Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: ACM1994

Author: Vincent Diepeveen

Date: 17:09:19 09/20/05

Go up one level in this thread


On September 20, 2005 at 16:35:43, Robert Hyatt wrote:

>On September 20, 2005 at 16:22:58, Vincent Diepeveen wrote:
>
>>On September 20, 2005 at 15:27:14, Bruce Moreland wrote:
>>
>>>On September 20, 2005 at 13:16:35, Robert Hyatt wrote:
>>>
>>>>On September 20, 2005 at 12:59:03, Uri Blass wrote:
>>>>
>>>>>On September 20, 2005 at 11:58:35, Robert Hyatt wrote:
>>>>>
>>>>>>On September 20, 2005 at 02:35:50, Bruce Moreland wrote:
>>>>>>
>>>>>>>On September 18, 2005 at 14:47:26, Robert Hyatt wrote:
>>>>>>>
>>>>>>>>On September 18, 2005 at 14:38:27, Uri Blass wrote:
>>>>>>>>
>>>>>>>>>On September 18, 2005 at 13:59:18, Robert Hyatt wrote:
>>>>>>>>>
>>>>>>>>>>On September 18, 2005 at 12:00:09, Vincent Diepeveen wrote:
>>>>>>>>>>
>>>>>>>>>>>On September 18, 2005 at 10:45:19, Robert Hyatt wrote:
>>>>>>>>>>>
>>>>>>>>>>># Name 1 2 3 4 5 P BU SB G
>>>>>>>>>>>1 Deep Thought II 2w1  5b0  7w1  3b1  5b1  4  13½ 11  5
>>>>>>>>>>>2 Zarkov          1b0  6b1  4w1  5b=  3w1  3½ 15  9¾ 5
>>>>>>>>>>>3 Star Socrates   10w1 7b1  5w1  1w0  2b0  3  12½ 5  5
>>>>>>>>>>>4 Now             6w=  10b= 2b0  8w1  9b1  3  10½ 5½ 5
>>>>>>>>>>>5 Mchess Pro      8b1  1w1  3b0  2w=  1w0  2½ 16½ 7¾ 5
>>>>>>>>>>>6 Cray Blitz      4b=  2w0  9w1  7b0  10w1 2½ 11  4  5
>>>>>>>>>>>7 Wchess 9w1      3w0  1b0  6w1  8b0       2  13½ 4½ 5
>>>>>>>>>>>8 Evaluator       5w0  9b0  10w1 4b0  7w1  2  10  2½ 5
>>>>>>>>>>>9 Innovation II   7b0  8w1  6b0  10b1 4w0  2  10  2½ 5
>>>>>>>>>>>10 Spector        3b0  4w=  8b0  9w0  6b0  ½ 12½ 1½ 5
>>>>>>>>>>>
>>>>>>>>>>>Zarkov from those days had no problems beating your 3 million nps Cray Blitz.
>>>>>>>>>>>Nor had Wchess problems beating your 3 million nps Cray Blitz.
>>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>>Ask John about the game.  First, this was a 500K program for rounds 2-4.  And
>>>>>>>>>>when you ask him, he'll tell you about our rather severe crash problem due to a
>>>>>>>>>>missing test to limit ply to 64 or less.  And in a couple of cute places, we
>>>>>>>>>>went beyond that limit, crashed, and burned.  We fixed it for the last round,
>>>>>>>>>>but it really didn't matter to the final results.
>>>>>>>>>>
>>>>>>>>>>But notice the issue was about deep thought, _not_ about Cray Blitz.  Did you
>>>>>>>>>>see any of the micros coming close?  (hint:  round 2 was a forfeit which is why
>>>>>>>>>>they were paired a second time, round 2 never got started for the DT MCP game).
>>>>>>>>>
>>>>>>>>>Deep Thought was significantly better than the micro of 1995
>>>>>>>>>
>>>>>>>>>Uri
>>>>>>>>
>>>>>>>>I am well aware of that.  And DB was 100x faster than deep thought 2, and also
>>>>>>>>had a better evaluation...
>>>>>>>>
>>>>>>>>that was my point in all this...
>>>>>>>
>>>>>>>I think that Hsu is a pretty arrogant person, after having read some stuff that
>>>>>>>he has said about himself and other chess programmers.
>>>>>>>
>>>>>>>Read his book, if you can stand it.
>>>>>>
>>>>>>I did.  It didn't particularly turn me off.  But then I have had dozens of
>>>>>>face-to-face conversations with him dating back to 1987 in Orlando at the ACM
>>>>>>event that year, continuing thru the point where he left IBM a few years back.
>>>>>>
>>>>>>>
>>>>>>>For years you have taken a few things he has said as truth, but given his
>>>>>>>personality I am not sure if they are true.  He may think they are, but this
>>>>>>>doesn't mean that they are.
>>>>>>
>>>>>>That is probably all in perception.  I took very little of what he said at face
>>>>>>value, without supporting evidence.  He was usually more than happy to sit down
>>>>>>with deep thought and play with positions to see how it would react.  And he
>>>>>>never wanted to "hide" the display so I could not see.  I didn't trust the SE
>>>>>>data, since they gave conflicting reports on the effect (was it +70 or +7 rating
>>>>>>points better) so I simply tried it for myself in Cray Blitz.  And I think that
>>>>>>+7 was closer to the truth although in the right positions it was much more than
>>>>>>that...
>>>>>>
>>>>>>
>>>>>>
>>>>>>
>>>>>>>
>>>>>>>He hung you out to dry by leaving you as primary defender of Deep Blue for
>>>>>>>years, while he left the trivialities of computer chess to mere mortals like us.
>>>>>>> The DB project doesn't deserve defense.  It hit computer chess like a
>>>>>>>carpetbagger, then left in the night when the money was gone.
>>>>>>>
>>>>>>>bruce
>>>>>>
>>>>>>
>>>>>>However, I would not attribute that to Hsu or Campbell.  They were active in
>>>>>>computer chess for many years.  IBM pulled the plug for obvious reasons after
>>>>>>they hit the peak of Mt. Everest...
>>>>>>
>>>>>>I'll always have a great deal of respect for the group.  PVS search was first
>>>>>>used in my program by accident, as Murray and I played with it at an ACM event
>>>>>>on a machine we were not using, but a machine we had to use due to a power loss
>>>>>>during a key round.  Singular Extensions was Hsu's idea, and it certainly
>>>>>>worked, since many are using it today in various forms.
>>>>>
>>>>>I know nobody who use it in the way that Hsu used it.
>>>>>
>>>>>Uri
>>>>
>>>>
>>>>Several have used PV-singular as defined by Hsu.  Kittinger was one, Lang was
>>>>another.  I'm not sure anyone did the FH-singular extension as defined by Hsu
>>>>because it is _very_ costly.  But I did it in Cray Blitz and it worked just
>>>>fine.  I tried it in Crafty and it didn't work just fine.  Whether that was a
>>>>result of the null-move stuff with R=2/3, vs Cray Blitz with R=1, I don't really
>>>> know since I was not interested in crippling Crafty's search to try them
>>>>further...
>>>>
>>>>One day I might, again...
>>>>
>>>>However, just because no one did FH-singular doesn't mean the idea is wrong.  No
>>>>one does it because of the cost.  When you are 1000x faster than your nearest
>>>>competitor, you can give up a factor of 10x to implement something like that,
>>>>and _still_ be 100x faster than them which is more than enough to avoid
>>>>trouble...
>>>>
>>>>People copy ideas that work well with low cost, they often avoid things with
>>>>high cost because that cost will more than offset the gain if you don't have
>>>>special-purpose hardware to hold the cost down.  I could make a list of a few
>>>>key endgame features that appeared in Crafty before they were in any commercial
>>>>programs.  Outside passed pawns in 1995 is one.  Then candidate passed pawns
>>>>(distant majorities).  etc.  If you go back to 1976 and "blitz" we had the
>>>>"passed pawn race" code already in place so that we could detect uncatchable
>>>>passed pawns in king and pawn endings.  Most everyone does that today.  Because
>>>>the costs are more than offset by the gains.  FH-singular might not be there
>>>>yet, but that is why no one is using it, not that it does not work.  Because it
>>>>is clearly correct theoretically.  Just too expensive practically...
>>>
>>>It also mean that it does work.  When your competition is out-gunned, everything
>>>you do works, including the stuff that doesn't work.
>>>
>>>This is not to say that the stuff doesn't work.  If you want an example of
>>>someone who does FH singular extension, albeit in restricted form, look at me.
>>>When I fail high I keep searching at reduced depth, although not all the time.
>>>
>>>DT/DB may have been a great system.  I don't know, because I don't trust the
>>>information that I have seen about it, and I can't examine it for myself.
>>>
>>>I don't trust the DB guys to make the determination.  You can listen to any
>>>number of people talk about how their program is super-advanced and does this
>>>and that, and when you play the program on equivalent hardware, it sucks.
>>>
>>>Hardware can mask a lot of problems.  Essentially you said so yourself when you
>>>talked about DB "giving up a factor of 10" to SE.
>>>
>>>I would be absolutely insane to "give up a factor of 10".  My program gives up a
>>>ply due to its use of SE, but it doubles its tactical speed.  So I gave up
>>>nothing -- SE doubles my program's speed, by my estimation, despite reducing its
>>>average depth.  That is the world I operate in.  If something makes my program
>>>suck, I have to remove it, because I can't afford to suck, because other people
>>>who don't suck have big enough hammers that they can break my skull if I do.
>>>This is very Darwinian.  Something that is occupies a niche by itself can afford
>>>to be inefficient.  Once it has competitors who make use of the same resources,
>>>it had better become efficient unless it wants to end up obsolete.
>>>
>>>Honestly, none of the big iron seems very efficient to me.  You invented a lot
>>>of techniques, and you deserve to be one of the first few guys in the CC hall of
>>>fame, but I bet you got a lot better when you had to fight Fritz on equal
>>>footing.
>>>
>>>bruce
>>
>>Please realize 1 factor. At 480 cpu's using SE is very difficult.
>>
>>The huge number of tiny searches it generates are a major problem.
>>
>>In fact i made special splitting code in order to be able to use SE at a quad
>>even. It scales the program better already at a quad. At 16 cpu's i couldn't run
>>with SE initially without making the special splitting code.
>>
>>Above 16 cpu's, forget using SE. Too many tiny searches.
>>
>>SE loses diep less than a ply usually. Usually about 0.5 ply. The problem i have
>>with SE is that in testgames it scores better without SE, because that 0.5 ply
>>is an extra "positional 0.5 ply" whereas SE is just tactical and at 14 ply
>>search depths there is no magic tactics that win the game suddenly.
>>
>>Only in testsets there is.
>>
>>Additional SE sees tactics after the game has been decided already. When you
>>need it, like in seeing perpetual check, something that does make sense to see,
>>then it doesn't see it.
>>
>>Vincent
>
>
>I don't believe your last  statement has anything to do with reality.  Bruce and
>I played lots of games when we were both fiddling with SE.  We played a big
>batch with him using SE and me not.  And in many games, suddenly BOOM, ferret
>would find some deep tactical shot that I would not see until a couple of moves
>later.  Without SE, our depths were pretty similar, and our tactics were as
>well.  With SE, he was quite noticably stronger.  However, Crafty on the other
>hand was not.  And before you start on q-search checks, this was at a time where
>_neither_ were doing them..  I don't know if he does today or not.
>
>I'll remind you that 5 years ago you continually said (a) clusters can't work;
>yet you used one;

Huh?

Want to call the 512 processor origin3800 a cluster?

It's a supercomputer in any kind of respect.

Note that todays clusters have a better latency than the supercomputers
used to have.

That changes things *bigtime*.

>(b) SE is no good;  yet you always claim to have it
>implemented but usually turned off;

Actually i had some selfinvented singular extension turned on in 1994-1996 then
a much improved one was there in 2001 (thanks to bruce) until world champs 2005
it was turned on always in tournaments. Usually this is a last second decision
to turn it on or off.

I plan to turn it off for next tournament though, as i'm in home tests running
right now without it and i like it.

But the real bottom line is:

If YOU believe SE works, knowing diep uses it, ferret uses it, what stops you
from turning it on in crafty?

You say it is a magic extension, yet you are the only one not having it turned
on in tournaments.

Then either your statement is false, or you're doing something wrong in crafty.

*my* statement, after running with it for 4 years now is that at todays search
depths in TOURNAMENTS at 4 cpu's or above, i see no point in using SE, because
it does tactically improve an engine *usually*, but there is simply no trick
there where you can win with. So it is a lost 0.5 - 1 ply search depth
effectively.

Vincent





This page took 0 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.