New killer idea

[elcabesa]

I implemented your idea in Vajolet, it's now playing a tournament against the old version in very fast games.

those are the results so far:
Score of vajolet2 2.3.5 vs vajolet2 2.3.4: 11 - 10 - 28 [0.510] 49
ELO difference: 7.09 +/- 64.31

Tomorrow I think we will have some more useful data, now it's impossibile to draw any conclusion

the only conclusion we can draw so far is that the patch is not a big ELO LOSS and hasn't really big bug

as soon as the test finished I'll try with longer TC.

[brtzsnr]

At LTC the score climbed to 11Elo (LLR 2.7) and overnight it dropped to 4Elo (LLR 0.8) which doesn't look good. At 4 Elo I will need a lot more games to accurately measure the improvement.

[brtzsnr]

Final status is that it passed LTC, too. It recovered a bit. This patch behaved really strange patch. Went to 11Elo (2.7 LLR) then dropped to 2Elo (0.3 LLR) then increased again to 5Elo (3.5 LLR) before SPRT accepted it!


18675 @ 40/60+0.05
5511 - 5237 - 7927
ELO 5.10±3.78
LLR: 3.49 [-1.87:+3.34]

[brtzsnr]

I ended up adding only the last move predictor which is very similar to the counter move heuristics. In other words, I reinvented the wheel.