Friday
Oct192018
The Final Four, TCEC-Style
Friday, October 19, 2018 at 2:47PM
This is not the main TCEC event, but I like the format of the TCEC Cup, a 30' + 10" knockout event that's down the final four. Very unsurprisingly, the semi-finalists are Stockfish, Lc0 (Leela), Houdini, and Komodo. The first semi is underway, featuring the first pair.
All the matches in the event are best-of-eight, with pairs of tiebreak games played if necessary. Stockfish has demolished its opponents so far while Lc0 has squeaked through its matches (scroll down the page linked above to see the brackets), but sooner or later one assumes Lc0 will devour all the conventional programs. Has that time come yet? Probably not, but we'll see.
Reader Comments (2)
It would be fun to be able to play like Stockfish did in game 5.1 against Lc0:
Stockfish 270918 (3519) - Lc0 18.11248 (3247), 2018.10.20 Sicilian, Alapin's variation (2.c3)
1. e4 c5 2. c3 Qa5 3. Nf3 e6 4. Bd3 d5 5. exd5 exd5 6. O-O Nc6 7. Bc2 Bd6 8. d4 c4 9. Re1+ Nge7 10. Ng5 h6 11. Qh5 O-O 12. Nh7 Re8 13. Nf6+ gxf6 14. Bh7+ Kxh7 15. Qxf7+ Kh8 16. Qxf6+ Kh7 17. Qxh6+ Kg8 18. Qg5+ Kf7 19. Qh5+ Ng6 20. Qh7+ Kf6 21. Rxe8 Qc7 22. Qh6 Kf7 23. Re3 Bf5 24. Rf3 Nce7 25. Qh7+ Kf6 26. h4 Bh2+ 27. Kh1 Bf4 28. Bxf4 Nxf4 29. Qh6+ 1-0
[DM: That was a fun game! A very easy one though; I used to win games like that in blitz pretty routinely. (I can supply references to prove it, if necessary!) 11...0-0 is just insane - very few masters would play that in anything slower than a blitz game, and even in blitz they'd be pretty nervous about it. Lc0 played that game with a glass jaw.]
[DM: 11...0-0 is just insane.]
I looked and it appears Lc0 computed 1.3 million "nodes" to play 11...0-0, which feels like more than enough computation for it to "know better" if it really had some sort of chess understanding. I'm not convinced there's any evidence that these NN engines play more positionally than traditional engines. It may just be humans' cognitive bias that makes them claim that.