So apparently AlphaGo came back with a vengeance and has been demolishing 9d players left and right...

i think you should read up more on how go works, how alphago works, and how neural nets work

i read the paper

>I think calling something "alpha" is different from calling it "master," especially in the context of technology.
merely a quip :)

that's great, now give yourself an actual background in the subject first and then read that article again
not using MCTS would likely make the bot worse and further obfuscate our already limited understanding of the AI (thanks DCNN!)
there is no reason for it to seem more interesting by using just nerual networks

i'm sure it would be worse. alphaGO was intended to be a high-performance bot, and that's what it is. but i am not interested in CNNs, nor MCTS

It could techinically work on a neural network. The netwrok would need to learn how to do a montecarlo search without that being hardcoded in. But our brain is a neural network just with 100 billion nodes... .

Like if we want an AI to play real time video games, montecarlo wouldnt be an option, it would have to solely rely on neural networks. the reason montecarlo is works in go and chess is that time freezes and gives the bot time to think. but video games move at a minimum of 30 fps so we would need a MUCH bigger computer with way more processing power.

>our brain is a neural network just with 100 billion nodes...
the rest of your post is pretty okay but this memeing right here is triggering me

>catching up to an AI

no

>The netwrok would need to learn how to do a montecarlo search

no, it would not. MCTS simulates a huge number of games, but people do not do this while playing go. i doubt that it would learn any method resembling MCTS.

MCTS is kind of a hack desu

This is actually great!

I can now blame my 100+ losses on AlphaGo beating my rather than some AI that I could beat.

i've actually heard that the policy network by itself can select pretty good moves

if that's the case, then that's already an example of a pure-nn solution, even if it is just a CNN

>people do not do this while playing go
confirmed for never playing go at any respectable level

you play out thousands and thousands of games until the end every single turn?

what do you think happens when it takes a player 10 minutes to make a move? they just stare at the board? they try out the "obvious" local moves, far away ones, and then a smattering of semi-random ones in pursuit of a clever move, and this is to try and look many moves ahead. in a way, MCTS is very similar to how go is actually played, it's just that bots would have less guidance and would need to try out more moves at each level. it's not a "hack," and any neural network would likely have to adapt a way to check out certain unpredictable moves that the initial training data would not have recommended. if anything, neural networks are a hack since the resulting AI is essentially a black box

mcts plays games to the end many, many times. you might do a few very shallow searches when you play, but it's not the same.

>you might do a few very shallow searches when you play
i don't think shallow means what you think it means, but anyways:
>in a way, MCTS is very similar
i did not say it was the same, i said that it is very similar to what we already do, and calling it a "hack" just indicates an unfamiliarity with how the game is played

what's your rating, if you don't mind me asking?

it already does that.

wolfram alpha knows math better than you ever will.

If they dont change it or update it at all it will be easy to catch up.

If they do then probably only another computer can do it or a computer assisted human.

i'm 3k according to KGS

better than me then. if it weren't so late, i'd ask you for a game.

anyway, i still kind of disagree with what you're saying, but i'll leave it at that, i'm tired.

What do you mean 'came back'? Didn't it already demolish the best player in the world some time ago?

what is your rank, user? sleep well!

i'm just a double-digit kyu pleb on OGS.

>wolfram alpha
Nice try Stephen.
How come Wolfram Alpha didn't help you become a valuable mathematician?

>Monte Carlo Tree Search
>typical
don't forget, MCTS is a new technique only a decade old, compared to the venerable alpha-beta depth-first search that dominates the chess-like games.

That said, it would also be interesting to develop an AI framework that could discover new things in the search arena, rather than just refinements of the evaluation function. That would require artificial creativity.

(AlphaGo still boggles my mind. Before it hit, everyone thought top-level computer Go was decades away.)

>Master
Wow that's pretty racist.

they missed an opportunity to use "Sai"

bump

deeply concerned senpai

Yes. The MCTS is not required for the system to work. It just works much better with MCTS, which is why they use it.

AlphaGo smurfing in high elo challenger