Wednesday 29 June 2016

Machine still beating Man

In an attempt to try and even the gap between human players and strong chess engines, chess.com organised a time handicap match between Komodo and GM Sergey Erenburg. If you click on the link you will see that the odds were quite large (90m+30s against 3m+1s, small opening book, Erenburg white in all games) and yet the final results was 3.5-0.5 in favour of Komodo.
For anyone who has developed their own chess engine, this should not come as a surprise. Even as far back as 1995 it was clear that chess engines were able to play quite strongly with very little thinking time. I observed this myself when we used to hold the Carbon v Silicon matches during the ANU Chess Festival. For the early part of the games, the human players would hold their own, but once time ran short, the engines would dominate.
This is because engines can usually get to 7 or 8 ply in under a second, and it is in these first 3 or 4 moves where tactical blunders occur. So engines don't make them, but they do catch them, and this is enough to win. Even my poorly written chess engines (Vanilla chess and bchess) were able to solve over 80% of the positions in '1001 Brilliant Ways to Checkmate' with 1 second thinking time per position.
So the match looks like a kind of 'confirmation' experiment, rather than a new approach to evening the odds. Based on the last few events at chess.com in this area, I still think that odds chess is the way to go, although even then, engines will have to start with less and less material as they get stonger.

1 comment:

Anonymous said...

Basically, a few fans were annoyed at "nonreal" (odds) chess, and wanted to dumb-down Komodo enough that the GM could fight it. Everyone (else) said the human still had little (no) chance, but chess.com agreed to it anyway as a demonstration. He might have drawn an extra game or two, and said afterward that Carlsen might be able to draw with White against 1-core 3-second Komodo nearly all the time, but beating it is completely different.