Giraffe Becomes an IM in 72 Hours !

The_Ghostess_Lola

It appears as though the brute force computer has left the chess parlor. Would someone pleez turn off its lights ?

Questions....comments....concerns ?

The_Ghostess_Lola

An MIT student did it. It's an amazing story ! Here is a partial write-up....Smile....

****

A student has built a computer that managed to teach itself chess, and in just three days attained a level of skill comparable with some of the world’s most skilful players.

In the last two decades computers have become exponentially more powerful, but the “brute force” method of evaluating 200 million moves per second that Deep Blue employed was still the norm. Matthew Lai instead trained a “neural network” using example situations from real games of chess.

Rather than searching through all possible future moves and calculating the best one, Giraffe was programmed to learn the game in a similar way to humans, at a greatly accelerated rate.

****

 

The_Ghostess_Lola

Here's more....

****

Giraffe apparently finds the best move from its top three move decisions for each round on 70% of moves. Lai found that after just 72 hours, Giraffe had the proficiency higher than about 98% of ranked human players.

****

When playing, Giraffe looks at the game in three ways, like a human might. First, it looks at whose turn it is and who has what pieces. Then it takes takes a holistic view of the entire board—what pieces are where. And finally it considers what moves each of its pieces can make. Based on the millions of moves it’s learned in the past (and whether they’ve had a positive, negative or neutral outcome), Giraffe will then make its move. With every new move, each new game, Giraffe learns a bit more and gets a bit better. Like humans do.

Right now, the system is not as fast as other engines, but Giraffe works smarter, not harder. Lai said he’s working to shorten the time it takes Giraffe to find the right move, and he’s confident he can get Giraffe up to Grand Master level pretty easily. “I still have many ideas that need to be explored,” he said.

watcha

A neural network can have gut feelings, like humans. The interesting thing about a gut feeling is that it has no horizon effect. All brute searches either have a horizon effect or go very deep at the cost of senseless pruning.

If neural networks can get better than brute searches, now, this is a revolutionary development.

MikeCrockett

Just a fancy version of the old matchbox tic tac toe engine published by scientific American back in the 70's.

mcmodern

how does it stack up against houdini komodo SF? 

Robert_New_Alekhine

@mcmodern, nowhere close yet. It's only an IM.

This is a great idea! But this engine seems to rely heavily on intuition...perhaps too heavily.

mcmodern

I guess long ways to go then, SF and Komodo are the class of engines these days.

MikeCrockett

neural nets deserve more attention and probably would be important for strategic planning. brute force ends up working out the specific tactical ideas but only after deep look ahead.

The_Ghostess_Lola

In (20) years, chess engines will be exclusively neural-driven and today's BF clumsy clanks will be somewhere in a comic-oddity museum next to that glass-brained robot Dr. Smith used to campily abuse.

kkl10

The_Ghostess_Lola, can you link the article from which you're quoting?

The_Ghostess_Lola

Sorry....umm, I made it all up....Embarassed....

(....just teezing kk....Smile....)

http://qz.com/502325/an-ai-computer-learned-how-to-beat-almost-anyone-at-chess-in-72-hours/

kkl10

Thank you, this is very interesting. Artificial neural networks are a huge step towards improving AI heuristics.

mcmodern
MikeCrockett wrote:

neural nets deserve more attention and probably would be important for strategic planning. brute force ends up working out the specific tactical ideas but only after deep look ahead.

They should try to work on go instead of chess which for all practical purpose no longer need any improvement for human use. Go is a totally different story, best engine cannot even compete against the weakest go pros.

Goram
WHat is PETA doing??
watcha

Interestingly enough if a neural network starts to perform better than humans in a field that cannot be approached through brute searches ( like Go ), this is a lesser sensation and achievement than when it starts outperforming engines in a field which was always considered to belong to brute search and calculation ( like chess ). For my part I'm more interested in neural chess engines than neural Go engines.

The_Ghostess_Lola
Goram wrote:
WHat is PETA doing??

I would think that they'd be somewhat concerned. Now they have a neural-driven giraffe that will eat the moves off of the top of candidate branches & trees....moves noone could even reach w/ a ladder....let alone a long orangy neck.

watcha

Humans are neural networks with biological limitations.

Neural networks are neural networks with technological limitations.

You cannot expect biological limitations to go away quickly, but nothing prevents technology from advancing and producing the same kind of neural network humans have, just much deeper and stronger.

It is kind of trivial that in activities that humans are good at neural networks will surpass humans. This is just the question of the strength of the network.

On the other hand it is not that trivial that neural networks just by deep learning can solve problems that belong to formalized calculations. Humans did this by building machines for the purpose ( for convenience we call them computers ). If neural networks can do this without building separate machines, just by the mecahnism of deep learning, this would be quite a sensation.

The_Ghostess_Lola

I would think much deeper & stronger 'cuz exhaustion isn't involved in decision making. Emotion kinda the same....altho' I'm not really sure about if controlled emotions over a chessboard is a good thing or a bad thing - yet.

The ability to harness electricity has really taken us off the hook !....yoohoo - what a party !!

Commander_Crunchy

Interesting, so Giraffe looks at a position and draws from a database of similar positions and selects the move with the most statistical probability of winning...it doesen't calculate at all?