All my ideas lately to improve the playing strength of my chess engine did not work out. Sometimes it looked promising but something that seemed like a small progress vanished if more games were played.
I have not the resources to run hundreds of thousand of games. Playing 1000 games is already quite a commitment and the error bar after 1000 games is still very big. So my latest version v2589 shows something like a small improvement but actually it can even be worse than the last one in the list. They are all very close together and I kept the code changes not for a proven ELO gain but because they fixed some small bug or simplified the code a bit.
Rank Name Elo + - games score oppo. draws
1 gaviota v0.85.1 199 35 35 368 78% -20 17%
2 cheng3 1.07 JA 52 14 14 1860 60% -21 17%
3 Daydreamer 1.75 JA 30 12 12 2358 57% -19 29%
4 iCE 0.3 v2589 -20 16 16 1295 52% -31 37%
5 iCE 0.3 v2459 -24 20 20 766 50% -25 36%
6 iCE 0.3 v2441 -25 9 9 3768 47% -1 33%
7 iCE 0.3 v2489 -25 18 18 1014 50% -25 32%
8 iCE 0.3 v2457 -25 8 8 5160 50% -22 33%
9 iCE 0.3 v2559 -27 10 10 3493 49% -17 31%
10 iCE 0.3 v2574 -31 9 9 4162 48% -16 33%
11 iCE 0.3 v2413 -34 12 12 2161 49% -28 33%
12 iCE 0.3 v2571 -35 11 11 2768 48% -18 32%
13 iCE 0.3 v2437 -35 9 9 4249 47% -15 31%
So I'm still looking for something that adds some ELO. But it seems most of the low hanging fruits have been picked.
No comments:
Post a Comment