If we can forget about the whole AlphaGo event, this will be a very big leap for computer go.... no, even with AlphaGo's existence, it's still impressive, consider that Zen run on just 12 cores mac book pro (account: Zen19A), while CrazyStone run on LG Nexus 5 (account: CrazyPhone). Both bots are playing with 15 seconds of 10 byoyomi periods time setting though. And I think we'll get the update of CrazyStone desktop version soon that may fix its bug in 7 dan level. (account: CS2016b)

Just like chess apps, some day your watch may give top pros 4 stones.

Oh man that would be so good.

Which bug is it? The one where it can't read out ladders?

Pretty impressive, just 6 months after Zen got to 7d. One caveat is that Zen19A started rejecting games with 5 or more handicap (it resigns the game right away, so it doesn't count).

It's currently playing against kingmaker 9d (KGS #2)!

Edit: And won by 7.5 points (no komi)! Wowzers. kingmaker killed a corner with a non-alive outside group and Zen got a lot of profit from hassling it. I think this is a common weak point of humans, we get angry and feel "Hey, you should have lived with your group in gote but you didn't, so I will kill it you naughty bot!". It's almost a trick of bots that exploits human greed. With more experience perhaps kingmaker will do better.

and won, even though it's not an even game (komi 0.5), still, it has beat many 8d players in an even game.
Also, Aja Huang is like a celebrity when he appear in computer go room, one thing he did say is that AlphaGo is still improving and is much stronger now, and I was like, oh my godddd...

Zen won two of three games against Kingmaker 9d

http://kgs.gosquares.net/index.rhtml.en ... 6&d=20&r=0

_________________
“The only difference between me and a madman is that I’m not mad.” — Salvador Dali ★ Play a slooooow correspondence game with me on OGS?

Maybe later on KGS will have to adjust the ranking system so that there can still be some slots left reserved for humans to be ranked as 9d.

_________________
http://tchan001.wordpress.com
A blog on Asian go books, go sightings, and interesting tidbits
Go is such a beautiful game.

I'd like to see a Haylee vs Zen match!

Zen is now 4-1 against kingmaker 9d (though no komi games) and 4-2 against cloudnotes (strong 8d, used to be 9d). It seems to do less well with handicap against weaker players, and there is a guest (TACO2, apparently a Tygem 9d) with a winning record in even games. It seems Zen would be 9d if it played even games, the handicap games might hold its rank bank (plus the inertia of KGS ranks).

That might very well be that the new version hasn't yet had the fine tuning adjustments for letting the algorithm play a handicap game in the first place. The mechanisms by which these programs are able to play handicap games aren't as straight forward as you might think.

For example, the "straight" MCTS programs might use something like "dynamic komi" << in deciding who has won a playout, a temporary reverse komi of X amount used >> with that dynamic komi decreasing as the game progresses. But what its initial value should be and the rate of decrease would need to be determined for each number of handicap stones. A lot of trial and error and that takes time and data (games played at that handicap).

Hideki said that "It's well known that MC bots are not good at (high) handicap games." True to his words, even now Zen begins to has problem with 4 stones handicap games.

So komi is 0.5... if Kingmaker had played with komi at 6.5, he'd have 6 points more; Zen would still have won by 1.5 points.

I think that the internet has to rewrite the phrase: "Computers can't play go well. The best programs only play between 5 and 10 kyu. Most talented amateurs can beat them after a few months of practicing."

Haven't these sentences rewrote ten years ago, since the implementation of monte carlo tree search in 2006?

On some sites yes, but there are still many that state that computers can't play Go above mid-range amateur level

That might be an easy to misunderstand over generalization. INITIALLY (without some parameters being set) the MCTS algorithm can barely play against a handicap at all. I just described ONE approach that enables them to do so, "dynamic (changing) reverse komi" used in deciding who won a playout". But the correct values, initial starting vale of the reverse komi and the rate of decrease would have to be determined (tuned) for each number of handicap stones. And that would have to be redone every time the MCTS program had a major improvement. TAKES TIME (and effort). Even with a team to put in the effort I wouldn't expect to be right for quite a few months after each major improvement. And about the other possible approaches to allow MCTS progrms to play against handicap I know nothing. But I doubt any of them wouldn't also require adjustment involving much trial and error.

In other words, the statement should more correctly be interpreted "much of the time" which during times where the programs are rapidly improving become almost all the time << the new, improved version might be coming into use before the handicap adjustment done for the previous version >>

Well... I just think it's funny that many sites still state that 'you are playing a game at which you can easily beat the very best computer programs in existence.' Even sites by Go associations and publishers still say that. It just makes them look as if they haven't been keeping up for 10 years.

Looking at ranked games with correct handicap for opponent, Zen wins more than it loses by more than ten games.
It does not seem all too affected by handicap.

My ipad server version of Crazy Stone was automatically upgraded to 2016 Deep learning KGS 7d, which will have to make do till I can buy a 12 core Xeon and an English version of Zen 6. As a SDK it should probably be enough

_________________
She generally gave herself very good advice, (though she very seldom followed it).
Lewis Carroll