Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: Robert------Deep Blue knowledge question

Author: Robert Hyatt

Date: 08:52:16 04/10/02

Go up one level in this thread


On April 10, 2002 at 09:59:07, K. Burcham wrote:

>
>Robert there have been so many comments about Deep Blue chess knowledge.
>the number three statement here is yours. are you refering to chess knowledge
>with this statement.
>
>1. Deep Blue had 50x the knowledge of todays programs.

I don't know that that statement is true.  I am certain, based on details
that I learned by talking to Hsu at the last couple of ACM events, that they
had knowledge that other programs did not have.  We had a giant discussion
one night about opposite bishops.  A version of chess genius was playing
(I think at least) and it was saying the position was a dead draw.  Hsu
pointed out that it wasn't a draw at all and that "deep blue prototype"
knew this.  He later showed us what it thought about how to win.  A GM
had explained this to them a couple of years earlier and they had built it
in to the program with good results.  They mentioned _lots_ of such special-
case evaluation terms that were suggested by the various GM players they had
helping...



>2. Deep Blue could beat todays programs on just the knowledge it had.
>3. Fritz is not a "genius" in terms of smarts...  DB knew much more about lots
>of things than it does...

That is something I wrote, yes...



>4. Todays programs have a very small fraction of the knowledge Deep Blue had.
>
>Robert:
>A. do you say this because you feel no other programmer today has employed a
>Grandmaster to assist with this chess knowledge the way IBM did?


No.  The reason was that DB was _hardware_.  In a software chess engine,
every bit of evaluation "knowledge" you add costs you in terms of speed.  So
it is a series of compromises...  gaining this bit of knowledge is more than
offset by the tactical loss of skill due to the engine running slower...  for
example.

DB was different, because the designed the evaluation in hardware, and they
could pipeline the whole thing so that many parts of the evaluation could
proceed in parallel.  Which means that they could add "knowledge" with no cost
in speed at all.  That gives them a great advantage in that any knowledge can
be added without regard to reducing the tactical skill of the machine, something
that the rest of us have to deal with daily...

Certainly their GM advisors helped.  But without the hardware (special-purpose)
much of the "knowledge" the GMs suggested would have been impractical to
implement because of the cost in search speed.



>B. are you saying this because you feel no other programmer today including
>yourself has the capability to program this chess knowledge into their program?

absolutely not.  I believe that I, or anybody else, could program any chess
knowledge we want into our engines.  It isn't a question of "could" it is a
question of "should" due to the search speed cost for additional knowledge.

IE try to drag-race with a $2,000 a year budget and see what happens when you
race against a factory team with an unlimited budget.  You make different
decisions when you have constraints.  DB had no "time constraints" since it was
a hardware design from the ground up...




>C. are you saying this because you know it would take a huge amount of money to
>resource a Grandmaster over a long period of time to attain this chess
>knowledge, to have available for ones program. and because you know the
>programmers of today do not have this amount of money to budget towards this
>goal?


Maybe or maybe not.  GMs are not super-interested in telling us how to make
the engines so strong that they have no chance.  But some are.  I have had
several giving me advice, for example, at no cost.  Just because they are
interested.  If you paid one, you would get more information from them, so
that is an issue.  But it is not necessarily a limiting issue...




>D. it would seem that if this is a major issue with todays programs, then if one
>of the programmers were to employ a Grandmaster to help with chess knowledge,
>then one of todays programs could far surpass the competition in strength. do
>you agree?  (i am aware of the hardware difference)


Again, the problem is that you can't include every bit of knowledge you want,
if you are dealing with software.  It takes a series of careful compromises to
balance search speed with chess knowledge, so that you get the best overall
engine.  In a hardware project, as the Dodge commercials like to say, "The
rules have changed" and this is no longer an issue.





>
>1997  IBM comments
>
>Deep Blue's "chess knowledge" has been significantly enhanced over the past 12
>months
> through the efforts of team consultant and international grandmaster Joel
>Benjamin. Garry Kasparov
> is certainly a great chess player -- perhaps the greatest in history -- but the
>new and improved Deep
> Blue offers a challenge that even the world champion has yet to experience.



That appears to be true.  Hsu re-designed the DB chess processors after the 1996
match that they lost.  And they spent the rest of the year trying to tune that
new hardware knowledge to produce better play against GM players.


>
>Deep Blue's chess knowledge has also increased over the past 12 months. "We
>spent a lot of time, several
> months, working with a grandmaster, Joel Benjamin," states Campbell. "There are
>sometimes things that a
> grandmaster knows that are sometimes difficult to put into a computer program.
>We are working hard to get
> to know as much about chess as possible."

This is probably true for everyone, although the DB team was certainly a special
case since they had several full-time PhD people working on it, plus the GM
"advisors".  So they probably got "more" than most of us from such arrangements.



>
>Joel Benjamin worked briefly with the IBM Deep Blue team in their preparation
>for the match with Garry Kasparov in Philadelphia in February 1996, and has
>worked with the team coaching Deep Blue in preparation for the rematch since
>last fall
>
>kburcham


This matches with statements Joel has made, so that sounds true.  Of course
there were other GM advisors as well...



This page took 0.01 seconds to execute

Last modified: Thu, 15 Apr 21 08:11:13 -0700

Current Computer Chess Club Forums at Talkchess. This site by Sean Mintz.