Hello,

in this month, computer go has made several big steps:

* the wins of Zen on 19x19 against Takemiya Masaki (9p)
at 5 and 4 handicap stones

* the win of Pachi2 in the KGS bot tournament at slow time controlls

* the impressive performance of Zen13 in many free games on KGS

******************************************************

Sometimes I like to ask provocative questions.
Here come a few. What do you think?

* When will be the first time that a human pro takes handicap from a
bot on 19x19? (Giving the bot White and komi=0.5 is counting as handicap.)

* When will be the first time that a human pro loses a game against a
bot on 19x19, when the human took handicap?

* When will be the first time that one of the strongest humans takes
handicap from a bot on 19x19?

* When will the first time that one of the top human player loses against
a bot on 19x19, when the human took handicap?

* Will it ever happen that a top human player loses against a bot in a
19x19-game where the human got 4 handicap stones?

I know that there are other hard steps which have to be mastered before
by the programmers. But nevertheless...

Ingo.

As I expressed elsewhere, I have trouble understanding how MCTS could play at the level of top professionals. What's your take on that?

_________________
Occupy Babel!

I guess it's still ~5 years ahead, at least6-7 years ahead. As soon as this is needed to keep even, a pro-slip can happenReal handicap, like black with two stones? Maybe in 10 yearsProbably neverNo. I think.

_________________
Geek of all trades, master of none: the motto for my blog mostlymaths.net

I wonder if man-vs-machine matches will disappear when go bots start getting better than top humans, the way they have with chess. I don't know much about chess but my understanding is that current chess engines would completely crush top pros in a straight-up match. They tend to win even win with one hand tied behind their back (e.g. no endgame or opening tables) while giving heavy odds like pawn and a move. Although elo ratings from automated computer-only matches are not directly comparable to human elo ratings, I think that it's fair to say that the chess machines are leagues stronger than the best humans.

It's been suggested that top pros would need to take three or four stones to play God, but given the experience from chess, I wonder if that estimate might be low.

_________________
KGS 4 kyu - Game Archive
Keyboard Otaku - Webmaster Fort Myers Go Club

This is very true, but it hasn't stopped chess engines being developed - it still seems as active as it's ever been in the last 10 years.

_________________
My Go videos | My Go Teaching Ladder (GTL) reviews
I also offer Go teaching lessons
-- Creator of OGS --
"I played a trick play and my opponent fell for it, unfortunately he didn't know the correct followup, so I died"

Top programs will reach 9D in KGS with normal or slow time setting before 2020.

Two and also still one year ago several MCTS programmers
had expressed their opinions that MCTS would run against
some threshold rather soon. I - simply - believed them.

Now I am unsure what to expect.
The heavens for computer go seem to be open.

Ingo.

Hello Ruben,

nice to meet you again. We had played in a little 9x9
go tournament 3.5 years ago on Little Golem (and you
beat me).




That will be indeed another interesting question.
Will top humans "simply" avoid playing bots when
bots become too strong.

Ingo.

That happened in a few experiments with chess bot Rybka, but
human chess masters might learn to do better. Unfortunately,
there is no real culture of handicap game in chess.



Indeed, (future) matches where go bots give handicaps to top humans
may provide lower bounds for the number of handicap stones God might
give to humans.

Ingo.

Hello pookpooi,



A question out of curiosity:
Is KGS 9D assumed to be pro level?

Ingo.

No.

as you said, there are some lesser challenges to be beaten before the time for your questions comes.

when will be the first time that a human pro plays an even game against bot on 19x19? and when will the bot win such a game for the first time?

i would put both answers close to each other, let say around a year apart, and i estimate it might happen in 10 years from now. then a human receiving handicap could follow again relatively shortly afterwards, like 2 years, but i can't really imagine going up to your suggested 4 stones

_________________
Spilling gasoline feels good.

I might be wrong, but probably not.

"Is KGS 9d a rough ruler for pro strength?": http://www.lifein19x19.com/forum/viewto ... =10&t=4398. The short answer is that it's not quite reliable, but it's very very close.

_________________
Occupy Babel!

For programs to play at solid pro strength (say 5p or higher) I think a new approach will be necessary. The reason is that even the Monte Carlo based programs don't really "understand" go. A good position evaluation function still has not been developed. The chess programs with opening and endgame dictionaries built in take advantage of the relatively small size of the chess universe so they can play well with a somewhat crude position evaluation function.

It is vaguely interesting that Takemiya lost those handicap games to Zen but I'd like to see what would happen over a longer, multigame match. Personally I am not very interested in how strong programs are or could become. For me there is some pleasure in regarding go as a sequence of puzzles to be solved, which is how playing against a program feels, but playing against a human partner is much more. The old nickname for go, "shudan" (hand talk), sums it up; a game is a conversation and a mutual appreciation of the marvels of the game. This mutuality is simply impossible when my "opponent" is a computer program. I might be interested in using a pro strength program as a training tool as I think some chess pros do with strong chess programs. But a good "teaching" program would have to understand how humans can think about the game to explain mistakes in a useful way. Rather than simply saying move x is better than move y it would have to explain why move x is better, not just say the probability of winning is higher with x than with y.

People have moaned and groaned about such issues as defending the human spirit against the machine and held up go as a game that will never be "mastered" by programs but this is actually irrelevant. No one complains that robot driven automobiles have demeaned the human spirit because they can beat human runners in 400 meter races. Maybe the best results for go will come from "hybridization" of human brains and computers, hardwiring computers into brains to provide assistance to the brain, as I expect we will see before too long with computer assistance to overcome sensory defecits or physical impairments such as paraplegia.

Why do you believe that computers need to understand go to play it any better?
Computers are not humans. They don't need to understand go, nor to rationalize about moves as we do. They only need to pick the best moves, that's all.

Yes and no. It's true that a program that understands "this position is thick" is the best thing that's imaginable. But a program that can find points where your move leads to a sharp drop in evaluation is quite helpful even without that possibility. Similarly, if you think a particular line refutes a move, it can show you why it doesn't.

I can also imagine a MCTS program where you say "what about aji here?" and it finds lines where the opponent attacks and shows you refutations. Doing that well would be almost as much a problem of UI as programming. As time goes on, expect more of that sort of thing to be developed, but right now, the race for strength is so pressing that it'll take time for those ideas to be implemented.

_________________
Occupy Babel!