Author: Albert Silver
Date: 09:18:00 09/01/04
The following is a copy of a post and discussion at SchachComputerWelt
(http://f50.parsimony.net/forum200336/messages/3149.htm) as the discussion may
interest others experimenting with Pro Deo's settings.
>>>man beachte die diametralen Bewertungen in diesem Bereich...
>>>
>>>[Event "pd mach93_18 - f8 , Blitz:25'+2""]
>>>[Site "ORION8"]
>>>[Date "2004.08.28"]
>>>[Round "17"]
>>>[White "Liseth 1.0"]
>>>[Black "Fritz 8"]
>>>[Result "1-0"]
>>>[ECO "E55"]
>>What are your settings?
>> Albert
>in this case e.g. it was 93_18
>I am experimenting with styles.
>i have nice evaluation stuff, but the search stuff is not the way i want it to be.
>reason is that most of the search stuff switches do not really bring an advantage.
>maybe we can make a discussion about this that leads to progress in the decision which search switches make sense.
>if wished i could publish the style behind the games and if needed, explain what is different.
I'd like that.
>i doubt that most of the search switches of ed make really sense. the potential is mainly in his evaluation switches.
>here pro deo has made progress IMO.
It's interesting that you mention this, because yesterday I spent quite a bit of
time examining exclusively the search options and carefully comparing the
results. One thing is clear, changing the [Search Selectivity] has only given
worse results. I did some extreme changes it is true, like reducing it to 50,
making it more Null-Move (0 = complete Null-Move) and then trying increasing
this to 150 to see what happened. In all cases the results were much worse, so I
leave this option alone. A big problem with this option is that Ed doesn't
explain clearly what happens if it is increased. If it is reduced to zero it is
full null-move, but what happens if it is increased?
Anyhow, I questioned his conclusion on the effectiveness of increasing the
[Adjusted Selectivity] for adjusted maximum selective depth. I have found that I
can get many times better results increasing them all by 2, so it reads 8-7-6-5
instead of 6-5-4-3 AND increasing the Chess Knowledge to 150.
[Chess Knowledge = 150]
[Adjusted Selectivity MIDG = 8] * Maximum selective depth (midgame)
[Adjusted Selectivity END0 = 7] * Maximum selective depth (early endgame)
[Adjusted Selectivity END1 = 6] * Maximum selective depth (endgame)
[Adjusted Selectivity END2 = 5] * Maximum selective depth (simple endgame)
More than that has shown worse results overall, and increasing the Adjusted
Selectivity also.
I compared this with the default settings with no [Adjusted Selectivity] and
also adding the proposed 6-5-4-3 to the settings aned leaving the Chess
Knowledge at 100. Both yield solid results so far, and it is not 100% clear to
me which is best. I have found one position where the 8-7-6-5 + 150 combination
simply doesn't find a solution in a reasonable time,
[D]8/6pk/4p2p/p1QpP1nP/Pr6/1P6/6P1/6K1 w - - 0 64
but it is hard to know if any conclusions should be drawn from this isolated
case. I haven't found others but I also haven't done enough testing.
I also tested the 3 settings (the default, default plus 6-5-4-3, and my 8-7-6-5
+ 150 knowledge) with and without the new:
[Pruning = MISC_03] * move ordering speedup
[Pruning = MISC_21] * Lazy Eval (tuning)
[Checks Depth = 1] * limit QS checks to 1 ply
and adding all 3 has systematically led to an overall icnrease in performance.
Alone the Lazy Eval has led to either no or worse performance, but
interestingly, in combination with the other two it improves more often than
not, the search than with the other two alone.
On the positional settings, I'll add that the new [King_Safety] option has led
to mixed results, and that the new Passed Pawn code:
[Pruning ? MISC_58] * Recalculate passed pawns
[M58_Table ? Normal] * normal|midg|endg|strong
should be left at "normal". I tried using the "strong" setting and Pro Deo
started doing the oddest things, over-evaluating passed pawns far too much.
Albert
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.