3 ms·
One question I never found an answer for was what AlphaGo thought komi should be. Since Black goes first, it has an advantage of x stones. But if Black is given
by tunesmith 7y ago
One question I never found an answer for was what AlphaGo thought komi should be. Since Black goes first, it has an advantage of x stones. But if Black is given a handicap of 2 stones (placed on the board before play), white then goes first.
Since AlphaGo plays itself, it could have been set to play itself with a handicap, such that the win rate was 50%. I want to know what that handicap would have been, because wouldn't that indicate a pretty accurate measurement of what komi should be?
- bluecalm 7y agoI agree. There are many interesting questions Deep Mind could have answered about both go and chess. I am still bitter they didn't take a few days more to run experiments of great interest for fans of both games. For chess the simplest one would be to take most popular positions from top level human play and let A0 play Stockfish from that (say 10 times from every of 100 positions). Then do the same for self-play and then most importantly publish all the games not selected sample. They kinda did something similar but the positions they selected were chosen for entertainment value. We still don't know how Alpha Zero plays from positions that are actually interesting for human players.
- monkeypizza 7y agoOpen source research is continuing. There are a few open source implementations of the alphago algorithm and some new work being done to improve it for go. Check out lizzie, which is a nice front end to leela zero. https://lifein19x19.com/viewtopic.php?f=10&t=17095 https://lifein19x19.com/viewtopic.php?f=10&t=17095