Google’s AI Wins Pivotal Second Game In Match With Go Grandmaster [Wired]

[nevets88]

Google’s AI Wins Pivotal Second Game In Match With Go Grandmaster

Quote

‘I Am In Shock’

Following Game One, Lee Sedol acknowledged he was “shocked” by how well AlphaGo played and said he’d made a notable mistake at the beginning of the game that led to his loss about three hours later. “The failure I made at the very beginning of the game lasted until the the very end,” he said, through an interpreter. “I didn’t think that AlphaGo would play the game in such a perfect manner.” It’s unclear what early mistake he was referring to. The match’s English language commentators didn’t see one. But they do feel the Korean made a rather large error late in the game, following some particularly skillful play by AlphaGo. In any event, Lee Sedol did resolve to change his approach in Game Two.

http://www.wired.com/2016/03/googles-ai-wins-pivotal-game-two-match-go-grandmaster/

 

[Abu3baid]

I thought this was about chess.. But it turned out to be a game that I dont know.

 

[rkim291968]

Am I the only one who watched both matches on YouTube?    

I only know the basics of Go but know a bit more about AI, given my background as a computer science major.   Back in 80's when I studied AI at CAL, it was given that what AlphaGO is doing now is not possible.   Even it is, no one at the time could imagine the HW computing power that the SW can utilize.   At last, both HW & SW are available.   The only way Lee can beat this computer is to find a new strategy that is not in AlphaGO's database.   What's the odd of that?  None.    Given how the 1st two matches went, I don't see anyone beating AlphaGO in go.  

[nevets88]

8 hours ago, rkim291968 said:

Am I the only one who watched both matches on YouTube?    

I only know the basics of Go but know a bit more about AI, given my background as a computer science major.   Back in 80's when I studied AI at CAL, it was given that what AlphaGO is doing now is not possible.   Even it is, no one at the time could imagine the HW computing power that the SW can utilize.   At last, both HW & SW are available.   The only way Lee can beat this computer is to find a new strategy that is not in AlphaGO's database.   What's the odd of that?  None.    Given how the 1st two matches went, I don't see anyone beating AlphaGO in go.  

The play by play described in the article was pretty good. I'll watch the video over the weekend. The scary thing is there's no database. It's learning based. This is mentioned somewhere in the middle of the article. I am rooting for the human tomorrow. Three in a row! Beat that machine!

There's no putting off the singularity, but we can delay it. Be... very... afraid...

I haven't been keeping up with chess and AI. Did they slow down developing AI players since Deep Blue defeated the champion? Chess AI must be very unbeatable by humans now, probably available on more commodity machines.

[saevel25]

9 hours ago, rkim291968 said:

At last, both HW & SW are available.   The only way Lee can beat this computer is to find a new strategy that is not in AlphaGO's database.   What's the odd of that?  None.    Given how the 1st two matches went, I don't see anyone beating AlphaGO in go.

A person would have to have an extremely strong middle and end game to beat the computer. I see it as the game progresses it can fine tune the outcome it wants. I think a good point was made. That computer is looking for the highest percentage even if that means a 1 point win.

I agree that it might come down to an unknown play in the mid to late game that forces the computer to have to be more intuitive which might be it's weakness.

[Big Lex]

I know very little about artificial intelligence, but even if I did understand it, I'm not sure I'd know what to make about chess-playing computers.

I consider myself a pretty intelligent person. I've achieved at a high level in academics, I have a good memory, can solve problems well, can compute things in my head quickly, etc. I say this not to brag, but to set the stage for why I am confused by chess computers. I'm terrible at chess. I love the game, but I'm simply no good at it. If I play against a software chess opponent, I lose at anything beyond the beginner levels.

I haven't practiced chess much, and so maybe it's just that. But I can't help but think that chess is a specialized skill. I think that to be good at it requires more than simply thinking ahead as many moves as possible.

We hear a lot about computers playing chess and beating grandmasters. It's impressive, I agree. I'd like to hear about a computer doing something more mundane, like maybe talking to members of a family, and then scheduling the family's appointments, making shopping lists, giving them reminders, etc.

 

[rkim291968]

1 hour ago, nevets88 said:

The play by play described in the article was pretty good. I'll watch the video over the weekend. The scary thing is there's no database. It's learning based. This is mentioned somewhere in the middle of the article. .

In both videos (3 plus hours per video), they invited AlphaGo folks for Q&A.   They described how AlphaGo choose its (I almost wrote "his") move.   It picks the highest winning percentage move from its understanding.   I called that knowledge base a "database," not in the traditional sense of the meaning.   Whatever it is drawing its AI from, I referred to it as the database.

20 minutes ago, Big Lex said:

We hear a lot about computers playing chess and beating grandmasters. It's impressive, I agree. I'd like to hear about a computer doing something more mundane, like maybe talking to members of a family, and then scheduling the family's appointments, making shopping lists, giving them reminders, etc.

 

They already do, a lot more than we realize. It's all around us.  

[nevets88]

36 minutes ago, rkim291968 said:

In both videos (3 plus hours per video), they invited AlphaGo folks for Q&A.   They described how AlphaGo choose its (I almost wrote "his") move.   It picks the highest winning percentage move from its understanding.   I called that knowledge base a "database," not in the traditional sense of the meaning.   Whatever it is drawing its AI from, I referred to it as the database.

They already do, a lot more than we realize. It's all around us.  

The terminology to me is confusing. It sounds like a database to me too, but here's the section of the article where it says it's not.

Quote

During Game One, match commentators Michael Redmond and Chris Garlock didn’t seem to understand that AlphaGo operated in this way. Redmond kept referring to AlphaGo’s “database” of moves—something it doesn’t really have. Once the system is trained using those machine learning techniques, it plays entirely on its own. By Game Two, Redmond and Garlock were wise to this, after some coaching from the DeepMind team over breakfast here at the Four Seasons.

Here's how it's described in Google's blog. It sounds to me like it's more a "dynamic database". Regardless, maybe the team isn't doing a good job of describing it because it's just hard to describe in layman's terms.

Quote

Traditional AI methods—which construct a search tree over all possible positions—don’t have a chance in Go. So when we set out to crack Go, we took a different approach. We built a system, AlphaGo, that combines an advanced tree search with deep neural networks. These neural networks take a description of the Go board as an input and process it through 12 different network layers containing millions of neuron-like connections. One neural network, the “policy network,” selects the next move to play. The other neural network, the “value network,” predicts the winner of the game.

We trained the neural networks on 30 million moves from games played by human experts, until it could predict the human move 57 percent of the time (the previous record before AlphaGo was 44 percent). But our goal is to beat the best human players, not just mimic them. To do this, AlphaGo learned to discover new strategies for itself, by playing thousands of games between its neural networks, and adjusting the connections using a trial-and-error process known as reinforcement learning. Of course, all of this requires a huge amount of computing power, so we made extensive use of Google Cloud Platform.

https://googleblog.blogspot.com/2016/01/alphago-machine-learning-game-go.html

[rkim291968]

1 hour ago, nevets88 said:

The terminology to me is confusing. It sounds like a database to me too, but here's the section of the article where it says it's not.

Here's how it's described in Google's blog. It sounds to me like it's more a "dynamic database". Regardless, maybe the team isn't doing a good job of describing it because it's just hard to describe in layman's terms.

https://googleblog.blogspot.com/2016/01/alphago-machine-learning-game-go.html

Dynamic database is close to what AI is, IMO.   Instead of calling it database, let's say that it has a tree it can navigate, each branch leading to a different decision.   The algorithm that decides how it traverse the tree is the core of the AI SW.  But at the end of the day, the SW relies on past information on how a move was played out and picks the most winning move.  I think that was mentioned in the 1st video somewhere.

I was taught that computer cannot teach itself and I still believe that.   That's why I mentioned that Lee needs to come with something the computer hasn't seen before (or not in its tree).   That is not going to be possible if AlphaGO folks has done their job which I think they did.

I am still hoping that Lee wins the match.  

[nevets88]

Humans! Humans! Humans! He took a match from AlphaGo.

Quote

 Share on Facebook  Tweet  Share  Pin     
AlphaGo wrapped up victory for Google in the DeepMind Challenge Match by winning its third straight game against Go champion Lee Se-dol yesterday, but the 33-year-old South Korean has got at least some level of revenge — he's just defeated AlphaGo, the AI program developed by Google's DeepMind unit, in the fourth game of a five-game match in Seoul.

AlphaGo is now 3-1 up in the series with a professional record, if you can call it that, of 9-1 including the 5-0 win against European champion Fan Hui last year. Lee's first win came after an engrossing game where AlphaGo played some baffling moves, prompting commentators to wonder whether they were mistakes or — as we've often seen this week — just unusual strategies that would come good in the end despite the inscrutable approach. (To humans, at least.)

http://www.theverge.com/2016/3/13/11184328/alphago-deepmind-go-match-4-result

 

[rkim291968]

The battle ended with Alphago taking 4 of 5 matches.   I watched the last match from the start to finish.  It's clear Lee was winning but made a few humanly errors.   And that's all it took for Alphago to come from behind to beat Lee.  

Some random thoughts:

• Even the best GO masters will make humanly mistakes in a course of 5 hour match.  Computers won't.   It will have bugs/uncharted moves but won't make mistakes which is not in the program.
• Alphago algorithm will only improve, more rapidly than the best GO masters to learn new moves.   Unless DeepMind folks stop working on Alphago program, humans will not beat the beast in GO.
• Would the outcome change if it was Alphago vs Ke Jie (sp?), the reining #1 GO master in the world?  I think not but it would have been more dramatic.  Will there be Alphago vs Ke Jie?   
• A huge victory for the GO world, Lee Sedol, DeepMind, Google, and AI.   No losers.  Everyone's a winner in this event.    

I am contemplating learning the game.   

[nevets88]

The meaning of AlphaGo, the AI program that beat a Go champ

Quote

Q: So, why is it important that AI triumphed in the game of Go?

A: It relies on a lot of intuition. The really skilled players just sort of see where a good place to put a stone would be. They do a lot of reasoning as well, which they call reading, but they also have very good intuition about where a good place to go would be, and that’s the kind of thing that people just thought computes couldn’t do. But with these neural networks, computers can do that too. They can think about all the possible moves and think that one particular move seems a bit better than the others, just intuitively. That’s what the feed point neural network is doing: it’s giving the system intuitions about what might be a good move. It then goes off and tries all sorts of alternatives. The neural networks provides you with good intuitions, and that’s what the other programs were lacking, and that’s what people didn’t really understand computers could do.

http://www.macleans.ca/society/science/the-meaning-of-alphago-the-ai-program-that-beat-a-go-champ/

[jkelley9]

Do you guys actually play Go, or are more into this from the AI perspective? I used to play a lot, got down to about a 4 kyu on CGoban. I don't even know if that software platform still exists now, it's been so long. I had a club I played in while in college which was loads of fun. When I moved to the Carolinas, those clubs are non-existent. 

It's very impressive what this AI is able to do. Go is a much more linearly progressive game. You are adding physical data to the board each move unlike chess where you're removing physical data. But alternatively, (almost) EVERY empty spot is an available move for either player, whereas in chess your "pieces" can only move to a limited number of spots.

Despite that there are more available combinations of moves in a game of Go than there are atoms in the entire universe (let that sink in for a moment, roughly ~361! (factorial), or 361x360x359x358x... etc with slightly [relatively speaking] more combinations than that given that there are captures that open spaces, and traded captures)... I actually think it would be easier to develop a bruteforce AI for Go than chess due to the nature of the game.

I think it's only taken this long to develop such a strong AI only because of the relative popularity of chess vs Go. There's chess clubs everywhere. However, I spent 6 weeks all over Japan last year and never once found a Go club. It's basically dead except for the few traditionalists and old folks who grew up with it.

[saevel25]

11 minutes ago, jkelley9 said:

I think it's only taken this long to develop such a strong AI only because of the relative popularity of chess vs Go. There's chess clubs everywhere. However, I spent 6 weeks all over Japan last year and never once found a Go club. It's basically dead except for the few traditionalists and old folks who grew up with it.

I wonder if it has to do more with the internet. I would suspect more younger GO players would just play on the computer then spend their time in a GO Salon. I am not how much actual popularity of the game has decreased. 

[jkelley9]

8 minutes ago, saevel25 said:

I wonder if it has to do more with the internet. I would suspect more younger GO players would just play on the computer then spend their time in a GO Salon. I am not how much actual popularity of the game has decreased. 

I suppose that's possible. But I will literally go 6-12 months at a time without ever hearing or even reading anything (internet or elsewhere) related to "Go." 

I'm all about using computers. If you only know how much of a tech geek I really am (I have 2 servers and a PC in my HOME office that I all built, configured, and maintain myself) but when it comes to boards games like Go, they're so much more fun playing against someone across a real board. Man... I miss that! I may stop by the local Atlanta club this year when I'm back in that area, if that club even still meets!