Author: Vincent Diepeveen
Date: 17:02:53 06/23/01
Go up one level in this thread
On June 23, 2001 at 18:19:18, Gian-Carlo Pascutto wrote:
Some people seem to be too stupid to understand
what the overhead is on numbers of nodes when you
search
a) fullwidth (so not a SINGLE line you ever extend gets
ever of a smaller depth, it only exponentially
adds up to the horror
b) without hashtable
As long as you don't understand the difference for search
depth, how can i make *anything* clear to you?
Because the only thing i wanted to make clear is the
big difference in search depth with nullmove + hashtable.
When i posted this in CCC that for diep i needed billions
of nodes to get 11 ply, then bob said that if he modified
crafty to fullwidth with SE, that he could get easily 17 to 19
ply (that's 11(6) to 13(6) meaning according to bob),
that it was only my horrible program that prevented to
search deep.
So even a very bad SE implementation in crafty was enough
to proof my point.
We're talking about a bound here. *for sure* DB will have
used more nodes as what i did to crafty (last 6 ply no
hashtable and SE at depths >= 6 ply).
Hope i'm very clear now.
In 1999 it still was not obvious to some scientists
that nullmove would outperform anything when you get
more nps, also nullmove would be too dubious, apart from
its reduction factor, it would give back scores so dubious
that it would mess up hashtable and give completely incorrect
information back to the root.
The CORRECTNESS of nullmove at depth n was simply doubted,
where i could depth n for a problem x.
Not to mention that important things as hashtables were neglected
and not seen as *that* important at bigger depths.
Also my statement that 17 to 19 ply with 200M nps was possible
WITH nullmove AND hashtables in 1997 i was laughed for around
1997.
"knuth proved that impossible".
Well, knuth applies to a search without hashtable and without nullmove
very well (so the last 6 plies of deep blue there Knuth's lemmas
are valid), but as soon as you add nullmove and hashtables to the
search they are no longer valid.
Nowadays in 2001 everyone is so used to the good branching factors
we have in 2001, that history has been forgotten.
Well, in WMCC 1997 and dutch open 1997 i searched 8 ply with diep,
i can't remember whether that was with last ply pruning, but definitely
with R=3 everywhere and with hashtables, and with a DOS interface.
I played against schach in dutch open for example and got completely
outsearched as schach got itself 9 to 10 ply or so. That was with
a quarter of a million nodes a second for Schach or so at a P133,
and WITH nullmove and WITH hashtable and WITH some kind of singular
extensions.
I do not know much details on schach, but if you give it an entire
night, then it still will show the same bad branching factor as
we were used to in 1997.
So posting that searching with big hashtables and with nullmove
would deliver a cool branching factor above 10 ply, that was
completely different as a statement would be made now.
Now amazingly something else gets claimed, namely that Deep Blue
was searching 17-19 ply. I'm AMAZED some people swallowed this
big joke.
in 1997 deep blue searched 12 ply FULLWIDTH (nullmove was too
dubious to take serious was the statement back then) and
its holy extensions would give a level increase which PC programs
would not get first 20 years.
Now the log files are released, we can see it searched 11 to 13 ply
and that it had loads of capture extensions and some extra check
extensions or something (which is of course giving tactics of
a bit bigger depth as 11 to 13 ply, but definitely no + 6 ply... ).
However now that PC programs all get above this search depth, then
suddenly the thing must look better as it was, and it is said now
that it searched 17 to 19 ply.
A bigger nonsense statement is not possible!
Knuth is true for the last 6 plies of DB, and considering they
had mobility and huge scores for things, this means obviously that
their sorting was definitely no better as mine is.
No one can argue against that!
Now a kid like you is intervening with stupid childish statements,
without realizing the historically facts and discussions.
All i wanted to show is what Bob defending Hsu by saying
crafty would proof me incorrect. I did modify
crafty for the last 6 plies and of course used
some extensions of yours. You will be the first
to admit that it doesn't cost as much as a full
implementation of SE is costing.
>On June 23, 2001 at 17:18:17, Vincent Diepeveen wrote:
>
>>Please play some 120 60 games (about 3 minutes a move)
>>or allow them to get played.
>
>It wouldn't prove anything either, would it?
>
>I knew in advance that you were going to defend yourself
>by using the timecontrol as an excuse, which is why I put
>in the 400x faster statement. We both know it's bullshit.
>
>We both know (I hope!) that calling this crafty 'Deep Blue'
>is bullshit too.
>
>You picked up our code that I gave you in good faith, crippled
>it and used it to make some statements about how bad Deep Blue
>was and how good Diep is.
>
>You didn't even bother to ask me about it. You didn't even
>bother to check the details of the SE implementation. Had you
>done that you would have known it was totally unlike Deep Blue
>and that I had a version that was much closer to what Deep Blue
>does.
>
>So this is my revenge. Two losses by Diep vs 'Deep Blue'.
>I know they don't mean anything. You know it too.
>
>But it's no different than your original post about 'Deep Blue'.
>
>I know you hate this post because you think it puts your program
>into a bad perspective in an unfair way.
>
>Yes, you're right. But it's no different from what you've been
>doing here. Perhaps even worse, as the Deep Blue guys can't speak
>for themselves.
>
>I know you have a great program, and I've been very happy to
>use it. I personally think it is _way_ underrated. If someone
>doubts it abilities they should check out the win over crafty
>in CCT3. (which you already predicted several months before,
>which also says a lot about _your_ abilities)
>
>So please, let's put this stuff behind us and get back to what
>we should be doing.
>
>I'll be happy to let FriarTuck play some 120 60 games vs Diep.
>
>But please don't use the result to say anything about Deep Blue,
>or I'll shoot you (*) when we meet IRL.
>
>(*) met een waterpistool, welteverstaan :)
>
>--
>GCP
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.