Predict AlphaGo on Future of Go Summit in Wuzhen, China

[hyperpape]

$$c Ke Jie is Black.
$$ ---------------------------------------
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . 3 . . . . . . . . . . 5 . . . . . |
$$ | . . . , . . . . . , . . . . . , 1 . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . , . . . . . , . . . . . , . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . 4 , . 6 . . . , . . . . . 2 . . . |
$$ | . . . . . . . . . . . . . . . . 7 . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ ---------------------------------------

Click Here To Show Diagram Code

[go]$$c Ke Jie is Black.
$$ ---------------------------------------
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . 3 . . . . . . . . . . 5 . . . . . |
$$ | . . . , . . . . . , . . . . . , 1 . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . , . . . . . , . . . . . , . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . 4 , . 6 . . . , . . . . . 2 . . . |
$$ | . . . . . . . . . . . . . . . . 7 . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ | . . . . . . . . . . . . . . . . . . . |
$$ ---------------------------------------[/go]

And on move five, we seem to have a new opening, at least from the GoGoD database in SmartGo Kifu.

[djhbrown]

3-3 the losing move? either way, Ke Jie demonstrates his commonsense by acknowledging that the dog has already passed him, although his linear extrapolation that she (or, rather, her clones, since DM are moving on to more challenging domains) will get better and better is a bit of an oversimplification.

1/10 of the processing power on a single machine in the cloud - a machine with the equivalent of 200 cpus + i forget how many gpus. tpus are just for dcnn, or can they be used for monte too?

since the dog is now 3 days from dead, retiring undefeated (except by Lee and Swim), the strength of her technology can only be established by one of her clones giving top pros handicaps, so still plenty to wonder about.

over to the speculators... oi give it foive

[Uberdude]

since the dog is now 3 days from dead, retiring undefeated (except by Lee and Swim)

You forgot Leela again, and Hajin Lee, and Kim Myungwan, and a load of Chinese pro commentators, and all the other people (possibly myself included, I would with have atarid from above or below, I can't recall which) who would have answered Lee's wedge correctly. Except they aren't so arrogant as to claim that means they defeated AlphaGo.

[mistakenot]

Final count had AlphaGo winning by 0.5 points after komi, though commentators noted that AlphaGo gave some points away toward the end (possibly to make the final outcome more certain or because it doesn't care about the winning margin) so the game wasn't as close as that number might suggest.

Here's the final board state and the score histogram from Leela (AlphaGo was White, Ke Jie was Black):

[djhbrown]

splutter

he who is without humour shouldn't live in a glasshouse

[Uberdude]

Ah, the claim Swim beat AlphaGo is meant to be humorous nonsense? I must admit my abilities to discriminate humorous nonsense from serious nonsense may fall short of what is required to interpret your posts as intended.

[Uberdude]

[Bonobo]

[..]
[totally on-topic so not hidden like your motives]

olddogsnewtricks.jpg

[..]

I’d appreciate if you’d enlighten me about my “hidden” motives, as all motivation I’m aware of was to post a (maybe silly) joke

[dhu163]

just looking up to move 83 so far, it looks like Ke Jie has played well.

the lower right seems reasonable. the upper left ended up alright despite alphago's brilliant new move. The lower left invasion seemed slightly desperate, but manipulating it towards the cut at 83 seems alright. I suppose the poor aji in B's positions on the lower side and left mean W is ahead, but it still doesn't seem clear.

[Baywa]

@Game#1 of the series: So by taking territory from the beginning this was Ke Jie's rope-a-dope-strategy (as in Muhammad Ali vs. George Foreman) but AlphaGo answered it with the Gloria-Gaynor-strategy (as in Killing me softly). Sorry, couldn't resist!

So, AlphaGo really squandered 6 to 7, maybe 8 points, or even more during the end game? For example, at move 210, W A10 - was that really the biggest move? Would be nice if someone could figure out the best sequence of moves for the weak-at-play among us!
Edit: Or showed us the biggest "safe moves".

[djhbrown]

enlighten me

oh, no, not you too. my post was a joke as well; just a word-play.

_______________
i joke, therefore i am stoned

[djhbrown]

So, AlphaGo really squandered 6 to 7, maybe 8 points, or even more during the end game?

this post is not a joke - no, she didn't squander anything. have a listen to David Silver's answer in the post-game Q&A to a question similar to yours. Like all Monte-carlo players that i know of, Alphago does not play to maximise difference, but to maximise - as she sees it - the probability of winning. Couple that to the fact that she knows nothing about eyes, or territory, or anything that people talk about when they think about the game, and her plays sometimes look bizarre, on "tilt", or weak, or slack. But they are none of these, ever - or, to be more precise, not weak or slack insofar as she can see, which is right up to the very end of the game, although only the ends of the possible futures she examines. the closer the game gets to the end, the less likely it is that she will overlook a way she could lose, but the more likely it is that a human player could see a "better" move, insofar as they see it in the terms that they characterise as "good" or "bad". i tried a while back to explain this in my blog http://lcipm.blogspot.com.au/

if there is an opportunity to outsmart a Monte-Carlo player, it will be long before the yose, except for bots such as Leela on a laptop with much less computing power than this Alphago, which is running on the computational power equivalent of 200 latest hi-tech PCs with multiple cores and GPUs.

[pookpooi]

Press conference is almost as exciting as actual match for me, as I can gather more AlphaGo data.
- Single Machine, but with TPU (on Google Cloud). Make it tenth power of Lee Sedol version. If the old single machine (without TPU) then it's only 1/20 power of Fan Hui version.
- Require much less human game record (Zen creator laugh at this, saying 'see! It's still need human!')
- Require much less training time, can be done in a matter of weeks instead of months

Will look forward to AI conference tomorrow as I expected it to reveal more, but with less details than future paper release DeepMind promised.

[Bill Spight]

I am glad that AlphaGo's team continues to make it stronger, to play it against pros, and to have it play matches such as this. It is a wonderful thing.

However, I do have what I hope is a minor nit to pick. From the official web site:

"Fan Hui believes that AlphaGo was telling us its own unique philosophy: "AlphaGo's way is not to make territory here or there, but to place every stone in a position where it will be most useful. This is the true theory of Go: not 'what do I want to build?', but rather 'how can I use every stone to its full potential?'"

Sorry, that philosophy is not unique. It was my philosophy as an SDK. Not that I was able to make full use of my stones, but I tried to do so. And I was hardly unique. What serious go player does not have such a philosophy?

Yes, AlphaGo is great. But let's not demean humans in order to praise it.

[Bill Spight]

if there is an opportunity to outsmart a Monte-Carlo player, it will be long before the yose, except for bots such as Leela on a laptop with much less computing power than this Alphago, which is running on the computational power equivalent of 200 latest hi-tech PCs with multiple cores and GPUs.

But that is not because of the Monte Carlo approach of estimating probabilities by semi-random playouts, but because the bot can read deeply enough not to squander a lead.