Author: Uri Blass
Date: 22:34:25 10/05/01
Go up one level in this thread
On October 05, 2001 at 16:45:36, Christophe Theron wrote:
>On October 05, 2001 at 07:57:43, Uri Blass wrote:
>
>>On October 04, 2001 at 21:47:46, Christophe Theron wrote:
>>
>>>On October 04, 2001 at 17:15:18, Francesco Di Tolla wrote:
>>>
>>>>Neglecting the proablby important (bu not easy to interpret) effect of the
>>>>different CPU models, this is actually perfectly logic:
>>>>
>>>>Junior is a fast searcher and shredder has more knowledge.
>>>>
>>>>The faster the hardware that less important the extra search depth!
>>>>If you can reach say 20% more knodes this are a given amount of plys at a speed
>>>>an less plys at higher speed due to the nonlinear growth.
>>>>
>>>>So what's wrong?
>>>
>>>
>>>
>>>What's wrong is that if search tends to be less effective with increasing speed,
>>>actually "knowledge" (probably meaning "evaluation" in your mind) has exactly
>>>the same problem.
>>>
>>>In chess, Search <=> Knowledge
>>>
>>>I believe in dimishing returns from improved search, and I also believe in
>>>dimishing returns from improved knowledge ("improved evaluation").
>>>
>>>
>>> Christophe
>>
>>I do not believe in diminishing returns from improved knowledge.
>>The opposite.
>>
>>I believe in increasing returns from improved knoweldge.
>>
>>I think that doing the next experiment may be interesting(I suggest it for tiger
>>but it can be also done with other programs):
>>
>>1)Generate a clearly weaker version of tiger by changing only the evaluation to
>>only material evaluation+some small positinal scores for the piece square tables
>>
>>The weaker tiger is not going to know nothing about passed pawns or double pawns
>>
>>2)play matches at fixed depthes between the original tiger and the weaker tiger
>>and get rating for both tigers
>>
>>I believe that you are going to find bigger difference in rating between the
>>tigers at big depthes.
>>
>>My example is extreme example but I think that we can learn to guess from
>>extreme example about the practical cases.
>>
>>The problem is that in the practical cases the difference in evaluation is
>>relatively small so we may need too many games to prove that you are wrong.
>>
>>I did not suggest only material evaluation because I am afraid that the only
>>material evaluation is always going to lose so we cannot learn much from it but
>>I believe that only material +some small scores for piece square table may be
>>enough for the weak tiger at depth 10 to beat the original tiger at depth 3 or
>>4.
>>
>>Uri
>
>
>
>It would be interesting to do the experiment, but I think it is going to be hard
>to prove the point. As hard as proving it for dimishing returns from search
>depth.
>
>However if somebody is really interested, I can provide two versions of Tiger:
>1) the normal version
>2) the same version with a weaker evaluation (as suggested by Uri by removing
>the evaluation of passed pawns for example).
>
>
>
> Christophe
I think that it may be more easy to prove increasing returns from evaluation
than deminishing returns from search because it is more easy to compare big
difference in knowledge.
In the case of knowledge I can easily play match like 3 plies of normal tiger
against 7 plies of weaker tiger and 6 plies of normal tiger against 10 plies of
weaker tiger.
I proved that at these depthes there is not a significant deminishing returns
from search in the nunn2 match so if I prove that 6 plies of normal tiger wins
convincingly 10 plies without knowldge when 3 plies of normal tiger is losing
convincingly against 7 plies without knowledge then it is going to be a
convincing proof of incresing returns from knowledge.
I suggest that you drop knowledge and stop doing it only when you see that 3
plies of normal tiger gets 15-25% against 8 plies of weaker tiger in the nunn2
match.
After doing it you can try 3 plies of normal tiger against 7 plies of weaker
Tiger and 6 plies of normal tiger against 10 plies of weaker tiger.
I suspect that you may see something like 30-20 for weaker tiger in the first
case and 30-20 for normal tiger in the second case.
This is not enough evidence and we need more games but doing the sweningen
tournament(normal tiger 3 plies,4 plies,5 plies ,6 plies) against weaker
tiger(7 plies,8 plies,9 plies 10 plies) may give 200 games for every program
and the relevant evidence
Note that I prefer the nunn2 match and not using opening books because I believe
that part of the practical deminishing return from depthes is because of the
opening book.
It is clear that there is an increasing returns from the opening book when the
depth is bigger(in the extreme case when the engines get infinite depth the
opening book has 100% influence).
Uri
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.