Google DeepMind Challenge Match


Caine

Hardcore Member
Joined
Jun 5, 2008
Messages
4,138
Location
Netherlands
Anyone following the ongoing Go match between AlphaGo (AI) and Lee Sedol (9-dan pro)?

Video link : https://www.youtube.com/channel/UCP7jMXSY2xbc3KCAE0MHQ-A/videos

I am not a Go player myself, so it is hard to fully appreciate the match in all its glory. That said, AlphaGo is performing very impressive in this matchup. Go is a game with an extremely high number of possible board configurations.

The AI is utilizing a neural network for determining hot-spots of where to play, a board valuation routine and a tree searching algorithm to determine its moves.
 

rygD

Nihilistic Mystic
Joined
Feb 28, 2014
Messages
7,464
Location
Everywhere
Ah, interesting. I will try to check it out in a bit. I like Go, but I suck at it.
 

Caine

Hardcore Member
Joined
Jun 5, 2008
Messages
4,138
Location
Netherlands
I am in no way an expert in neural networks nor artificial intelligence, but let's take a brief look at why this outcome is significant.

Traditionally, computer "intelligence" was realized by brute-forcing problems. I.e. evaluate all possible combinations and choose the optimal one from there.

Naturally this does not scale at all to larger problems. Not even the exponential growth in computer speed is sufficient to cope with the state explosion in many interesting games or problems. This is also why AI is not commonly seen as very intelligent.

The reason why Go in particular is considered to be a holy grail of artificial intelligence is because it has an absolutely massive amount of possible board states. I.e. on the 19x19 board each grid can be either black, white or empty and thus there are about 3^(19x19) possible board setups (which is a number with over 170 zeroes). The sun will probably explode before your computer has finished calculating all of those combinations.

The naive approach to resolve such problems is to use a branch-and-bound strategy (usually heuristics based) to cut off several paths and reduce the state which has to be searched. Again Go poses a serious challenge. It is very tricky to define proper heuristics as a stone can go practically anywhere and still be a valid move.

AlphaGo uses a neural network, which is a simplified simulation of what happens in a human brain. These networks are trained on problems and subsequently learn in a relatively "natural" way. Or to simplify matters, it just played a shitload of matches.

It takes humans about a decade to become really good at Go. AlphaGo managed to defeat a top-level professional human Go player after half a year of training.

In terms of breakthroughs in artificial intelligence, this match is writing history.

Or in layman's terms, DeepMind is skynet :p
 

_jr_

Advanced Member
Joined
May 5, 2013
Messages
1,170
I don't think it is adequate to claim that AlphaGo is not brute-forcing, just because the heuristic happens to be based on a neural net with some learning capabilities. From the available information it is still using vast quantities of computing power to evaluate individual positions (and there is nothing wrong with that, because the human opponent is doing it, too). Besides, these 'holy grail of whatever' claims are annoying in any context. It was always clear that a brain is some form of computing device and replicating it's features 'just' requires adequate algorithms (and there is no obvious indication that AlphaGo's is a breakthrough in that regard, though it may be).
 

levi

Still fresh, damnit!
Joined
Oct 6, 2008
Messages
12,993
Location
Somewhere off the coast of the EU
The approach of combining a traditional (as established at least by Big Blue the chess computer) simulated lookahead with pattern matching of previously observed matches and a neural net (I'm not entirely sure what the neural net does in this project) is novel though, and apparently now worthwhile.

I don't have a problem with the 'holy grail' description, as long as people realise there will always be a series of holy grails (and only one of them is guarded by the Killer Rabbit of Caerbannog). I think it's a useful way of measuring this sort of progress in a media-friendly fashion.

My only concern with this as usual is who's funding it. While I respect Demis Hassabis for many things, it's Google money that paid for AlphaGo. Run a similar algorithm over gmail archives, and who knows what they could discover about us all?
 

Caine

Hardcore Member
Joined
Jun 5, 2008
Messages
4,138
Location
Netherlands
I don't think it is adequate to claim that AlphaGo is not brute-forcing, just because the heuristic happens to be based on a neural net with some learning capabilities.
I suppose that is a true statement. Yet there is a significant difference between human-picked heuristics and heuristics learned by training a neural net.

Besides, these 'holy grail of whatever' claims are annoying in any context.
Heh, yeah. It's an over-simplification. Perhaps I should leave over-simplifying technology to the mass media, they are better at it :p

It was always clear that a brain is some form of computing device and replicating it's features 'just' requires adequate algorithms (and there is no obvious indication that AlphaGo's is a breakthrough in that regard, though it may be).
True, and neither are neural networks anything new. That said, it is starting to look like neural networks are slowly crawling out of the Trough of Disillusionment in the Hype Cycle.
 

Mr_Loon

Can't Remember
Joined
Aug 30, 2010
Messages
2,329
Besides, these 'holy grail of whatever' claims are annoying in any context.
Is that because you've already got one?

There is some justification in this context as Go is seen as the most difficult of board games for an AI to beat an expert human player, still think I'd be in with a chance for snakes & ladders though. ;)

The next holy grail for AI should be cards against humanity.
 

ZXDunny

Deep avatar
Joined
Oct 12, 2010
Messages
2,562
It appears that the neural network was trained by showing AlphaGo a shedload of real human vs human matches, so it had an idea of what a human would do in certain board configurations, and then they set it up to play against itself, which further refined and reinforced certain behaviours.

The brute force comes when it sits there looking at the board and decides what move the human is likely to make, and what the possible consequences are - then it askes the neural net "what would a human do in this situation?" and uses that "advice" to determine the optimal next move.

So the more matches it plays, the better it gets. The look-ahead tree "brute force" is very limited in nature and the real "genius" of this system is the neural net that has been trained to mimic human player in a natural manner.

D.
 
Top