the structure of the specification and the structure of the resulting program appear to be isomorphic, i.e., each specification corresponds to a single program. The larger system is presumably structured in a way that enables this program-by-program approach. While this was a nicely packaged and clearly presented demonstration, we did not see anything that is pressing the state-of-the-art. After our visit, our hosts explained that there was insufficient time to describe the entire system and they had not yet disclosed the technology in the SOFTEX shell. NEC has in fact synthesized other programs beyond the state transition model software that was demonstrated to us and has published a paper describing the system (Yamanouchi 1992). One of NEC's demos was a corporate-wide case-based system. This was an application in support of software quality improvement. For many years, NEC has collected examples of quality problems and improvements submitted from the organization. The traditional book form of distribution has become useless as it has grown large. What we saw was a simple retrieval example with emphasis upon the design of an index. There was no "repair" function. This application was more of a cultural statement than an interesting application of technology.
The payoff of ES applications wasn't (or couldn't be) quantified. NEC's people found the usual oft-stated limitations of the technology. They found it a very useful tool for internal use by the company. Their applications improved quality by allowing more trials by improving the speed of iteration. They also had some applications with a big ROI. They had no applications they were able to sell in multiples to customers.
The Basic Technologies Laboratory is in charge of VLSI CAD tools, knowledge-based systems, databases, and software architectures. The group does basic research as well as applications to support field personnel. (The term "basic" appears to mean something different in the Japanese labs from we generally mean. Perhaps the difference is not between the Japanese and the U.S., but between company labs and universities.)
Note that although diagnostic programs are fielded, most Japanese do not consider them as being in routine use since "the need for diagnostics do not arise very often." (Is the zero defect policy actually working?)
We saw a demonstration of a distributed, model-based diagnostic program running on a 16-processor ICOT machine. Overall, NEC's research program is similar to what one would find at an average AI lab in the United States.