Author: Christophe Theron
Date: 11:35:09 11/07/00
Go up one level in this thread
On November 07, 2000 at 01:48:44, Uri Blass wrote:
>On November 07, 2000 at 00:44:43, Christophe Theron wrote:
>
><snipped>
>>Programs of 1990 include the Mephisto line on 680xx. In 1990, it was the
>>Mephisto Lyon. I guess it was running on a 68030.
>
>I find Mephisto lyon on 68030 36 mhz in the ssdf list(rating of 2151 that is 479
>elo weaker than Fritz6a)
>
>I believe that it was an hardware dedicated for chess and it means that it
>cannot run on 1Gh computers.
No, it was simply a 68030 processor+RAM+ROM+I/O. A classical computer actually.
><snipped>
>>It's just an exception. Generally, 12 plies of extensions is more than enough.
>
>It was more than enough for slow hardware but I believe that it is often not
>more than enough with the hardware of today in tournament time control.
>
>Remember that the 12 plies of Genius included also the Qsearch.
What you need to know is that Genius uses a SEE. That means that it simulates a
QSearch (capture search) after each leaves.
So at the end of the ply 12 of the QSearch, it takes into account the captures
sequences that start at the 13th ply and on.
So in practice his QSearch is not limited to 12 plies.
><snipped>
>>If doubling speed accounts for only 30 elo (to take into account an hypothetical
>>"dimishing returns"), that still makes 299 elo points.
>>
>>I don't believe there is 300 elo points of difference between current best
>>programs and the best program of 1990 if they are run on equal hardware.
>
>The difference between 500Gh and 1 Gh is almost 9 doubling and it means 270 elo
>points difference(assuming 30 elo for doubling).
The formula is elodiff=30*log(speedratio)/log(2)
In our case 30*log(1000)/log(2)=298.9735285398626... which I rounded to 299.
>Here is some information about the difference in rating based on the ssdf list.
>
>Fritz6 is 78 elo better than Fritz5.32 on the same hardware(450Mhz).
>Fritz5.32 is 56 elo better than Rebel9 on the same hardware(200Mhz)
>Rebel9 is 31 elo better than Genius4 on the same hardware(p90)
>Genius4 is 109 elo better than Fritz3 on the smae hardware(486 50-66 mhz)
>
>Fritz3 came some years after 1990 and the difference is 78+56+31+109=274.
But Fritz3 was not the best at this time.
To be fair, we should evaluate the software improvement between the Mephisto
Lyon and the more recent Genius programs.
I think it is possible to do so, because Lang has ported back Genius to the
68030. I think this Mephisto computer is called Mephisto Genius 68030.
Here are the relevant lines in the August 2000 SSDF list:
71 Meph Genius 68 030 33 MHz 2198 45 -44 248 55% 2161
84 Mephisto Lyon 68030 36 MHz 2151 28 -27 716 68% 2016
The difference in elo is 47. If you take into account the MHz difference, which
account for approx. 9 elo points (70*log(36/33)/log(2)), then the elo difference
is 56 points. I know the error margin is big, but we have only this.
The Mephisto Genius 68030 has been released in 1994 and is supposed to be a port
of Genius 3 for PC to the 68030 architecture.
Now I leave to you the task to evaluate how Genius 3 would perform on current
SSDF top hardware (K6-2 450).
I would be surprised to see a 300 elo difference between Genius 3 and current
best programs.
And if it is the case, remember that 300 elo has been evaluated assuming a big
dimishing returns effect of 30 elo per speed doubling. The actual value on
current computers is 70 elo per speed doubling.
>>Branching factor at very slow time controls can also be an issue, but I don't
>>believe 1990 programs had such a terrible branching factor.
>
>The only way to know is by testing.
>I did not test programs of 1990 but I know that hiarcs7.32 starts to have
>terrible branching factor at long time control(I guess that 6 or 7 is the
>average branching factor of hiarcs at long time control but I did not do
>statistics about it).
>
>My experience with the branching factor of Genius3 at long time controls is
>better and 4 is the typical branching factor.
>
>I suspect that part of the programs have terrible branching factor at long time
>control because they were not tested at this time control.
It comes from the pruning algorithms they were using.
These algorithms were very effective near the horizon, say on 3 to 5 plies near
the horizon, but very inefficient after that.
So on a 10 plies search, for example, they had a good branching factor on the 5
plies near the horizon, but the BF was terrible on the 5 first plies near the
root.
So the deeper the search, the closer to 5 the branching factor will be.
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.