Google AI algorithm masters ancient game of Go

[The_Assimilator]

http://www.nature.com/news/google-ai-algorithm-masters-ancient-game-of-go-1.19234

A computer has beaten a human professional for the first time at Go — an ancient board game that has long been viewed as one of the greatest challenges for artificial intelligence (AI).

...

DeepMind’s program AlphaGo beat Fan Hui, the European Go champion, five times out of five in tournament conditions, the firm reveals in research published in Nature on 27 January. It also defeated its silicon-based rivals, winning 99.8% of games against the current best programs.

...

“This is a really big result, it’s huge,” says Rémi Coulom, a programmer in Lille, France, who designed a commercial Go program called Crazy Stone. He had thought computer mastery of the game was a decade away.

...

But the game has long interested AI researchers because of its complexity. The rules are relatively simple: the goal is to gain the most territory by placing and capturing black and white stones on a 19 × 19 grid. But the average 150-move game contains more possible board configurations — 10^170 — than there are atoms in the Universe, so it can’t be solved by algorithms that search exhaustively for the best move.

...

Rather than follow the trend of the past 30 years of trying to crack games using computing power, DeepMind has reverted to mimicking human-like knowledge, albeit by training, rather than by being programmed, he says. The feat also shows the power of deep learning, which is going from success to success, says Coulom. “Deep learning is killing every problem in AI.”

This is a massive step toward computers that can actually think, as opposed to simply using brute force to figure things out.

 

[konfab]

This is a massive step toward computers that can actually think, as opposed to simply using brute force to figure things out.

I disagree with that. Deep learning is a brute force way of solving a problem.
Edit:
Not that it is a bad way to do it.

 

[cyberbob1979]

That is seriously impressive - well done to the team and the programmers behind it

 

[saor]

This is a massive step toward computers that can actually think, as opposed to simply using brute force to figure things out.

I don't know about that. The system only seems to learn how to maximize a particular variable of the space it inhabits. In this example:

[video=youtube;V1eYniJ0Rnk]https://www.youtube.com/watch?v=V1eYniJ0Rnk[/video]

The AI has no concept of it playing a game or of the concepts 'ball', 'bounce' and 'blocks'. All it does it gain familiarity with the system and performs those actions which maximize the variable 'score'. If the value of score increases, then it leans toward executing more of those actions.

 

[EtienneK]

I don't know about that. The system only seems to learn how to maximize a particular variable of the space it inhabits. In this example:

[video=youtube;V1eYniJ0Rnk]https://www.youtube.com/watch?v=V1eYniJ0Rnk[/video]

The AI has no concept of it playing a game or of the concepts 'ball', 'bounce' and 'blocks'. All it does it gain familiarity with the system and performs those actions which maximize the variable 'score'. If the value of score increases, then it leans toward executing more of those actions.

Holy smokes. That is ... scary.

 

[konfab]

I don't know about that. The system only seems to learn how to maximize a particular variable of the space it inhabits.

Intelligence is maximising a variable without testing every single input parameter.

Gradient decent systems are partially better than randomly guessing a solution, but they don't really do it with a model of reality (the concepts of ball, bounce and blocks)

 

[K3NS31]

I don't know about that. The system only seems to learn how to maximize a particular variable of the space it inhabits. In this example:

[video=youtube;V1eYniJ0Rnk]https://www.youtube.com/watch?v=V1eYniJ0Rnk[/video]

The AI has no concept of it playing a game or of the concepts 'ball', 'bounce' and 'blocks'. All it does it gain familiarity with the system and performs those actions which maximize the variable 'score'. If the value of score increases, then it leans toward executing more of those actions.

Perhaps, but think about it - how do you think you learnt things when you were a couple of months old?

 

[saor]

Perhaps, but think about it - how do you think you learnt things when you were a couple of months old?

What kinds of things are we talking about here - motor functions, language, hearing certain sounds, rote memorization etc.? The means by which knowledge comes about in those domains follows slightly different pathways. For instance - learning to crawl is more a realization that "Oh! My body can do this!?" & doesn't require much of an external reference of crawling (babies learn to crawl without having seen other creatures crawl), whereas the acquisition of language, or rather - the building blocks of your parents language - is dependent on external references ie. a baby won't say 'banana' without first having heard it.

With regards to DeepMind: Crawling and saying 'banana' don't have much immediate feedback as to their value or application; they are initially a form of novelty seeking which later gets utilized in order to navigate the world and get some piece of it. DeepMind seems to be motivated from the very start by an acute sense of goal-directedness, of maximizing a particular value.

 

[K3NS31]

What kinds of things are we talking about here - motor functions, language, hearing certain sounds, rote memorization etc.? The means by which knowledge comes about in those domains follows slightly different pathways. For instance - learning to crawl is more a realization that "Oh! My body can do this!?" & doesn't require much of an external reference of crawling (babies learn to crawl without having seen other creatures crawl), whereas the acquisition of language, or rather - the building blocks of your parents language - is dependent on external references ie. a baby won't say 'banana' without first having heard it.

With regards to DeepMind: Crawling and saying 'banana' don't have much immediate feedback as to their value or application; they are initially a form of novelty seeking which later gets utilized in order to navigate the world and get some piece of it. DeepMind seems to be motivated from the very start by an acute sense of goal-directedness, of maximizing a particular value.

Well, that's not how I saw it when my son learned to crawl and walk, for example (which happened recently - he's only 1.5) - up to about 10 or 11 months, almost every single thing he did was goal directed. He desperately needed to crawl to get himself around. He wanted to do it so badly, and would practice any type of movement which allowed him to propel himself. (eg. bear crawling, which not all kids do). The same with walking - the goal was to be higher, see more and move faster. And he worked towards it and maximised whatever movement / technique would help towards those goals.
Ditto talking - he's learning to talk because he's finally decided that he needs to communicate more effectively with us. We could see the change. Just 3 or 4 months ago he would hardly vocalise anything, as he didn't care about communicating as much as he cared about walking, running and climbing (and eating, of course).
Then there was a switch, and he started practising towards his goal, to the point where he's constantly vocal now, even though he still can't say many actual words.

The point I'm getting at is my experience from watching a baby learn basic skills which I'm too old to remember learning, is that it's a lot more similar to what deep mind is doing in that breakout game than you think.

 

[cguy]

I don't know about that. The system only seems to learn how to maximize a particular variable of the space it inhabits. In this example:

[video=youtube;V1eYniJ0Rnk]https://www.youtube.com/watch?v=V1eYniJ0Rnk[/video]

The AI has no concept of it playing a game or of the concepts 'ball', 'bounce' and 'blocks'. All it does it gain familiarity with the system and performs those actions which maximize the variable 'score'. If the value of score increases, then it leans toward executing more of those actions.

The idea behind deep learning is that those concepts would be emergent as higher level features in one or more layers of the net.

 

[ponder]

In Saudi the computer will be banned and the programmer executed....

 

[SoulTax]

Well, that's not how I saw it when my son learned to crawl and walk, for example (which happened recently - he's only 1.5) - up to about 10 or 11 months, almost every single thing he did was goal directed. He desperately needed to crawl to get himself around. He wanted to do it so badly, and would practice any type of movement which allowed him to propel himself. (eg. bear crawling, which not all kids do). The same with walking - the goal was to be higher, see more and move faster. And he worked towards it and maximised whatever movement / technique would help towards those goals.
Ditto talking - he's learning to talk because he's finally decided that he needs to communicate more effectively with us. We could see the change. Just 3 or 4 months ago he would hardly vocalise anything, as he didn't care about communicating as much as he cared about walking, running and climbing (and eating, of course).
Then there was a switch, and he started practising towards his goal, to the point where he's constantly vocal now, even though he still can't say many actual words.

The point I'm getting at is my experience from watching a baby learn basic skills which I'm too old to remember learning, is that it's a lot more similar to what deep mind is doing in that breakout game than you think.

I have a similar experience.

Even something as simple as grabbing hair is a stumbled upon action. A young baby will accidentally grab your hair and then realise that grabbing at your hair is fun, probably because you go "ow ow ow!" in ever increasing volume and pitch. It then becomes a goal of sorts, to grab at your hair on purpose.

All learning is goal driven. And in actual fact, this Deep Thought thing has displayed learning in both primary ways that humans learn.
1) It learned by watching and remembering, with the GO game.
2) It learned by doing and failing, with the attached video.

Those are literally the only two ways that we learn as humans.

The only difference as far as I can see. Is that our learning allows us to apply knowledge outside of the closed bubble in which we attained that knowledge. If for instance a computer, without any coding to tell it to draw parallels between chess pieces and a military strategy board, could learn to play chess and then apply that knowledge dynamically and abstractly to another area; That to me would be another big step.

 

[cguy]

All learning is goal driven. And in actual fact, this Deep Thought thing has displayed learning in both primary ways that humans learn.
1) It learned by watching and remembering, with the GO game.
2) It learned by doing and failing, with the attached video.

Those are literally the only two ways that we learn as humans.

The only difference as far as I can see. Is that our learning allows us to apply knowledge outside of the closed bubble in which we attained that knowledge. If for instance a computer, without any coding to tell it to draw parallels between chess pieces and a military strategy board, could learn to play chess and then apply that knowledge dynamically and abstractly to another area; That to me would be another big step.

This already happens to a degree - if you had to take a deep net trained to recognize pictures of dogs, and then train it to recognize pictures of cats, it would converge much faster than if it was just a blank or random net to begin with. The reason being that many of the requisite high level features are so similar - e.g., the edge detection filters, and such generated at the first hidden layer(s) would be nearly identical.

 

[Knyro]

Looks like Google touched the Japanese on their studio. Dwango, partnered with the University of Tokyo and an association of professional go players in Japan have teamed up on a project called DeepZenGo. It is a project specifically created to defeat AlphaGo at go. They aim to pull it off in a timeline of six months to a year.

The new version of Atlas by Boston Dynamics also rustled some jimmies because it makes Asimo look like a piece of junk.

 

[K3NS31]

Looks like Google touched the Japanese on their studio. Dwango, partnered with the University of Tokyo and an association of professional go players in Japan have teamed up on a project called DeepZenGo. It is a project specifically created to defeat AlphaGo at go. They aim to pull it off in a timeline of six months to a year.

The new version of Atlas by Boston Dynamics also rustled some jimmies because it makes Asimo look like a piece of junk.

Geez that new Atlas is amazing. When it walks in the snow!

 

[The_Assimilator]

AlphaGo is now 2-0 against the 18-time Go world champion Lee Se-dol:

http://www.theverge.com/2016/3/10/11191184/lee-sedol-alphago-go-deepmind-google-match-2-result

Google stunned the world by defeating Go legend Lee Se-dol yesterday, and it wasn't a fluke — AlphaGo, the AI program developed by Google's DeepMind unit, has just won the second game of a five-game Go match being held in Seoul, South Korea. AlphaGo prevailed in a gripping battle that saw Lee resign after hanging on in the final period of byo-yomi ("second-reading" in Japanese) overtime, which gave him fewer than 60 seconds to carry out each move.

"Yesterday I was surprised but today it's more than that — I am speechless," said Lee in the post-game press conference. "I admit that it was a very clear loss on my part. From the very beginning of the game I did not feel like there was a point that I was leading."

...

DeepMind's AlphaGo program uses an advanced system based on deep neural networks and machine learning, which has now seen it beat 18-time world champion Lee twice. The series is the first time a professional 9-dan Go player has taken on a computer, with Lee competing for a $1 million prize. AlphaGo's victory today means it leads the series 2-0; Lee had predicted he'd win 5-0 or 4-1 at worst, but he now needs to come out on top in all three remaining games whereas AlphaGo could wrap up the series by Saturday.

The singularity is coming. Are you ready to become obsolete?

 

[The_Assimilator]

AlphaGo 3 | 0 Lee Se-dol

AlphaGo takes the series, although the two remaining games will still be played. This is already history-making, but if AlphaGo can cleanly win the series, it will be even more important.