SF 29.07. improvement, remis parameter could be not good?

[Frank Quisinsky]

Dear Stockfish programmers,

since 2 years (after a longer break) I testing a bit SF 29.07.2013 x64 8 Cores. First opponent was Spark 1.0 x64 8 Cores. Games with ponder = off.

Result was 8,5 : 1.5 for SF.
Interesting are two of the three remis games.

Both games ended very fast (under 30 moves) with remis.

Situation:
SF had a little Advantage, around 0,3 - 0,4 Points. Lost this Advantage in the following 3-4 moves up to 0,0 and the games ended with remis after SF is willing to make scrifices moves ended in 3 moves rule.

Same Situation I saw in SWCR against weaker engines.

In my opinion the remis code should be a lot harder, because SF should win against weaker engine the late middlegame / endgame. No reason to give games to fast remis. I think this one can give an Advantage from around 5-10 Points. But I am not sure, because I don't tested this one in detail.

Against stronger opponents very hard to see.

Perhaps a good idea to improved SF. If I remember I had spoken about it in the time I test the PHQ Parameter for SF 2.1.1. SF 2.1.1 had the same Problem I think.

By the way ...
What a great developement in the last time.
Again and again, have thanks for this great piece of chess Software.

Best
Frank

[Frank Quisinsky]

The other PHQ tester gave me in CSS forum the hint, that SF have a contempt Parameter. I think this one is good but perhaps an other way could be a simple programming ...

Attentione:
Frank a non programmer is now speaking ... sorry to all the great programmers.

Perhaps a good idea could be ...
If the ponder hits after the first 15 moves the engine calculated (after opening book) is high ... then the contempt factor should be lower.
= SF is playing against a stronger opponent

If the ponder hits lower after the first 15 moves the contempt factor should be higher.

Engine can calculated after the first 15 own moves, 20 own moves, 25 own moves and so one a new contempt factor self one. Perhaps better as a fixed Parameter, I don't now.

Best
Frank

[Frank Quisinsky]

Hi there,

one idea in complement.

If such an intelligent code is available (engine find out how strong is the opponent with the own eval or ponder moves during the still running game, a lot of other ideas are possible) the result of this own learning during the game can be give in a text file.

So with an intelligent code engine can give the Information:

Opponent are playing with:
- around the same Level.
- 100 ELO weaker
- 200 ELO weaker
- 100 ELO stronger
- 200 ELO stronger

Ende of a test with perhaps 100 games I can look in the file and can see very easy in 90% the intelligent contempt factor is good or wrong. So with a good additional code to "contempt" it's easy to improved a bit I think.

Best
Frank