Author: Christophe Theron
Date: 14:02:52 12/05/99
Go up one level in this thread
On December 03, 1999 at 20:48:17, Vincent Diepeveen wrote:
>On December 03, 1999 at 16:42:31, Will Singleton wrote:
>
>>On December 03, 1999 at 08:32:39, Vincent Diepeveen wrote:
>>
>>>Current DIEP version, so no tricks, no extra extensions turned on
>>>as the ones that are currently turned on (all are out except check
>>>and a passed pawn extension), limited at
>>>6 ply against amateur at 6 ply.
>>>
>>>So limiting *all moves* it to 6 ply, including endgame.
>>>then DIEP at 10 ply against amateur at 10 ply.
>>>
>>>No pondering as we play with near to infinite time.
>>>If you want to i can also run everything on a single processor, so
>>>that transpositiontable luck because of parallellism
>>>at 10 ply is not gonna get blamed for the higher % afterwards by some
>>>scientific dudes that agree with you.
>>>
>>>Just searching n ply against n ply, without tricks as suggested some
>>>years ago like modifying eval to be an eval that did a 4 ply search.
>>>
>>>Just a 2 minute change that turns off time-check in program and forces
>>>it to
>>>use the depth=n setting.
>>>
>>>What i predict is a much closer to 50% score for the 6 against 6 ply
>>>match,
>>>and a complete annihilation of amateur when both searching at 10 ply.
>>>
>>>I assume here already of course that DIEP's eval is better than
>>>Amateurs.
>>>If i would consider the evaluation of amateur a lot better then i would
>>>obviously predict the opposite outcome (amateur annihilating diep at
>>>10 ply and nearer to 50% score at 6 ply).
>>>
>>>Of course things like learning and such must be turned off.
>>>Preferably both versions playing the same openings. However by just
>>>playing a large quantity of games we can measure statistically accurate
>>>what has happened.
>>>
>>>To make the experiment even better we should log the evaluations too
>>>and see where things went wrong and calculate whether that had
>>>statistical influence on the outcome.
>>>
>>>Also we could start with the same book if you want to.
>>>
>>
>>Vincent,
>>
>>An interesting offer, sounds like fun. However, it would be worthwhile only if
>>
>>1) Our programs were bugfree. Mine is not.
>>2) Extensions are identical.
>>3) Search and Qsearch are identical.
>>4) HT and null-move are identical.
>>5) Books identical.
>>6) Independent verification of all this.
>
>Except for the books this is not needed.
>
>>What you are trying to measure is the proposition that, given unequal evals,
>>games between deep searchers will enable the better eval to win more often than
>>games between shallow searchers, due to the vagaries of tactics (which you
>>contend are pretty random at 6 ply, but not at 10 ply). I contend that tactics
>>would continue to be significant at 10 ply, though *perhaps* not to as great a
>>degree as 6 ply.
>>
>>So I think it would be a worthwhile experiment, but not using unequal programs.
>>Use unequal evals, with the same program.
>>
>>I will make my eval suck even more than it does now, and play a 100 game match
>>at 6-ply and then at 9-ply levels (mine won't reach 10 too often, besides
>>there's some bug that causes an overflow in something when I get too deep).
>>
>>If you or others wish to do the same experiment, then we might be able to get
>>some usefuls results.
>
>This experiment is not good, as we don't talk about an eval that's
>tuned to play well with little info, but just a sucking eval.
>
>You need to tune your sucking eval first like nimzo and junior
>and chesstiger are tuned before starting the experiment.
Or degrade your sucking eval to the level of Diep.
If my evaluation sucks, Vincent, what should I say about yours?
Christophe
This page took 0.01 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.