Not logged inRybka Chess Community Forum
Up Topic Rybka Support & Discussion / Rybka Discussion / Paramiters Sugestion
- - By mindbreaker (****) [us] Date 2010-07-24 00:06
I keep running into the problem of the value change increments of the pawns being to great.  This interferes with trying to find the optimum settings. Moving a pawn one tenth can totally destroy a setting.  I suggest that the settings of material be in % of the default value.  And before everyone comes down on me for basing that on results that have considerable statistical randomness and such, I run my tests 900+ games or until it has messed up so bad an Experiment can't possibly recover to catch the leading Experiments.  I did have one wild statistical anomaly.  After 250 games Exp 21 was only loosing 4% but after another 250 it was worse than Default.  Imagine you are a chess professional and you join the elite for a few years then play lousy for just as long but actually your strength never changed at all...that would be hard to consider.  I think suicide would be higher on the consideration list ;)

Method: 60/1, 60/1, then 1 sec inc. Gauntlet consisting of 9 engines from different authors averaging over 3100 elo.  Except that I included one R3 clone. Perfect 13 book. 3-4-5-man endgame tables on camera SD card using a card reader.  100 rounds.  No head to head as I am looking for a rating increase that would be reflected on a typical ratings site. Hardware: Q6600 @3465.  Chip has been too abused to go 3600 anymore ;)  XP 64-bit OS.

After over 16,000 games the conclusions seem to be:

1. None of the parameter settings given in this forum or another, other than Human are as good as the Default setting.
2. Any deviation to default Rook Endgame Scaling is destructive.  I ran my experiment on RES against StockFish 1.8 because I was seeing a high frequency of rook endgames with it.
3. The 100 150 timings seem to be better for most attempts but not for the Default settings for some reason. (I have only tested three timings settings...that aspect is next on my list to investigate)
4. Contempt seems to be effective in increasing performance much more than the settings given in the forums.
5. It appears that the engine can be optimized in different directions: low draw, high wins, high points, or low loss.

For reference, Default achieved 72.52% of game points, 6.76% loss, 51.80% win and 41.4% draw.

Experiment 15 (not to be confused with Beta 15) achieved 74.22%, 6.22%, 54.67%, and 39.11%
Experiment 18 achieved 74.69%, 9.41%, 58.79%, and 31.81%

The bold values are the best values by any engine tested so far.

Exp 15
--------
3
100
150
1
2
23
25
23
25
23
25
14
10
4
2
100

Exp 18
--------
3
100
150
4
3
10
8
10
8
10
8
17
13
32
28
100

For some reason I did not think the games of either were particularly entertaining but Experiment 7 was.  It is only a smidgen better than Default: 73.07, 6.79, 52.93, 40.28
Here is 7 for fun:

Exp 7
-------
3
100
150
3
4
8
10
8
10
8
10
17
13
32
28
100

Exp 8 was also fun to watch

Exp 8
-------
3
100
150
6
8
8
10
8
10
8
10
17
13
32
28
100

You can take or leave the results.  They are only based on 900 games per engine.  Perhaps another 900 for Exp 15, Exp 18, and Default would be more conclusive but I thought other testers might enjoy trying some interesting settings now.
Parent - - By Uly (Gold) [mx] Date 2010-07-24 02:29

> I suggest that the settings of material be in % of the default value.


There's consensus that % would not work (one point is that the user doesn't know what's the actual value from where the % comes from), also, GUIs don't allow to have periods on the parameters, so for inserting a 99.996% piece value you'd need to enter 99996 which would be a very confusing value. However, Vas said Rybka 4 Update will allow Millipawns offsetting (10 times as much freedom as now).

>I have only tested three timings settings...that aspect is next on my list to investigate


This is the setting that has the most potential from all of them. At least 30 elo can be won over defaults depending on conditions.
Parent - By mindbreaker (****) [us] Date 2010-07-24 02:50
Millipawns would be fine.

I just have not gotten to the timing aspect because it is strongly tied to the game time control.  I was going for the optimum in the material as that should be less subject to the specific circumstances of my time control, my computer, etc.
Parent - - By Labyrinth (****) [us] Date 2010-07-24 12:18

>one point is that the user doesn't know what's the actual value from where the % comes from


This one is a deal breaker, but I think the rest could be done though. Like you said it would be confusing but it could be made to internally decipher 99996 to 99.996%.

Of course both could not be supported right? How would it know you were entering a percentage vs. a millipawn value? The gui could not be expected to understand a % sign I don't think. Maybe percentages could be prefixed or suffixed somehow? 9999600?
Parent - - By Uly (Gold) [mx] Date 2010-07-24 22:20
What benefit would percentages have over millipawns?
Parent - By Labyrinth (****) [us] Date 2010-07-26 09:43
Well I have no idea, just found the puzzle of implementation interesting.
Up Topic Rybka Support & Discussion / Rybka Discussion / Paramiters Sugestion

Powered by mwForum 2.27.4 © 1999-2012 Markus Wichitill