Author: Vincent Diepeveen
Date: 17:43:39 06/18/01
Go up one level in this thread
On June 18, 2001 at 19:12:18, Gian-Carlo Pascutto wrote:
The DB team itself wrote that SE is hardly giving them
extra playing strength. They wrote that when they searched
a bit deeper. that is, deeper as 8 ply. chiptest only got 8 ply
with 500k nodes a second.
How many people here nowadays get 8 ply with 500k nps at 3 minutes
a move?
Of course when playing 6 ply searching pc programs this 500k nps
at 8 ply kicked butt.
I have in DIEP several implementations tested of SE and similar
extensions that do forced things. What i'm using now i'm not
too happy to reveal. Only that it takes 0.5 ply when i'm
getting 10 ply in a quiet position and that it takes more like
5 ply when i'm getting a 15 ply search.
Obviously it has to do with threats and such :)
In any way, i see no difference in score in any auto232 run of
diep versus a load of different programs.
It only might score worse in fact.
the *only* reason this stuff is in diep is because of 2 reasons:
a) at tournament levels i hardly get above that 10 ply,
and getting 10 ply WITH extensions is no much different
to getting 10.5 ply without from positional viewpoint
b) it of course is tactical hell better, this is the main reason
i do it. To solve TESTSETS. Diep is 20 times slower as other
programs, so i MUST solve testsets with some other tricks.
Using Amir's approach is 1, using a bit more extensions is another
possibility of course
As soon as i'm getting above that 10 ply at tournament level i definitely
have to rethink the way worse b.f. i have nowadays, even though it still
is about what most programs have as b.f., i used to have such a
much better b.f. a few years ago, that i am already fearing the time
that i am getting each search >= 12 ply, because the overhead is
definitely not worth it then from positional viewpoint. Losing a ply or
2 to 3 then is definitely a big price for finding tactics a bit sooner.
DB never was in that position of course. The first time in their history
they got > 10 ply were in those games.
If i rememberwell the deep blue chessprocessors were only ready 2 weeks
before game start.
No way to test with that 12 ply search depth then a lot or to rethink
extension choices.
They searched deeper as they did before.
The whole discussion about SE is kind of stupid here, it's simply a
huge overhead, but obviously *no one* ever reported it to be an increase
in playing strength.
It probably hurts me more as it is helping me. For sure it doesn't help me
at standard time controls.
The only other program in the world that's combining SE with fullwidth
search is genius. But it also forward prunes a lot near the leafs.
So obviously a lot of the extra overhead which SE generates when never
pruning, this overhead is taken away by forward pruning inside genius.
How deep does genius search at 4 hours a move, WITH forward pruning?
Still not very impressive deep!
12 ply or so?
>On June 18, 2001 at 11:19:41, Bas Hamstra wrote:
>
>>Now you sound exactly like Bob. Noone is disqualifying their program. At the
>>time unbeatable. But it *is* possible to compare search model A with search
>>model B and conclude that B is better. DB is not a magical black box that we
>>know absolutely about.
>
>If you don't know the exact details your comparisation is going
>nowhere.
>
>The SE is a nice example of this. The SE that Vincent tested is
>totally unlike what DB used. It's tuned for a chessprogram like
>crafty, not DB.
>
>I have one that is closer to DB, but Vincent didn't test that of
>course. Also, depiste all publications MANY of the details are
>totally unsure. SE is VERY complex. And you're not ever going
>to get something close to what DB did because they simply left
>out way too much details.
>
>I am telling you now that you are never going to get this
>comparisation to work. Not unless Hsu fills in the missing
>details.
>
>>We know they didn't prune. So they could have even been
>>stronger.
>
>Hello? How can you know this? The better your eval is the
>more pruning is going to hurt.
>
>--
>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.