Go game is nice and interesting, but AlphaGo made it a bit irrelevant. What will be the next milestone to target in AI? Is there anything more complex? (I mean something with full information and zero random chance.)
markets
>>7932449
so you mean telling the future?
>>7932450
yeah making better predictions based on models
i mean if i had deepmind thats the shit i'd be doing
6 dimensional tic-tac-toe
>>7932444
They're tackling starcraft next. After that the only more complex thing is the real world.
>>7932457
I only see three levels there
POMDPs
O
M
D
P
S
>>7932444
Take your shitty go meme and fuck off. Everyone knows Smarties don't even make black and white ones. Mongole.
>>7932464
Beating a human at starcraft would be incredibly easy due to the physical limitations of the human
You could just give the AI a rush tactic, even a worker rush, and with perfect unit control it would win easily.
>>7932444
Rock, Paper, Scissors.
>>7932444
Proving mathematical problem.
>>7932444
Go is pretty the pinnacle of perfect information games, at least in terms of strategic complexity.
The next step IMO would be imperfect information/probabilistic games. I think I read in an interview somewhere that Hassabis was interested in doing something with no limit hold'em.
>>7932540
https://en.wikipedia.org/wiki/Automated_theorem_proving
>>7932527
They won't be giving the AI anything.
The entire point is that the AI will learn to play from scratch. It's actually an extremely hard problem.
>>7932444
Financial markets
The philosophy and strategy of war
World domination.
>>7932542
go is less than tic tac toe compared to league of legends
Just automate my philosophy up senpai.
>>7932612
War is the ultimate game.
>>7932444
composing
>>7932454
>If I was part of a company with a literal monopoly then I'd use my AI to predict the market... that my company already controls.
10/10 reasoning, would give my PhD.
>>7932527
>Give the AI a specific strategy so that human players only have to counter it
10/10 reasoning, would give my PhD.
I bet google just wants to become machines and enslave everyone in a transhumanist "brain net" (super hive mind, facebook and myspace in your brain) while their skynet ai and machine army do all the work and keep everybody in line.
>>7932653
Well, it's Starcraft. Micro gives an advantage. With perfect, superhuman micro, you don't need to be unpredictable.
Computer-controlled Starcraft has already been done, and fucks human players over with tactics a human just couldn't use, like individual mutalisk micro, or marines that just can't be hit by lurkers.
>>7932444
>AlphaGo made it a bit irrelevant
You're irrelevant.
>What will be the next milestone to target in AI?
"World In Flames"
>>7932449
>>7932454
Already a thing. Live human trading is an extremely small percentage of the market trading.
>>7932705
So give the AI input lag comparable to human reflex time, and give the neural network an attention system so it can only view a comparable amount of available information at a time.
>>7932616
Learn to read retard
>>7932734
>ai already solved it
>so make shittier AI
What
>>7932589
automated theorem proving AI still sucks, you dumb ignorant
>>7932764
Something like that would be interesting.
Make an AI that can defeat professional human players but loses to Alpha Go.
Right now as it stands, specially after Lee Sedol's raping hand in game 4, Alpha Go just seems to be a dumb computer that crunches a bunch of data and then calculates a bunch of probabilities. Totally dumb.
If you'd make a human like AI that defeated humans but was totally inferior to Alpha Go that woudl show that Alpha Go is actually skynet.
>>7932444
Basketball or MMA.
>>7932787
>my AI is too strong
>let's make a slightly worse one so it's more fun :^)
10/10 reasoning, would give my PhD
>>7932734
Starcraft is a really terrible RTS to test it with to be honest. Mostly based on the speed at which you can amass a blob of units and overtake the enemy blob of units. A better RTS for this kind of system would be something like Supreme Commander, where battles are decided by intelligence gathering and employed strategies on a large scale.
>>7932805
That's Starcraft 2 by the way. Brood War doesn't have all the same issues, but it still isn't ideal.
>>7932794
This guy said it, the next challenge is a set of neural nets that can handle a robotic body and play sports to a human level.
Like it or not, it seems that for neural nets it is easy to optimise for simple games. Guess what science is a simple set of rules, I'm sure neural nets can play that with large datasets, no problem and better than 99.9999% of men.
Shogi
It's better than chess according to an IM.
>>7932764
The point is, beating humans at Starcraft unrestricted is not any kind of expression of intelligence; you just have to act faster than the humans can react. In order for this to actually be an interesting strategic problem, you have to act under the same physical limitations as the opponent.
>>7932924
>it's not suitable for AI
>so let's try to hack it into a real problem and then solve that one :^)
10/10 reasoning, would give my PhD
>>7932923
Same shit, different name.
>>7932650
Well if you're drawing on large data sets for models it helps if you'e running the company everyone gets their information from
I mean imagine how many data points Google can tap, no one else can compete, you're right
>>7932444
A game in which you are given a computer algorithm executable by a Turing machine and must determine if it terminates and at what point it will terminate.
>>7933134
:^)
>>7932454
Markets are different, because players have limited information and there are random factors... But I would be surprised if there was not any intention to make money with this AI. I would used it on markets too... I can't wait to see AlphaStock. :)
The Pro GO players should *just* switch over to 25x25 boards and be able to beat the AI again.
>>7933365
kek. i like you
>>7933370
There are also some 37x37 GO online communities.
That would certainly be enough to totally cripple the AI
>>7933365
What about a norvig x norvig board
Probably unbeatable by a computer
>>7932805
What about company of heroes?
THAT would be something.
memes
Literally all objections and "problems" with Starcraft that people here like to point out can easily be solved by limiting the AI APM to 100 or 200.
>>7932663
wud gib mai phd
What about 37x37x37 cubic go? Does anyone have an idea how hard this would be for an AI?
>>7932444
Nice trips. Checked
>>7932805
Why not EVE?
>>7934061
i don't think you would get enough captures to make it playable
>>7932449
>something with full information
>>7934403
you don't have to deal with the whole world
you can have a model just look at stock prices, play and adpat
>>7932467
it is 6d, pic related
I didn't highlight 1d, but it's a row of three cells, if you get three in a row, you win, but that's impossible
2d: a 3x3 board, this is highlighted in red and is a normal tic-tac-toe board, and if you get a
3d: highlighted in pink, if you see this in real life, you'll probably see the three boards stacked on top of eachother. if you get 3 in a row going in the x,y, or z direction, you win.
from here on it gets hard to visualize, but here we go
4d, highlighted in orange, impossible to actually represent, but if you get 3 in a row in the x,y,z, or k directions (k being the direction perpendicular to all three previous directions) you win.
I don't know the standard variables used for axes beyond this point
5d, highlighted in yellow, axes are x,y,z,k,i(?) i is perpendicular to all previous axes, three in a row in any of the 5 directions wins
6d is highlighted in green, we'll call this axis j, and it's perpendicular to all previous axes, three in a row in any of the six directions wins.
and that is 6d tic-tac-toe
It sounds shit, honestly
>>7932444
I hear it's starcraft.
I'd prefer civ or something, desu
>>7934493
>2d: a 3x3 board, this is highlighted in red and is a normal tic-tac-toe board, and if you get 3 in a row in either direction you win
fixed
>>7934493
Is that even a hard problem? For example, isn't 3D Tic-Tac-Toe easily solved?
>>7934527
I doubt it would be difficult.
It's probably even trivially solvable by just looking through the game tree by hand.
>>7932590
i dont think you understand how this works, finding a way to counter strategies in starcraft is way easier than in go, that and the immediate reactions of a computer will crush a human easily...
I don't think having a great go AI makes game irrelevant.
As someone said in a same thread about chess: Any car wins Usain Bolt in speed. That doesn't make olympics meaningless.
>>7932449
> implying so fucking hard that financial markets are an adapted process
why do you even breath oxygen
>>7934410
>How would this affect things?
The market would become fully efficient and the prices of securities will match their true values. It is hardly a world ending scenario.
Unfortunately a trading algorithm that has "solved the market" will reduce any arbitrage opportunities in the market hence rendering itself obsolete (i.e. it will slowly drive prices towards the true value. Once it has reached this value there is no longer a price discrepancy and an opportunity for abnormal market returns)
>>7934412
>you can have a model just look at stock prices, play and adpat
People spend their entire career in the finance world trying to determine the inputs for such models. It is by no means an easy feat and is more of an art than a science. It certainly doesn't help that the inputs are continually changing. Obtaining "full information" for an insider within a firm is extremely difficult let alone for an outside trader.
