Site: SEKISUI CHEMICAL
Umeda Center Building
Osaka 530, Japan
Date Visited: March 26, 1992
Sekisui Chemical, which developed the first plastic mold in 1947, has expanded into the fields of housing, building materials, industrial and agricultural equipment, and packing and home building materials. Sekisui's annual revenue is $1.35 billion, and it has 6,200 employees.
Sekisui Chemical has a housing division that sells modular housing, i.e., semicustom housing that is made 80 percent in the factory. This housing division is the fourth or fifth biggest seller of houses in the country (they sell 20,000 homes per year). This constitutes about 50 percent of the company's business. The rest of the company makes plastics (PVC, etc.). This division, called Sekisui Heim, competes with another company, Sekisui House, in which Sekisui has a 21 percent ownership position. Sekisui Heim has annual sales of $488 million.
The leader of the Sekisui team meeting us was Mr. Kawai. He is Director of the company's Information Processing Department. When Mr. Kawai was director of Sekisui's housing division, he was experiencing a problem of a five percent error rate in parts selection, resulting in project delays. He heard Ed Feigenbaum's talk and decided to try ES. The company started by buying 70 percent of a small company now named ISAC (International Sekisui AI Corporation). ISAC does contract work for Sekisui and also sells software, e.g., K-PROLOG, the best selling PROLOG in Japan, and METHODLOG, an object-oriented frame system on top of PROLOG. They are now getting into relational databases with a multi-media interface, and other software products. Two members of the JTEC team visited their offices in Tokyo on the Saturday after the JTEC visit week and saw demonstrations of most of their expert system applications.
Sekisui's main system involves semi-custom configuration (actually they think of it as a parts-picking application). They have to make a specific material list for a house, and those materials have to be delivered to workstations on the factory floor. There are 300,000 different parts in the inventory, of which only several tens of thousands participate in any particular house building. The house itself, as designed by the customer working with the sales office, has to be transformed into a set of parts lists (so many 2 by 4s, so many wall modules, so many bolts, so many plumbing pipes of length L, etc.) The entire process is a rule-driven and constrained process, where the rule book is provided by the Sekisui Design Center, which designs the major classes of houses that are possible (e.g., the Parfait line, the Gloire line, and the Avante line). It is a huge rule book. When a new class of house is introduced, there are many errors made (much inefficiency) until the staff learns the new system very well. Now they simply convert the rule book to a knowledge representation and the expert system immediately performs flawlessly.
Sekisui has a version called HAPPS for its steel frame houses, another for wood frame houses (TAPPS), and a third for high rise houses (MAPPS) -- see Chapter 2 for additional details. HAPPS, TAPPS, and MAPPS are used to figure out which parts from an inventory of about 300,000 parts are needed for each particular house and where they go in the assembly process. About 5,000 different parts are needed for any particular home.
Sekisui has spent 450 million yen building the systems, and the payback is 1 billion yen ($8 million at approximately 128 yen/$) per year!
The systems are written in K-PROLOG and METHODLOG, but run in C on UNIX workstations.
In addition there is a system called APEX that interfaces with the customer to aid in the design of the house. APEX has a CAD system and list of all the different sales options (degrees of freedom of customer choice). The customer designs (with the help of the sales person) by playing what-if games at the workstation. The APEX can even give the customer roughly estimated costs of any specific design, as well as a 2.5 D rendering of what the house will look like when finished (in addition to the floor plans of course). APEX is designed to help salesmen work with the customer to design a modular home. It has been in use for two years, for 10-20 percent of sales. There has been some resistance by sales force -- the problem is apparently with the user interface. We saw this system, much enhanced in drawing and rendering capability, being run on a portable workstation (essentially a very large laptop) and there is a version of this in color.
After a house has been ordered, Sekisui has 2 months to completion including 40 construction days (80 percent completion in the factory). The house must be finished at the site.
Opinion of Mr. Kino, founder of ISAC, AI fever is over in Japan, but ES is now accepted as a legitimate technology/methodology just like FORTRAN or COBOL. There is increasing awareness of what ES can/cannot do; steady increases in orders are expected.
These systems are motivated by the cost of training people and by past mistakes in part selection. Manual methods took 3 years to train experts and result in about a five percent error rate. Now a few days training is enough. Knowledge acquisition for this system came mainly from Sekisui's housing design books. This is a "big win" system -- it saves about 1 billion yen/year. That number seems a little low given that Sekisui does about 350 billion yen/year in housing sales. This is perhaps in the top five of ES applications in terms of impact in the world. Another motivation was lots of problems every time they sold a new type of house. Development began in 1986. Operational use began in 1989.
ISAC has sold about 800 copies of K-PROLOG, about 20 copies of METHODLOG -- an object oriented PROLOG built originally for internal use.
SEEDS/STREAM is an application that provides design support for chemical adhesives. This sounds very much like the 3M application. The system outputs ratios of chemicals in final process for synthesizing adhesives.
Sekisui also has a scheduling system for scheduling use of plastic extruders in its factory. There is the need to re-schedule 2-3 times per week. This system supports human experts and is in use in 8 factories. First operational three to four years ago. It took over a year to develop the system. The result was a reduction in the size of the scheduling team from four to two people per factory and a reduction in scheduling time from four hours to ten minutes.
We asked our hosts why they have developed no process control applications. The answer: that is simply not a main part of Sekisui's business.
Our hosts commented on their lessons learned:
Comments on Sekisui's future plans: the company is interested in case-based reasoning and constraint-based scheduling for more complex problems.
Sekisui has added 37 new employees in information systems this year and has been teaching them UNIX and C. This is a way of getting new ideas (including ES) into the mainline (and mainframe) information systems world.
See Chapter 2 for further details.