Computer Chess Club Archives


Search

Terms

Messages

Subject: Re: NPS, EBF, Knowledge and Elo

Author: Ricardo Gibert

Date: 18:09:55 08/15/03

Go up one level in this thread


On August 15, 2003 at 14:44:18, Sune Fischer wrote:

>On August 15, 2003 at 14:03:12, Ricardo Gibert wrote:
>
>>
>>So if hypothetically we were to actually have 32-man EGTBs, you would expect the
>>program using them to play equally strong to a program not using any EGTBs?
>>
>>It helps to know the reasons why EGTBs do not significantly add strength to
>>programs. Applying empirical results without understanding the reason for the
>>result will inevitably lead you astray.
>
>Because 99 times out of 100 the game has already been decided when they reach
>the EGTBs. And the tables are too slow to be used at the leafs from great
>distances.
>
>
>>>Yet another type of knowledge is search depth.  So I don't think it is possible
>>>to separate the two cleanly.
>>
>>
>>I'm actually very interested in not separating them cleanly. I'd like to know
>>the relationship between search depth knowledge and eval knowledge. It is useful
>>to know which one to try to emphasize.
>
>They go together like yin yang.
>
>What is too expensive to search you evaluate, and what is too complex to
>evaluate you search.
>
>-S.

Question: What happens if you reduce the number of bits alotted to the eval
score by eliminating the lower order bits used in the eval score? In other
words, you make the score more coarse. This effectively reduces the knowledge
used in eval by a quantifiable amount. What is the impact on playing strength?

For example, you could compare using deci-pawns instead of centi-pawns.

For a more concrete example, you could compare using the upper 8-bits of the
eval score instead of say 12-bits. The amount of knowledge reduced would be
4-bits or 33%.

For a given percentage reduction of bits alotted to the eval score, what is the
impact on strength in terms of elo points? What would the formula for such a
relationship look like? If we plotted elo scores for 6, 7, 8, 9, 10, 11 and
12-bit evals, what would the resulting curve look like?

I suspect the lowest order bits used in evals might just be noise (e.g.
millipawns), but as you move up, a measurable effect on playing strength can be
discerned. I would expect a coarser eval to gain in search depth, but not enough
to offset the loss of information. I would expect some type of useful
conclusions could be drawn. What would you expect?



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.