[Boards: 3 / a / aco / adv / an / asp / b / biz / c / cgl / ck / cm / co / d / diy / e / fa / fit / g / gd / gif / h / hc / his / hm / hr / i / ic / int / jp / k / lgbt / lit / m / mlp / mu / n / news / o / out / p / po / pol / qa / qst / r / r9k / s / s4s / sci / soc / sp / t / tg / toy / trash / trv / tv / u / v / vg / vip /vp / vr / w / wg / wsg / wsr / x / y ] [Search | Home]
4Archive logo
Google AI beats humans in Go
If images are not shown try to refresh the page. If you like this website, please disable any AdBlock software!

You are currently reading a thread in /sci/ - Science & Math

Thread replies: 62
Thread images: 4
Google has built an AI that is capable of beating human professional players in Go boardgame. What next? Skynet?

https://www.youtube.com/watch?v=SUbqykXVx0A
>>
> What next? Skynet?
you wish it was that easy. some AI beating some dumb guy means shit.
>>
>>7815212
AI takeover is the best thing that could happen to us. Political corruption squanders so much money that we'd all be millionaires by now if the system weren't so inefficient.
>>
File: stockphoto.jpg (26 KB, 403x312) Image search: [iqdb] [SauceNao] [Google]
stockphoto.jpg
26 KB, 403x312
Well, go is pretty difficult
>>
>>7815212
>The same had been done with chess decades ago
>It became so mundane and easy to make machines 'beat' humans that people just lost interest in it

>Google afraid no one cares about them anymore and just know them for their search
>W-we have to make something cool!
>Make the same shit that already existed for decades but now for another game.

When I was a retarded piece of shit retard practicing programming I made an AI that played connect 4. It is 'almost' perfect. Where is my fucking article?

Where is everyone's fucking article because anyone with half a brain can make an AI that plays a fucking game.
>>
>>7815252
>Whoosh.

Playing chess can be made by brute force. Go have a combinatorial explosion that requires a different approach.

In your retarded world there's no difference between cutting a tree branch with a stone axe and CNC laser cutting a pattern in a steel sheet.
>>
>>7815246
What makes you think the AI wouldn't just get rid of us?
>>
>>>/g/
>>
>>7815273
>What makes you think the AI wouldn't just get rid of us?

No loss if it kills us, for anyone like myself that have a value as a human being it'd be killing 1000 total morons that serve no purpose on this planet.

But as it's going to be designed by people it'll most likely be an optimizing agent acting in generalized human interests but without the vulnerability of bias and corruption.
>>
Here is a paper describing their ANN approach.
https://storage.googleapis.com/deepmind-data/assets/papers/deepmind-mastering-go.pdf
>>
File: meh.jpg (14 KB, 449x328) Image search: [iqdb] [SauceNao] [Google]
meh.jpg
14 KB, 449x328
>>7815212
Come back when they solve it.
>>
Monte carlo tree search is very cool. It's also very simple to understand.
>>
>>7815212
>AI
>Actually just ML and lots of training data
>>
>>7815252
>what is combinatorial explosion
come back when you are no longer a freshman
>>
>>7815331
It used a search algorithm together with an ANN. But ML is also a part of AI. If you're some popsci faggot crying about your general AI meme you should gtfo of my Taoist Monastry meme image board.
>>
>>7815260
Chess has a combinatorial explosion just like go. It's just a slightly smaller game tree. It's no surprise that top-tier go computers would emerge shortly after top-tier chess computers.
>>
>>7815401
This uses a significantly different approach than say Deep Blue used.

Because every step increases moves exponentially the branching factor is extremely significant. Having 400 moves per turn instead of 100 makes an enormous difference (more than a million more possibilities if you check ten moves and then run some heuristic evaluator).
>>
Combinatorial explosion isn't the problem. The problem is that there is no good, robust evaluation function for Go, unlike Chess.
>>
>>7815425
> The problem is that there is no good, robust evaluation function for Go
Which is due to combinatorial explosion.
>>
>>7815428
Not really. Chess positions can be evaluated relatively easily using the material value of pieces, the amount of space each player controls, the pawn structure, etc. Go doesn't have that, and that has nothing to do with combinatorial explosion.
>>
Call me when it can beat this guy

https://en.wikipedia.org/wiki/Lee_Chang-ho

>>7815331
>implying ML isn't ai
>>
>>7815541
It's going to beat Lee Sedol in March who is arguably in better form than Chang-Ho.
>>
I could totally beat this thing at beer pong, though.
>>
>>7815425
>The problem is that there is no good, robust evaluation function for Go
They just invented one
>>
>>7815401
>slightly
>>
>>7815217
The sad part is that this is still a specific environment which any hardcoded software could be programmed into beating a human in.
But afaik DeepMind is using Machine learning which is quite different to hardcoding a algorithm.
>>
>>7815212
So? There are people who can't beat godlike bots in unreal tournament.
>>
Why don't we use neuron-simulation to create a human-like brain by chance?
Even if one generation needs 30min to be simulated we could simulate 1-2million years (first human appearance up until now) in 114 years and considering that technology improves over time this date will probably get closer to our current time.

It's not a method that guarantees a brain when we reach this point but at least it does something.
>>
Computers have been able to beat players in games for the last 50 years.
>>
>>7815856
Because we don't know how to model neurons properly.
>>
>>7815979
Are you sure about that?

https://en.wikipedia.org/wiki/Neuron_%28software%29
>>
>>7815987
Yes, I am sure.
>>
>>7816003
I argue that we do understand neurons pretty well. Its just that we do not understand how the interaction inside a neuron network creates consciousness, thoughts and the rest.
>>
>>7816010
We don't actually.

Any moron that things we have a good model of how a neuron functions is mentally retarded.

None of those models are actually correct, they only appear correct because the outputs are "close enough."
>>
>>7816015
But we do what they are made of, which molecules trigger information signals etc.
>>
>>7816025
You do not.

MIT is doing mad research right now trying to figure out how the internals of the neuron work, turns out there are tons of micro tunnels inside for some reason.

Their function is unknown for now.
>>
Not to mention that we discover new types of neurons basically every time someone comes up with a new imaging technique.
>>
File: 152919079.png (81 KB, 591x368) Image search: [iqdb] [SauceNao] [Google]
152919079.png
81 KB, 591x368
Japanese pro player looked through the games and tweeted that Sedol might be having hard time when facing AlphaGo in March.
>>
>>7816030
>>7816015
neurons use their DNA as a turing machine tape.
ya heard it here first m8
>>
>>7816261
Do you have a single fact to back that up?
>>
>>7815217
Well the cool thing is they were able to beat them with thousands less board states that deep blue used to beat kasparov.

That alone is pretty fucking cool. Many problems can be reduced to tree search. Even engineering problems like designing machines and what not
>>
>>7815252
do you even know what Go is and how it's played
>>
>>7815260
>computers just brute force chess its easy shit
kek
>>
>>7817156
It's true though. Even my microwave running a chess engine would beat Magnus Carlsen.
>>
>>7817164
there's much more to computer chess than just bruteforcing
search and evaluation algorithms are still being worked on, it's not like your pc can calculate his way to the checkmate or even winning a piece
>>
>>7815212
Wow. Just a year ago people thought that defeating humans at Go would take decades due to its being immensely more complex than chess.
>>
>>7817172
The mathematician's perspective:
>Chess and Go are games that terminate in finite time
>Therefore they are deterministic, and if every end-state is evaluated as a win for either player I or player II, for each game one of the two players has a winning strategy obviously calculable in finite time
>Therefore these games are both trivial and trivial to solve
>>
Why do the retards in this thread whose maximum level of mathematical achievement is learning calculus keep pretending that they know better what Google is working on, then, uh... Google itself?

At least spend some time reading up how Google DeepMind works before bashing it, you imbeciles. It uses human-like learning algorithms, derived from models of the human brain and behavior. It's one of the most promising AI systems up to date.
>>
Is this actually a new development, or is it just alpha beta pruning running on faster hardware like every other AI win for the past few decades?
>>
>>7815212
What the fuck is Go?
>>
>>7817235
It's a completely different kind of approach. It uses neural networks and Monte Carlo tree search.
>>
>>7817242
Ah, then that is interesting.
How did they train the network?
>>
>>7817246
Their paper is available here:
https://storage.googleapis.com/deepmind-data/assets/papers/deepmind-mastering-go.pdf
>>
>>7817201
>Chess and Go are games that terminate in finite time
B-but what about triple ko?
>>
>>7817704
Only a problem with inferior rulesets. Whoever makes a move that repeats a previous board position should lose automatically.
>>
>>7815217
>some dumb guy
was a 5-0 wipeout of a 2-dan professional, mate. Granted, he's just the European champ (king of the kindergarten), but that is still stronger than pretty much all amateur Go players.

This is a huge step up from the previous state of the art! Google claims AlphaGo utterly wipes out all other top Go programs, a feat not seen in computer chess since Deep Thought.
>>
>>7815552
>It's going to beat Lee Sedol
don't count your chickens

Many mistakes were made on both sides, according to strong players. There is a lot of difference between a 2-p and a 9-p world champ.
>>
>>7818016
another astonishing thing: according to the paper, AlphaGo doesn't use fuseki or joseki! In chess terms, it plays this well without even an opening book.
>>
>>7818095
But it does use a neural network that has been trained by moves made by actual humans.
>>
>>7817241
A very old and very fun game.
>>
>>7817242
That in itself isn't new anymore. MCTS has been out for a decade, and many other Go programs have also dabbled with neural networks for evaluation. People are saying this is like the 2005 World Computer Chess Championship, where two unknown programs, one open source, crushed the field of established programs. They took a fresh look at existing ideas and simply implemented them better than anyone had before.
>>
>>7815212
>Go is more intuitive
>Chess is more logic-based

This is just wrong. The mental filter that determines which moves a good chess player even begins to consider is just as intuitive as for Go players and always the result of having played thousands of games. Humans can't just start playing by brute-forcing their way through with logic and reason. It's always about building an intuitive knowledge base. Once you have that, Go positions can be considered just as analytically as chess. There literally is no difference except in the details of the moves.
Thread replies: 62
Thread images: 4
Thread DB ID: 476216



[Boards: 3 / a / aco / adv / an / asp / b / biz / c / cgl / ck / cm / co / d / diy / e / fa / fit / g / gd / gif / h / hc / his / hm / hr / i / ic / int / jp / k / lgbt / lit / m / mlp / mu / n / news / o / out / p / po / pol / qa / qst / r / r9k / s / s4s / sci / soc / sp / t / tg / toy / trash / trv / tv / u / v / vg / vip /vp / vr / w / wg / wsg / wsr / x / y] [Search | Home]

[Boards: 3 / a / aco / adv / an / asp / b / biz / c / cgl / ck / cm / co / d / diy / e / fa / fit / g / gd / gif / h / hc / his / hm / hr / i / ic / int / jp / k / lgbt / lit / m / mlp / mu / n / news / o / out / p / po / pol / qa / qst / r / r9k / s / s4s / sci / soc / sp / t / tg / toy / trash / trv / tv / u / v / vg / vip /vp / vr / w / wg / wsg / wsr / x / y] [Search | Home]

All trademarks and copyrights on this page are owned by their respective parties. Images uploaded are the responsibility of the Poster. Comments are owned by the Poster.
This is a 4chan archive - all of the shown content originated from that site. This means that 4Archive shows their content, archived. If you need information for a Poster - contact them.
If a post contains personal/copyrighted/illegal content, then use the post's [Report] link! If a post is not removed within 24h contact me at [email protected] with the post's information.