AlphaGo vs Lee Sedol

[RaBind]

Google Deepmind's AlphaGo vs Lee Sedol tomorrow morning at 03:30 AM for anyone interested.

https://www.youtube.com/watch?v=vFr3K2DORc8

It'd be nice to use the chat box to discuss the game while it's going on. Any one who knows a thing or two about either the game Go or the AlphaGo system giving a commentary would be interesting.

 

[Ex-User (9086)]

Heh, commentary is much better on kgs or any other service with voip, chatroom and spectating the real-life board to simulate the moves and variation branches.

There will be a few pro players streaming and commentating the game so anything they say is going to be a lot more interesting and in-depth imo.

I'm kind of excited to see these games, despite this, Lee Sedol isn't in his best shape and there are a number of better players (as of now), it's a nice opportunity to bring some popularity to Go.

There's little to no chance they will succeed at their first attempt, since they couldn't have improved so much in such a short time frame, but it's only a matter of time until the computer will master this game.

I highly doubt anyone here is going to sit through a 4 hour long game, I'm going to get some sleep and watch a condensed review later.

 

[onesteptwostep]

He lost :[

I know how to play Go but playing it on a higher level is hard. I'm actually surprised there was a foreign following for Go. I know in Korea they have cable channels dedicated to it but generally no one watches them.

 

[Ex-User (9086)]

He lost by his mistake by komi - 6.5 points (if you count territory prior to resignation). The game was close but black made more mistakes and it's possible alpha go is much stronger but played more lightly because it felt it was winning.

After seeing this game I think it's possible alpha will win every game, or Lee Sedol won't make mistakes in his 2nd game, we'll see.

He lost :[

I know how to play Go but playing it on a higher level is hard. I'm actually surprised there was a foreign following for Go. I know in Korea they have cable channels dedicated to it but generally no one watches them.

There's a growing Go scene in Europe who had their own pro association created 3 years ago. US has had their own Go Association for a very long time.

 

[Ex-User (9086)]

2nd game wasn't even close. AlphaGo won by resignation again, but point advantage was getting over 15 - komi, decisive victory and the endgame was getting better and better for it as well.

This might mean an end to worldwide professional Go on the scale it is today, it'll exist but not as great as it used to be.

Sucks to be one of the aspiring pro players right now, knowing the public won't give them as much attention and they can only dream about sponsorship for what they were doing. Their most valuable skill they've been improving their whole life rendered obsolete.

AlphaGo is a scientific project, so they'll likely soon move on to bigger goals, like medicine, cancer research, etc.

 

[PaulMaster]

AlphaGo is a scientific project, so they'll likely soon move on to bigger goals, like medicine, cancer research, etc. Leaving the entire Go community devastated.

This is likely the ultimate goal. Go is just a warm up. A testing ground. Similar to Watson and Deep Blue.

 

[Haim]

possible alpha go is much stronger but played more lightly because it felt it was winning.

AlphaGo is an AI it just calculates the best possible move, it does not go easy or "felt it was winning".

Due it will take a lot of time for a Go program at that level to work on consumer computer or reasonable sized server, AlphaGo was using 1920 CPUs and 280 GPUs!

 

[Ex-User (9086)]

AlphaGo is an AI it just calculates the best possible move, it does not go easy or "felt it was winning".

Not really. AI algorithms (in chess and other abstract games) may choose safe (less efficient/less aggressive towards the opponent) variations over risky ones when the evaluation function shows they are in the winning position. Also I think saying they "feel" is very accurate because neural learning net function very much like our brains, hence "felt it was winning". I don't think it's accurate to think in terms of the "best move" in a game of Go where AI had to use search narrowing simplifications to choose the most worthwhile variations to ponder. Best move is something when the whole probability space has been sifted and even then particular chains down the road make it questionable to think of individual moves as best disregarding the whole chain of variations.

Due it will take a lot of time for a Go program at that level to work on consumer computer or reasonable sized server, AlphaGo was using 1920 CPUs and 280 GPUs!

Yes, as expected, it will take something in the neighbourhood of 10 years to scale it down. Provided people would still remain interested enough in 19x19 go and won't jump to 3d go or 25x25 or other variants.

 

[Haim]

Not really. AI algorithms (in chess and other abstract games) may choose safe (less efficient/less aggressive towards the opponent) variations over risky ones when the evaluation function shows they are in the winning position. Also I think saying they "feel" is very accurate because neural learning net function very much like our brains, hence "felt it was winning". I don't think it's accurate to think in terms of the "best move" in a game of Go where AI had to use search narrowing simplifications to choose the most worthwhile variations to ponder. Best move is something when the whole probability space has been sifted and even then particular chains down the road make it questionable to think of individual moves as best disregarding the whole chain of variations.

Yes, as expected, it will take something in the neighbourhood of 10 years to scale it down. Provided people would still remain interested enough in 19x19 go and won't jump to 3d go or 25x25 or other variants.

Of course I meant trying to make the best possible move, which is accurate for AlphaGo because it is working using statistical data of better/worse ratio of boards states given a certain move.

First you need to understand that neural networks behave in inspiration from the brain, not like the human brain.
The neural networks in AlphaGo does not have feelings, AlphaGo does can not make it's own goals, it does not have "feel" data.Any part of dog style learning(which is the closest thing to feelings) AlphaGo had was done before the match.Also it would not make sense for the programmers to put "play safe" algorithm(which still wouldn't be feelings but pure algorithm).After some number of moves AlphaGo stops using traditional algorithms and rely upon its NN and Monte Carlo(both don't have "safe" mode)


Other thing I didn't mention is that part of AlphaGo training data was Go matches by Lee Sedol himself, which is like trying to win your lifetime training partner which knows your every move while you lose all your memory of him, literally AlphaGo has neural network that predicts his every move(which then used as an input to other neural network).

 

[Architect]

This might mean an end to worldwide professional Go on the scale it is today, it'll exist but not as great as it used to be.

Years ago I didn't take up chess because a computer had beat a human. When I discovered Go in Japan, as hard as it was to imagine how we'd do it, I knew eventually the computer would win which is why I didn't devote much time into improving. Not to say it isn't worthwhile to play, but I didn't get that much pleasure from just improving.

Not really. AI algorithms (in chess and other abstract games) may choose safe (less efficient/less aggressive towards the opponent) variations over risky ones when the evaluation function shows they are in the winning position. Also I think saying they "feel" is very accurate because neural learning net function very much like our brains, hence "felt it was winning".

Not true. DNN's (deep neural nets) or the old search tree approach used in chess don't pass the Turing test so aren't conscious. They are very narrowly defined intelligence, only doing one thing. Human consciousness has many more 'parameters', using DNN parlance, which (somehow) gives rise to consciousness. Alpha go, with all of it's millions of parameters is still operating at a much lower level.

So no it doesn't feel an answer (intuitively, emotionally or what have you) as a human does. We don't know enough about consciousness yet to define what the difference really is.

So AlphaGo won. Big day, one I've been watching for decades, now we're finally here. Interestingly, just like in the Kasparov chess match, Humans lost because of an emotional reaction, apparently Lee lost his nerve, or gave up or something on the third game.

 

[Haim]

I wonder if AlphaGo could win over 2-4 9 dan go players playing in turn against it.
That way AlphaGo wouldn't be able to predict her opponent's moves and would create super unique playstyle that AlphaGo isn't fitted for, also it would be more fair as AlphaGo use shitlods of cpu and gpu.

It was really interesting to watch the first match, such an interesting board state.

 

[Ex-User (9086)]

Not true. DNN's (deep neural nets) or the old search tree approach used in chess don't pass the Turing test so aren't conscious. They are very narrowly defined intelligence, only doing one thing. Human consciousness has many more 'parameters', using DNN parlance, which (somehow) gives rise to consciousness. Alpha go, with all of it's millions of parameters is still operating at a much lower level.

Thanks for pointing out the obvious. I know it isn't conscious. I said it 'feels' figuratively, I wasn't using the word 'feel' literally.

Although you are correct in clearing up any misconceptions that this program has any intuition or feelings, yes it does compute the board states just like most other algorithms do it in chess except with an excellent database and search narrowing.

So AlphaGo won. Big day, one I've been watching for decades, now we're finally here. Interestingly, just like in the Kasparov chess match, Humans lost because of an emotional reaction, apparently Lee lost his nerve, or gave up or something on the third game.

Yeah, seeing that breakdown was kind of a telling sign that the game won't be as good as the previous ones for the human player. He could have continued the ko fight but maybe his morale wasn't enough to keep him in the game. In the conference after the 2nd game he was 'accused' of playing along and not fighting a ko, which definitely was upsetting to hear and so he at least tried a different approach to fight a ko once at the end.

So in the first 2 games alphago won but Lee Sedol got to keep his territory and it was relatively close to alphago, first game about 6.5 points and second game about 10. So in the last game he gave up completely, even though he could keep on playing and in the best case scenario lose by a closer margin, just like in the previous 2 games.

I wonder if Lee Sedol played the most recent database of alphago, or if they let him play an older version so that the games would be closer than if he played the most advanced version so far. I think there's a case to make that argument, assuming it learns every day and seeing as the games were relatively close in terms of points (10 is crushing in human vs human, but it's not huge in terms of point advantage that is possible on the board).

 

[RaBind]

Other thing I didn't mention is that part of AlphaGo training data was Go matches by Lee Sedol himself, which is like trying to win your lifetime training partner which knows your every move while you lose all your memory of him, literally AlphaGo has neural network that predicts his every move(which then used as an input to other neural network).

Spoiler

https://www.youtube.com/watch?v=qUAmTYHEyM8&feature=youtu.be&t=22m23s

I wonder if Lee Sedol played the most recent database of alphago, or if they let him play an older version so that the games would be closer than if he played the most advanced version so far. I think there's a case to make that argument, assuming it learns every day and seeing as the games were relatively close in terms of points (10 is crushing in human vs human, but it's not huge in terms of point advantage that is possible on the board).

Spoiler

https://www.youtube.com/watch?v=qUAmTYHEyM8&feature=youtu.be&t=24m7s

 

[Haim]

Spoiler

https://www.youtube.com/watch?v=qUAmTYHEyM8&feature=youtu.be&t=22m23s

Spoiler

https://www.youtube.com/watch?v=qUAmTYHEyM8&feature=youtu.be&t=24m7s

I knew it already, still it does have a NN which job is to predict its opponent's next move and in it's training database among the rest there is every recorded match of Lee Sedol, it is to some extent fitted to predict his moves.AlphaGo is using deep NN, interly in order to predict it's opponent move it might divide it's predictions into playstyle or even specific player, the number of nodes it has is bigger than the number of games it has in the database, making it possible for specific predictions for Lee.By predictions I don't mean to predction of the next specific move but the probability of Lee to play at certain location, even very little improvement to the prediction accuracy like 0.0001% can prevent checking billions if not more of play states.

 

[Architect]

yes it does compute the board states just like most other algorithms do it in chess except with an excellent database and search narrowing.

You might not be saying this, but chess classically has used exhaustive tree search, deep neural nets have only been around since 2013. I'm not aware of any DNN chess programs and don't see the point as that's already better served by trees, whereas Go can't be searched like that.

I wonder if Lee Sedol played the most recent database of alphago, or if they let him play an older version

According to the commentary (available on YouTube) he's playing a vastly updated version. Which is what I'd expect, if I was on the engineering team I'd have him playing latest.

 

[Ex-User (9086)]

You might not be saying this, but chess classically has used exhaustive tree search, deep neural nets have only been around since 2013. I'm not aware of any DNN chess programs and don't see the point as that's already better served by trees, whereas Go can't be searched like that.

You are right. Although I think there's some search narrowing done anyway. It's true that Go being much more permutative and complex as it is, requires search narrowing by default and not simple tree search.

I'm still thinking there is some built in databases for openings and early game opening search narrowing, but I may be wrong I would have to get my facts straight at this point.

According to the commentary (available on YouTube) he's playing a vastly updated version. Which is what I'd expect, if I was on the engineering team I'd have him playing latest.

Yes, I read about it later, it was the latest version frozen in progress for the purpose of playing the 5 game match.

 

[Haim]

Architect do you have a guess how they trained their value network?The hardest part of Go AI is that it hard to detriment who is winning based on board state, I wonder how they solved it.

 

[Ex-User (9086)]

Probably a conclusive revelation. Recently alphago won 60 - 0 games against top pro players in online blitz matches. This makes computers the uncontested go champions. 60 consecutive wins in pro matches is unheard of and most if not all of them by resignation.

The games it plays are influence/potential oriented and it has a great sense of move urgency as well.

It seems to either force the opponent to risk losing the reducing groups or give alphago too much secure territory during the attack. Which would indicate that human sense of potential and urgency is flawed and/or lacks the necessary execution skills.

There will be some full duration official matches with alphago this year.