Hacker News Reader

Comment by loup-vaillant | Hacker News Reader

Some commenters on Twitter are still saying 0.5 points is "very close".

It would be if both players were human: in human play, score differences tend to correlate with differences in actual skill, and probability of outcome (who wins the game).

Not so with Alpha go. That machine just takes the surest path to victory, with no regards to its magnitude. It doesn't care about winning by only half a point. It cares about securing at least half a point.

It may have been a crushing victory for all we know.

View on HN

It was a crushing victory actually. I watched the entire game, and followed professional commentary. Before the end game started, Alphago was leading by between 10 to 15 points -- an enormous lead in professional games. Most players would have simply conceded, which is not to say this is what Kejie should have done. Actually Bravo to him for sticking it to the very end, and let us all watch how a computer would handle end games. As it turned out, Alphago routinely picked the marginally safer move while yielding a bit of its lead. A style of playing that's not consistent with human players.

It would be interesting to see how Alpha Go's performance varies with different goals, balancing between maximizing score and maximizing probability of winning.

I have to think they've discussed that internally, and they probably just want to make sure alphago can win consistently, period, before they start playing around with allowing slightly riskier moves as long as the win probability stays sufficiently above 50%. (But how much more? You can't know ahead of time how much stronger you are than the other player that day, or you wouldn't be playing, at least not for money.)

It's just like adjusting komi to give the human an advantage, right?

An indirect way to do something similar-ish to what I was interested in would be to play with varying numbers of handicap stones with the current goal unchanged (maximizing probability of victory).

Actually, this mostly works. http://pasky.or.cz/go/dynkomi.pdf has details how to do so.

Didn't they deny that assumption last time?

Not that I recall, but "close victories" were mentioned as well. Moreover, professional commenters didn't know why Alpha Go won: they qualified some of its moves "poor" and were dumbfounded as to how it won anyway. (Alpha go did lose one game, but that was because it didn't manage its time properly: it used the same amount of time for every move, even in high-uncertainty situations.)

I don't see them making the same mistake again though. By now the machine is most definitely superhuman. Its moves will be studied, and this will likely improve human play as well.

Reply on news.ycombinator.com