World's top Go player loses second match of five against Google AI

[kurros]

Games are won by massive search trees. Computers will always be able to do this better. No news. No surprise.

Again, no they aren't, at least Go isn't, and that isn't what AlphaGo does, not entirely, or even primarily. It does some searching, but so do humans. But it has two neural networks trained to figure out whether is winning, and what it's next move should be. These are the key elements to its success. The brute force search tree method has been tried and failed at Go for decades, and computers are still nowhere near powerful enough to win via this tactic, and may never be. The tree to be searched is just too enormous.

Compare just the state space complexity of chess vs Go. I.e. the number of board configurations. For chess this is 10^47. For Go it is 10^79. Have computers become 32 orders of magnitude faster in the 20 years since a brute force search defeated Kasparov at Chess? No. They have become, at best, two or three orders of magnitude faster. So are still roughly 30 orders of magnitude too slow to tackle Go this way. This is crude estimate of course but surely you get the point.

https://en.m.wikipedia.org/wiki/Game_complexity

[wonderer]

Kurros,

Please tell me what I'm getting wrong. Help me.
(And don't just say...oh you don't understand....read moar..it's really special...neural networks...

Why think that a forum post could lead you to having any sort of intuitive sense of how neural networks work?

The neural networks in your own brain haven't been exposed to a relevant training set.  A few words in a post here aren't going to change that.  Basically, until you go to the effort of trying to comprehend the operation of neural nets, you will be left with poor intuitions about what is going on in the case of AlphaGo.

Of course, if you are in denial of the relevance of the neural networks in your brain to your own thinking, I can't expect you to take this admonition seriously. 

[Biep]

Self-sufficient, autonomous AI writing its own AI software script. THAT would send shivers up my spine.

Something like Douglas Lenat's Eurisko?  That program also became world champion in quite a few (obscure) fields, including one game where they kept changing the rules in order to prevent Eurisko from taking advantage of loopholes it found, and Eurisko kept finding new ways to win the game.

And Eurisko was written as a generalisation of AM (1976), a program that from basic set theory discovered numbers, algebra (including the unique prime factorisation theorem) - partly to counter criticism about the opaque nature of the AM innards.  His current ongoing project is Cyc, which is about representing all the knowledge as bright humans have and more - including all the vague and informal, redundant, and/or contradictory facts that we carry with us.

Maybe interesting for this forum is that, based on his experience of seeing evolution in action within Eurisko (as opposed to biologists who haven't seen much evolution, and certainly not inspectable in detail), Lenat has intelligently critiqued standard (neo-)Darwinism.

Unfortunately I am separated by more than 10 000 km from my books and articles (which are boxed and stored in a container anyway :-) - so I cannot provide you with all the juicy details.

[aleph naught]

Games are won by massive search trees. Computers will always be able to do this better. No news. No surprise.

AlphaGo is not doing that, which is exactly why it's news.

[RichardChad]

Games are won by massive search trees. Computers will always be able to do this better. No news. No surprise.

AlphaGo is not doing that, which is exactly why it's news.

Incorrect.


AlphaGo combines two AI methodologies:

Monte Carlo tree search: This involves choosing moves at random and then simulating the game to the very end to find a winning strategy.

Deep neural networks: A 12 layer-network of neuron-like connections that consists of a "policy network" that selects the next move and a "value network" that predicts the winner of the game.

[kurros]

Games are won by massive search trees. Computers will always be able to do this better. No news. No surprise.

AlphaGo is not doing that, which is exactly why it's news.

Incorrect.


AlphaGo combines two AI methodologies:

Monte Carlo tree search: This involves choosing moves at random and then simulating the game to the very end to find a winning strategy.

Deep neural networks: A 12 layer-network of neuron-like connections that consists of a "policy network" that selects the next move and a "value network" that predicts the winner of the game.

I don't know how you can read the quoted text and still stick to what you are saying. Yes there is a tree search, but the bot would *lose* to a grand master human player if that is what it relied on, because it cannot search thoroughly enough. The deep learning neural networks are absolutely essential to how AlphaGo achieved victory here.

[TheBigOhMan]

This is part of how humans learn, but it doesn't have the ability to freely choose that humans have.

No one ever said it did. We're talking about artificial intelligence, not free will.

Ive always suspected that the reason why no machine will ever have free will is because free will itself is impossible.

Were you free to come to that conclusion?

Probably not,

Therefore you cannot rationally affirm it

¿What do you mean by rational?

[RichardChad]

Games are won by massive search trees. Computers will always be able to do this better. No news. No surprise.

AlphaGo is not doing that, which is exactly why it's news.

Incorrect.


AlphaGo combines two AI methodologies:

Monte Carlo tree search: This involves choosing moves at random and then simulating the game to the very end to find a winning strategy.

Deep neural networks: A 12 layer-network of neuron-like connections that consists of a "policy network" that selects the next move and a "value network" that predicts the winner of the game.

I don't know how you can read the quoted text and still stick to what you are saying. Yes there is a tree search, but the bot would *lose* to a grand master human player if that is what it relied on, because it cannot search thoroughly enough. The deep learning neural networks are absolutely essential to how AlphaGo achieved victory here.

What I said is completely accurate.

Neural networks are simply more sophisticated types of networks, computers are deterministically executing these weighted search algorithms. What enables the technology is simply HW horsepower. There is not, and will never be, any sentience as computers will always execute deterministically.

[kurros]

Games are won by massive search trees. Computers will always be able to do this better. No news. No surprise.

AlphaGo is not doing that, which is exactly why it's news.

Incorrect.


AlphaGo combines two AI methodologies:

Monte Carlo tree search: This involves choosing moves at random and then simulating the game to the very end to find a winning strategy.

Deep neural networks: A 12 layer-network of neuron-like connections that consists of a "policy network" that selects the next move and a "value network" that predicts the winner of the game.

I don't know how you can read the quoted text and still stick to what you are saying. Yes there is a tree search, but the bot would *lose* to a grand master human player if that is what it relied on, because it cannot search thoroughly enough. The deep learning neural networks are absolutely essential to how AlphaGo achieved victory here.

What I said is completely accurate.

Neural networks are simply more sophisticated types of networks, computers are deterministically executing these weighted search algorithms. What enables the technology is simply HW horsepower. There is not, and will never be, any sentience as computers will always execute deterministically.

Neural networks are nothing like a tree search. And no-one was talking about sentience so I don't know why you bring it up. But anyway your argument makes no sense since quantum computers do not execute deterministically. And in fact quantum random number generator cards already exist and are used industrially in normal computers as true random number generators, for generating encryption keys and such. So already ordinary computers can run true non-deterministic algorithms. And even before quantum technologies people have long used measurements of the environment, from temperature sensors and so on, to introduce true randomness into computers.

[RichardChad]

Neural networks are nothing like a tree search. And no-one was talking about sentience so I don't know why you bring it up. But anyway your argument makes no sense since quantum computers do not execute deterministically. And in fact quantum random number generator cards already exist and are used industrially in normal computers as true random number generators, for generating encryption keys and such. So already ordinary computers can run true non-deterministic algorithms.

Please do some research on the various types of networks, they are all deterministic and they are all variations on the same basic theory.
Quantum computers produce predictable results, if they didn't they would be useless.
What does quantum level indeterminism have to do with anything?

Always the same with atheists, challenge them to explain why the treat others as if they could have done otherwise and they start appealing to quantum indeterminism as a refutation of determinism. As clearly articulated by atheist August Berkshire: The Illusion That Free Will is Not an Illusion
none of that gets you the ability to choose. A computer will never have that ability, and that is the entire point.

[kurros]

Neural networks are nothing like a tree search. And no-one was talking about sentience so I don't know why you bring it up. But anyway your argument makes no sense since quantum computers do not execute deterministically. And in fact quantum random number generator cards already exist and are used industrially in normal computers as true random number generators, for generating encryption keys and such. So already ordinary computers can run true non-deterministic algorithms.

Please do some research on the various types of networks, they are all deterministic and they are all variations on the same basic theory.

I didn't say they weren't deterministic. And how about you do some research because it sounds like you have no idea how a neural network works. I have programmed one so I know how they work just fine.

Quantum computers produce predictable results, if they didn't they would be useless.
What does quantum level indeterminism have to do with anything?

You made a weird claim based on determinism, I showed that it was based on a false premise with that example. You brought it up.

Always the same with atheists, challenge them to explain why the treat others as if they could have done otherwise and they start appealing to quantum indeterminism as a refutation of determinism. As clearly articulated by atheist August Berkshire: The Illusion That Free Will is Not an Illusion
none of that gets you the ability to choose. A computer will never have that ability, and that is the entire point.

I don't know why you think this is relevant. No one in this thread claimed that AlphaGo has free will or is conscious or any such thing. My issue was with your claim that this is not news because "obviously" computers can beat humans via tree searches. This is just wrong in lots of ways, that have since been explained to you. You then claimed that neural networks are "just like" tree searches, which is also just wrong. Now you are talking about free will for some reason. You are just all over the place.

[RichardChad]

I've worked for decades with various network typologies and search methodologies including bayesian networks and DAGs, they are all variations on the basic notions.

To the layman the statement "Games are won by massive search trees. Computers will always be able to do this better. No news. No surprise."  is absolutely correct. It is also specifically correct in the case of AlphaGo due to the use of Monte Carlo.

[Trinity]

Lee Sedol wins match 4 against AlphaGo.

http://www.ibtimes.com/lee-sedol-vs-alphago-after-three-straight-losses-korean-champion-strikes-back-maiden-2335472

[kurros]

I've worked for decades with various network typologies and search methodologies including bayesian networks and DAGs, they are all variations on the basic notions.

To the layman the statement "Games are won by massive search trees. Computers will always be able to do this better. No news. No surprise."  is absolutely correct. It is also specifically correct in the case of AlphaGo due to the use of Monte Carlo.

Ok well I violently disagree with that and think it grossly misrepresents the situation, especially to laypeople.

[Lion IRC]

...But it has two neural networks trained to figure out whether is winning, and what it's next move should be.

It 'figures out' what its next move should be based on stats. And those data indicate nothing more or less than the probability of a win should it take that particular pathway. Where the probability is 50:50 the game can be programmed to 'toss a coin'. (In fact it is forced to toss a coin because the game would freeze otherwise.) And the only reason the software might encounter a 50:50 choice is because IT HAS NO DATA INPUTS TELLING IT whether one path is better than the other.

"Neural network" is lovely jargon and I don't mind that it is used as a metaphor. I don't even mind a little bit of human vanity/ego getting involved and patting ourselves on the back at our own cleverness. But let's not look at a tromp L'oeil and mistake it for the real thing.

The Alpha-Go software isn't patting itself on the back is it?