Predict AlphaGo on Future of Go Summit in Wuzhen, China

[pookpooi]

Let's heard from what Lee Sedol, last year AlphaGo's opponent, said today after the match ended.
"I think Ke made some good unexpected moves in his game, but overall it was a complete loss," Lee said in a phone interview with Yonhap News Agency. "In the end, the game turned out as I expected."
"He just finished the first game with black stones," he said. "Since he will play the second game with white stones, he will do better."
"When AlphaGo gets into a good rhythm, it's really difficult to beat," he said. "He should make some moves in the early stages of the game."
"I noticed that Ke Jie had a lot of time," he said. "He should have used his time in the early part of the game."
http://english.yonhapnews.co.kr/news/20 ... 00315.html

[Baywa]

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.

I know the reasoning. But still, experienced players would still play differently with only little or no risk, because they know about eyes and liberties and don't have to "roll it out" every time.

Still, I'm just curious and my question still stands!

P.S. It seems it has not been the case yet, where a game AlphaGo vs. Human was very close at the start of the endgame. OTOH, that must have happened during AlphaGo vs. AlphaGo games. Would be interesting to see how the program handles that case when there is no safe margin.

[djhbrown]

[my question still stands!...safe margin.

in general, my philosophy is "if at first you don't succeed, give up" but being what the French call "tetu" i rarely adhere to it, so here goes another try (but this time, following in the footsteps of Ke Jie, it will be my final attempt): Alphago doesn't know what you are talking about when you use the phrase "safe margin" because that is not in her vocabulary. All she knows about is probability of >= +0.5

This does raise an interesting question which the mathematicians among you may (or may not, as the case may be) care to ponder: which of you is right and which is wrong?

Given that Alphago is better even than Ke Jie, the evidence would seem to favour her opinion.

However, personally, i don't concur, for reasons i have given previously.

Imagine you are lying on the operating table, and the surgeon says "i think there is a 95% probability that if i cut here it will come out all right"...

[Uberdude]

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....

Was the part I underlined not a joke? (and also not some philosophical gotcha on the ability of a program to "know")

[djhbrown]

she knows nothing about eyes...Was the part I underlined not a joke?

correct.

(and also not some philosophical gotcha on the ability of a program to "know")

correct

[Uberdude]

From page 11 of the AlphAgo Nature paper listing input feature to the neural networks:

Code: Select all

Feature              # of planes          Description
Sensibleness         1                    Whether a move is legal and does not fill its own eyes

This is presumably also why AlphaGo doesn't fill in its own eyes/territory at the end in Chinese rules if it is still winning, as some other Monte-Carlo based bots do (and look silly).

[Bill Spight]

Imagine you are lying on the operating table, and the surgeon says "i think there is a 95% probability that if i cut here it will come out all right"...

Well, that's why they put you to sleep.

[djhbrown]

From page 11 of the AlphAgo Nature paper listing input feature to the neural networks:

Code: Select all

Feature              # of planes          Description
Sensibleness         1                    Whether a move [b]is legal and does not fill its own eyes[/b]

this is an example of poetic licence - to be sure, checking that a move is legal is done. that, however, does not require a software expression of the general concept of an eye, which as a 4-dan player yourself, you will know requires a recursive procedure. Alphago manages very comfortably just checking that moves do not create self-capture, because her rollouts ensure that she doesn't put herself in the kind of pickle you are in right now.

If you wish to press this point, or any other point concerning Alphago, please take it up with the authors of the paper, since i do not speak for them.

One of the side-effects of a rich form of expression such as natural language is its inherent ambiguity. Which is why Lewis Carrol wrote: "When I use a word," said Humpty-Dumpty in a scornful tone, "I use it to mean what I want it to mean. Neither more nor less"

[dhu163]

there are mutterings about move 54. it does look very strange a play to cut in gote. Ke Jie said no human would play there, but that post game review showed it was a good move. Supposedly alphago's own winning percentage dropped from 58.3% to 56.3% after it chose it compared to the direct and natural attack at L3.

My source is https://www.youtube.com/watch?v=1_QXFJl0UJE listing which moves caused 2% shifts in alphago's winning percentage.

Next move 69 etc was a big mistake, the percentage going to 62.8%.
And Move 97, called an attempt to mimic alphago's style my playing in the centre (too late though) lost ke jie 4%. Though after move 99, alphago lost a little confidence. But after finding 110, it recovered. And it agreed with Redmond saying 111 should be G12.

at 118, B's winning percentage was 32%


Since I can't find any other source for this, I'm not sure what to think. He also mentioned alphago was leading by 18 points at one point.

[Kirby]

Supposedly alphago's own winning percentage dropped from 58.3% to 56.3% after it chose it compared to the direct and natural attack at L3.

I understand that AlphaGo's winning percentages may change based on what the opponent does (e.g. if the opponent plays an unexpected move).

But, why does the winning percentage change based on a decision that is made by AlphaGo (assuming Ke Jie's previous move was one that was somewhat expected)?

Or was Ke Jie's previous move less expected?

[hyperpape]

Supposedly alphago's own winning percentage dropped from 58.3% to 56.3% after it chose it compared to the direct and natural attack at L3.

I understand that AlphaGo's winning percentages may change based on what the opponent does (e.g. if the opponent plays an unexpected move).

But, why does the winning percentage change based on a decision that is made by AlphaGo (assuming Ke Jie's previous move was one that was somewhat expected)?

Or was Ke Jie's previous move less expected?

I'm not sure what the comparison is (between AlphaGo's evaluations after moves 52 & 54 or between its evaluations after moves 53 & 54), but it makes sense to me either way.

You're deciding between three moves, and you read a bit for each of them. You pick the one that looks best. Now it's your opponent's turn, and you know what your last move was, so you can ignore two of the three. You read a bit deeper, and see the problem you missed before.

There's a related problem for AIs called the horizon effect, though I'm not sure to what extent it affects current go programs. If the machine can see a potential problem in one line of play, but can play a stupid move that pushes the problem down the road, it will often choose the stupid play, even if it ultimately makes things worse.

[Kirby]

I suppose that makes sense, especially if you compare the knowledge known now to the knowledge known 20 moves ahead. A 2% difference in assessment with just two additional moves on the board seems pretty big, but I suppose this is relative.

Somehow, I thought that AlphaGo's probability assessment was based on inputs fed to the function trained through its value network - I didn't realize where reading came into play here.

I've never fully read the nature paper, though, so maybe I should get a better idea of how it works. :-p

[zermelo]

How would the authors of that video have access to Alphago's internal winning probabilities? I don't understand the language but I'd guess they are using some other engine, maybe Crazystone that's mentioned in the video title. Then it makes perfect sense that any move changes winning probabilities.

[hyperpape]

Somehow, I thought that AlphaGo's probability assessment was based on inputs fed to the function trained through its value network - I didn't realize where reading came into play here.

I don't remember for sure either. But if the value network isn't perfect, it can still have the same effect: a position looks good until you see the reply.

How would the authors of that video have access to Alphago's internal winning probabilities? I don't understand the language but I'd guess they are using some other engine, maybe Crazystone that's mentioned in the video title. Then it makes perfect sense that any move changes winning probabilities.

That would also make sense and make my story unneeded.

[dhu163]

I think he said it was official alphago data, but I haven't found anything to corroborate that ...

I also wondered if it was some other bot, but he only said it was what alphago was thinking at the time ...