rick@cs.arizona.edu (Rick Schlichting) (06/11/91)
Richard D. Schlichting (rick@cs.arizona.edu)
Associate Professor
Department of Computer Science
The University of Arizona
Tucson, AZ 86721, USA
[This report represents the personal opinion of the author, who was on
sabbatical in Japan from Dec. 1989 through July 1990. The sabbatical
was supported in part by grant INT-8910818 from the NSF U.S.-Japan
Cooperative Science Program.]
---------------------------------------------------------------------------
On Wednesday, I visited Hitachi at the invitation of Dr. Hiro Ihara,
a person I've known for a number of years by virtue of our common interest
in fault-tolerant computing. Although Ihara currently works for Hitachi's
relatively new Space Systems Divisions, he arranged for me to visit
Hitachi's Systems Development Laboratory since the work there is
generally more in tune with my own interests. He has also indicated
that he'd like to show me some of the other labs, an offer that I
intend to take him up on during some return visit.
I arrived at the Tama Plaza station on the De'en Toshi line on
the west side of Tokyo about 2:00pm, where I was met and taken to
the labs by taxi. Upon arrival, I was led to a conference room, where
I sat for few minutes before being greeted by Dr. Kinji Mori, a senior
researcher who is currently director of the ADS (Autonomous
Decentralized System) project. We talked briefly about the lab and
his work, and then saw a short slide show giving an overview of the
work going on at the lab. There were a number of interesting projects
underway, with the two most relevant to my own work being ADS, a distributed
and decentralized system for real-time process control, and the fuzzy
logic based subway control system.
After making my research presentation, we returned to the conference
room, where I was shown a videotape on ADS. It was interesting and
informative, despite being mostly non-technical. The tape progressed by
making the analogy between the cooperating processors in ADS and the cells
in the human body. I originally thought this was mostly hype, but
interestingly enough, it turned out to be historically accurate -- the
work on ADS was originally motivated by conversations with an MD turned
computer advocate at the University of Tokyo who made this analogy!
The videotape and subsequent technical discussions highlighted
a number of interesting technical aspects of ADS, especially in the
use of decentralized algorithms and what they call a "Data Field"
for communciation between processes. Although I have not yet sorted
out all of the technical details, the DF concept seems very reminesent
of the tuple space found in the Linda system developed by David Gelertner,
et al at Yale. Although the DF was probably developed prior to Linda
(the ADS project started in 1977, with the first patent in 1979), it is
not nearly as well publicized among academics (as Mori remarked somewhat
ruefully, he actually spends most of his time writing patent applications
rather than papers.) He did give me a paper on the software structure of
ADS that appeared in 1986 Fall Joint Computer Conference in Dallas, which
has pointers to a number of other earlier papers as well. (He also said
that there is a more technical videotape on ADS, although it is
unfortunately only in Japanese.) From a commercial standpoint the system
has apparently been very successful in the real-time process control arena
for which it was originally designed. They have also been successful in
obtaining a number of patents worldwide on this decentralized control
scheme (including the US), receiving some sort of yearly award for best
patent in Japan in the process. My impression is that this is a prime
example of a very worthwhile and interesting project that is virtually
unknown in the US.
Ihara and Mori then accompanied me into another room where a stock
trading optimization program based on neural nets was demonstrated. The
person doing the demonstration was Dr. Ikuo Matsuba, a senior researcher
who I believe is head of this project. The basic thrust of the
program was to select 5 stocks out of 100 so as to maximize the return
but minimize the risk. The basic strategy seems to be to represent
historical correlations (both positive and negative) between pairs of
stocks as neural links, and then iterate to determine which neural
connectsions strengthen as time progresses and which connections
weaken. The strenghened links indicate "good" combinations, while
the weakened connections indicate "bad" combinations. The result could
then be used to select the particular combination of stocks that will
maximize the return for a given level of risk. To demonstrate that this
was a good choice, they plotted this curve on a graph of risk vs. return,
and then selected 5000 (?) combinations at random and plotted them;
all 5000 points were below the curve suggesting that optimal
combinations had indeed been chosen.
The demo was run on a PC level machine, although the calcuations
had been actually run beforehand on a supercomputer. The research group
has also successfully fabricated a chip specicially designed for doing
this type of neural net calculations. Given all this, I was
surprised to learn that the number of people working on this in
the whole corporation was in the range of 20-30. The particular
program I was shown seems to be designed solely to demonstrate
the neural net approach, and I got the feeling that there were no
plans to commercialize it (despite a lot of jokes about getting
rich using the program as advisor.)
At this point Ihara, Mori and I adjourned to dinner at a nice
combination French/Japanese restaurant. The discussion during dinner
ranged from the usual discussions of Japan and the US, to discussions
about Hitachi's efforts in space-based computers. According to
Ihara, they started 3 years behind other Japanese companies in this area,
but are aiming to be competitive in the early 21st century. This work is
being done in partnership with TRW in the US.
--------------------
Dr. Hirokazu Ihara
Senior Chief Engineer
Space Systems Division
Hitachi, Ltd.
216 Totsuka-machi, Totsuka-ku
Yokohama 244, Japan
Dr. Kinji Mori
Senior Researcher, 1st Research Dept.
Systems Development Laboratory
Hitachi, Ltd.
1099 Ohzenji, Asao-ku
Kawasaki-shi 215, Japan
Dr. Ikuo Matsuba
Senior Researcher, 1st Research Dept.
Systems Development Laboratory
Hitachi, Ltd.
1099 Ohzenji, Asao-ku
Kawasaki-shi 215, Japan