Author: Christophe Theron
Date: 20:25:51 04/13/98
Go up one level in this thread
On April 12, 1998 at 11:54:07, Peter Kasinski wrote:
>Tactics dominate chess. Many strategically healthy games are lost to
>calculable tactical shots. At all stages of the game, various forms of
>contact develop between the pieces and the squares they control. In this
>sense, a combination is a coordinated series of contacts. How long are
>they? The following is a basis for discussion.
>
>1. Let’s start with an arbitrary number n (n=30 is fine).
>
>2. Legal chess produces positions where combinations of less than
>n-plies are present. Right after 1. g4 e5 2. f3? a known possibilty of
>length=1 is available.
>
>3. In theory, a tactical analysis of all master-level games is possible.
>Just take Deep Blue and demand a tactical check (to the depth of
>n-plies) of all positions reached in those games.
>
>4. Presumably, this analysis discards many “quiet” positions, and
>returns a large number of cases where a tactical shot of k-plies (k < n)
>exists.
>
>5. After exhaustive analysis, we get a distribution showing how many
>combinations were found for each number between 1 and n-1.
>
>Hypothesis: There is a number q (q < n) which limits most combinations,
>i.e. longer shots are physically possible, but statistically
>insignificant. Consequently, slow searching programs will benefit more
>from each additional ply (speed of hardware) than programs already
>approaching q in their search.
>
>Hypothetical distribution:
>
>q-3 75% Program A (slow searcher)
>q-2 85%
>q-1 92% Program B (fast searcher)
>q 94% of all tactics found
>q+1 95%
>q+2 96%
>
>If A and B are equally strong today, I like A’s chances much better if
>we double the speed of both CPU’s.
>
>I recently played a few games between Fritz 5 on PII 375MHz,160Mb RAM
>(128Mb hash) and Rebel 9 on AMD K6-300 MHz. (Yes, both machines are
>overclocked.) Level 40/120, 20/60.
>Hardly a scientifically sufficient sample, but for the first time Rebel
>9 didn't seem to be (to quote Amir) "hopelessly outgunned tactically"
>and, in fact, produced a positive score. Something to it?
>
>cheers,
>Peter Kasinski
It may also be that when you are able to compute deeper you need less
knowledge. So you are able to keep in your program only well verified
knowledge, and it gives you a play of better quality.
I take an example. Maybe this particular example is irrelevant, but I
hope you can catch my idea.
Most of the programs have these 2 pieces of knowledge:
1) put your rooks on open files
2) move your rooks inside the opponent's side (7th or 8th ranks are
great)
If I have a very deep searcher, maybe I can take knowledge #1 out of my
evaluation and still have a very good play. Maybe even better.
Because if my program is able to compute deep enough, it will see that
it is possible to move its rooks into the opponent's side by first
putting them on open files! And that it is useless to put its rooks on
open files if, for any reason, they will never be able to move to the
7th or 8th ranks.
Just an example. I think it is not true in this particular case, but
it's the general idea.
The general accepted idea is that, since we have computers fast enough
to compute deep, we should now add knowledge. Even if it slows down the
programs.
Maybe it's wrong? Maybe it's time to take some knowledge off?
Another brick in the knowledge/speed debate...
Christophe
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.