Discussion of anything and everything relating to chess playing software and machines.
Moderator: Ras
Leto
Posts: 2152Joined: Thu May 04, 2006 3:40 amLocation: Dune
Post by Leto Fri Apr 22, 2016 6:02 pm
Laskos wrote: hgm wrote: Indeed, contempt should just reflect how likely it is your opponent plays poorer than you, (e.g. as Elo difference), and any effect of position score and game phase on the desirability of a draw in this light should be automatically calculated. It would only make sense to adjust it during the game when you know your opponent is (say) especially good in the end-game and below-average for his rating in the middle-game.
Yes, without some special knowledge about the opponent, this dynamic part of Contempt should be in eval. Stockfish seems to behave normally with its Contempt. Against Komodo the optimal Contempt is close to 0 (they are close in general strength, and there is no special knowldge about particular strengths), independently of the opening advantage:
Code: Select all
SF7 Opening Advantage: 120-160cp
Contempt=-30
+361 =108 -31 83.0%
Contempt=0
+379 =87 -34 85.5%
Contempt=30
+379 =79 -42 83.7%
SF7 Opening Disadvantage: 120-160cp
Contempt=-30
+56 =118 -326 23.0%
Contempt=0
+74 =109 -317 25.7%
Contempt=30
+42 =135 -323 21.9%
Good find Kai
Nordlandia
Posts: 2845Joined: Fri Sep 25, 2015 9:38 pmLocation: Sortland, Norway
Post by Nordlandia Fri Apr 22, 2016 7:51 pm
White: 8-Core i7-5960X
lkaufman
Posts: 6300Joined: Sun Jan 10, 2010 6:15 amLocation: Maryland USAFull name: Larry Kaufman
Post by lkaufman Fri Apr 22, 2016 8:15 pm
Were the unbalanced positions you used materially unbalanced (a pawn or more), or just positionally unbalanced?
Komodo rules!
Nordlandia
Posts: 2845Joined: Fri Sep 25, 2015 9:38 pmLocation: Sortland, Norway
Post by Nordlandia Fri Apr 22, 2016 9:59 pm
lkaufman wrote: Were the unbalanced positions you used materially unbalanced (a pawn or more), or just positionally unbalanced?
[d]rnbqkbnr/pppppppp/8/8/8/8/1PPPPPP1/RNBQKBNR w KQ - 0 1
Two rooks pawns for no castling rights for black.
I used contempt of +20 and +25 respectively.
I can reverse the colors, e.g Houdini play black (8-Core), - Komodo on (2-core) to see how it turns out.
Last edited by Nordlandia on Fri Apr 22, 2016 10:03 pm, edited 2 times in total.
Laskos
Posts: 10948Joined: Wed Jul 26, 2006 10:21 pmFull name: Kai Laskos
Post by Laskos Fri Apr 22, 2016 10:00 pm
lkaufman wrote: Were the unbalanced positions you used materially unbalanced (a pawn or more), or just positionally unbalanced?
Some 1500 positions Stockfish evaluated as 120-160cp. The majority are material or obvious captures, but there are some just positionally unbalanced too.
mjlef
Posts: 1494Joined: Thu Mar 30, 2006 2:08 pm
Post by mjlef Fri Apr 22, 2016 10:57 pm
Laskos wrote: lkaufman wrote: Were the unbalanced positions you used materially unbalanced (a pawn or more), or just positionally unbalanced?
Some 1500 positions Stockfish evaluated as 120-160cp. The majority are material or obvious captures, but there are some just positionally unbalanced too.
We are trying to reproduce your results. Were the positions selected via a static eval, or a full search? If search, how deep?
The hard part of using the best score form the root, and changing the Contempt, is that each time you do it you would have to clear the various evaluation hashes and main Hash table, which is going to kill search efficiency. But it could be worth it maybe once during the game. If the positions were selected with a static eval, then this does not matter.
Mark
Laskos
Posts: 10948Joined: Wed Jul 26, 2006 10:21 pmFull name: Kai Laskos
Post by Laskos Sat Apr 23, 2016 12:25 am
mjlef wrote: Laskos wrote: lkaufman wrote: Were the unbalanced positions you used materially unbalanced (a pawn or more), or just positionally unbalanced?
Some 1500 positions Stockfish evaluated as 120-160cp. The majority are material or obvious captures, but there are some just positionally unbalanced too.
We are trying to reproduce your results. Were the positions selected via a static eval, or a full search? If search, how deep?
The hard part of using the best score form the root, and changing the Contempt, is that each time you do it you would have to clear the various evaluation hashes and main Hash table, which is going to kill search efficiency. But it could be worth it maybe once during the game. If the positions were selected with a static eval, then this does not matter.
Mark
Search 100ms of Stockfish. Depths were 10-12 or so. Also, positions were early middlegame. I built these kinds if opening suites about half a year ago for other purposes, they came in handy to test what Larry said.
Nordlandia
Posts: 2845Joined: Fri Sep 25, 2015 9:38 pmLocation: Sortland, Norway
Post by Nordlandia Sat Apr 23, 2016 2:06 am
White: Komodo 9.42 | i5-5200U [2-Core]
Laskos
Posts: 10948Joined: Wed Jul 26, 2006 10:21 pmFull name: Kai Laskos
Post by Laskos Wed Apr 27, 2016 8:11 pm
lkaufman wrote:
Based on Kai's tests, a value of about 40% of the score, with the sign reversed (negative contempt when ahead) should help, which is roughly what I use in the handicap matches.
I tested for a simulation of your handicap matches: Komodo much stronger and in significant opening disadvantage.
Code: Select all
Komodo 12s+0.12s vs Stockfish 2s+0.02s
Disadvantage 200-240cp
Contempt=0
+282 =267 -451 41.6%
Contempt=+30
+318 =241 -441 43.9%
Contempt=+60
+343 =225 -432 45.6%
Contempt=+90
+355 =222 -423 46.6% +35 ELO points
Contempt=+120
+336 =207 -457 44.0%
The Contempt seems to improve chances of win by 26%.
lkaufman
Posts: 6300Joined: Sun Jan 10, 2010 6:15 amLocation: Maryland USAFull name: Larry Kaufman
Post by lkaufman Thu Apr 28, 2016 4:28 am
We did make a change to Contempt (for Komodo 10) based on your findings, but it is a small change and appears to be worth just one or two elo. But we'll take it, thanks!
Komodo rules!
