Giraffe, the little chess engine that ...

September 14, 2015 3:42 PM

Deep Learning Machine Teaches Itself Chess in 72 Hours, Plays at International Master Level

Giraffe: Using Deep Reinforcement Learning to Play Chess (the study in pdf form)

posted by phoque (45 comments total)

on september 14th at 6:42 PM, Chessnet became self aware
posted by feckless fecal fear mongering at 3:44 PM on September 14, 2015 [18 favorites]

Yes, I'm old. Old enough to remember the MCP when he was just a chess program!
posted by fearfulsymmetry at 3:52 PM on September 14, 2015 [12 favorites]

Makes sense. When you come down to it, Giraffe is a far more chilling name than Skynet. And I bet it will have a cute logo.
posted by rtimmel at 4:02 PM on September 14, 2015

And drawing psychedelic fish-dogs!

I bet if you feed my chess moves to this thing while it runs backwards, it will even let out a similar-sounding whimper to the one I usually make when I get beaten.
posted by pulposus at 4:04 PM on September 14, 2015 [1 favorite]

This seems very useful for creating AI opponents for more than just chess; imagine being able to feed the rules of an arbitrary strategic board game (like any kind of fairy chess variant) to Giraffe and get out a competent engine for playing it. Now if we had some way to measure, based on how this engine performs, various qualities of the game itself, we could have another program that designs the board games, searching for new "verified interesting" chess variants.
posted by wanderingmind at 4:10 PM on September 14, 2015 [2 favorites]

imagine being able to feed the rules of an arbitrary strategic board game

e.g. global thermonuclear war
posted by thedaniel at 4:12 PM on September 14, 2015 [8 favorites]

In a world first, an artificial intelligence machine plays chess by evaluating the board rather than using brute force to work out every possible move.

This sentence does not give me any confidence in the author's understanding of how traditional chess engines work.
posted by Wolfdog at 4:21 PM on September 14, 2015 [19 favorites]

This result seems pretty big to me. Is it big? I have to think people have tried neural networks on chess many, many times with less success. What made this one work? The MIT Tech Review article implies it was his choice of a training set. Guess I should read the paper to really know.

It's amazing how quickly statistical machine learning is taking over the world. It feels like in the last five years things have really turned a corner, with results as mundane as better search suggestions to serious AI problems like Siri's speech interface or Google's self driving car. I'm just wrapping up taking a Coursera course in machine learning just because it feels like a tool I should have in my belt when approaching problems. Between the better packaging of well known algorithms and the ease of doing data crunching on terabytes of data, it feels like we're at an inflection point.
posted by Nelson at 4:24 PM on September 14, 2015 [3 favorites]

"i, for one, welcome....etc. etc."
posted by salishsea at 4:26 PM on September 14, 2015 [4 favorites]

I suggest somebody teach it tic-tac-toe, before it's too late.
posted by Professor Falken at 4:28 PM on September 14, 2015 [14 favorites]

Reading the paper, it looks like rather than simply stopping the search at a specific depth, they stop when a move is not sufficiently probable. Farther-out moves are naturally less probable, but they also take into account their neural net. I guess this is what is meant by "not just brute force".

My understanding is that we've recently gotten really good (because of theoretical as well as hardware improvements) at training deep neural nets, so now you can feasibly use very deep nets on very large training sets. E.g. rather than using standard (stochastic) gradient descent to optimize the weights, this paper found that a technique called AdaDelta was much more effective, and it was only published in 2012. That kind of advance, plus cheap fast GPUs, is why they're suddenly becoming state-of-the-art at every task (and finally living up to the 90s hype!).
posted by vogon_poet at 4:29 PM on September 14, 2015 [4 favorites]

So, traditional chess engines like Deep Blue weren't that great at pruning their search trees, but they made up for it by exploring the tree both deeply and broadly. ("While Deep Blue was searching some 200 million positions per second, Kasparov was probably searching no more than five a second.") Deep learning neural networks are a different paradigm that can afford to explore much less of the search tree because they're better at evaluating positions and only exploring the best ones—but they have to be trained on huge and carefully selected datasets of millions of positions. This is impressive, but it still doesn't match whatever humans are doing. Kasparov did not play millions of games of chess to reach his skill level. I suspect that we're at least one more paradigm shift away from computer programs that not only match human performance, but do so using the same methods.
posted by Rangi at 4:44 PM on September 14, 2015 [9 favorites]

AFAIK the same techniques are being used to play Go, albeit with zero lookahead. Lots of success has been had using a tree search and training the eval function on local patterns. Though I wish this article would just tell us how what ply on average the chess engine descends to.

Also, someone on the chessprogramming wiki is a bit of a comedian, as they linked to Giraffe - Internal Systems
posted by RobotVoodooPower at 4:50 PM on September 14, 2015 [1 favorite]

Also, I think training on a massive set of real chess games is a *bit* of a cheat. It's like, *all* the apriori knowledge. But ya know, it's not how you play the game, it's whether you win ...
posted by RobotVoodooPower at 4:54 PM on September 14, 2015

I loathe how chess is constantly trotted out as a metaphor for every activity signaling intelligence. This machine could use a moral teacher. "The lesson is, if anyone who looks on to the world as if it is a game of chess, deserves to lose."
posted by Apocryphon at 5:22 PM on September 14, 2015 [3 favorites]

Straight out of the box, the new machine plays at the same level as the best conventional chess engines, many of which have been fine-tuned over many years. On a human level, it is equivalent to FIDE International Master status, placing it within the top 2.2 percent of tournament chess players.

Something's not right here. The best conventional chess engines have an effective Elo rating well over 3000, significantly better than the best humans in the world. An International Master has a rating of about 2400. The rating scale is logarithmic, with each difference of 200 points corresponding to about a 3:1 ratio of results.
posted by dfan at 5:25 PM on September 14, 2015 [6 favorites]

I'll be worried when it masters Calvinball.
posted by gottabefunky at 5:40 PM on September 14, 2015 [5 favorites]

But can it play Go at the dan level?
posted by redbeard at 5:42 PM on September 14, 2015

A neural network that plays Go at the dan level
posted by dfan at 5:49 PM on September 14, 2015 [2 favorites]

Pfft. A decade ago, I had a little robot/program that would enter a MMORPG, teleport to an area, fight off various minions, find its way through a small maze, then kill a boss character, grab any loot (good drops from boss characters), teleport back to my loot chest, deposit loot that had dropped, exit game, create new game, etc.

I would leave that running overnight and have cool gear the next morning.

I was much more impressed with that robot that chess playing bots. Prolly not as many variables as a chess game, but much more AI involved.
posted by CrowGoat at 5:53 PM on September 14, 2015 [1 favorite]

Chessnet became self aware

The survivors called it Checkmate Day
posted by nubs at 5:57 PM on September 14, 2015 [5 favorites]

Whoa! This is actually my friend's thesis. Sorta surreal to see it on mefi, though come to think of it he would probably find the metafilter crowd quite interesting..
posted by btfreek at 6:25 PM on September 14, 2015 [2 favorites]

In the year one million and a half, humankind is enslaved by giraffe.
posted by George ROFLington at 6:29 PM on September 14, 2015 [2 favorites]

Yeah, well I've got a few billion years of evolution under my belt. My neural net may comparatively slow in terms of raw speed but I am optimized like crazy. For chess? No. For the survival of my genes.

And if all else fails, even me and bacteria share a good percentage in common.

Give it a few more years as neuroscience advances and more of the basic structures that make the human mind so great become part of computing. At point it will be a human mind modeled and enhanced and it will be unstoppable.

Butlerian jihad to follow. Then the kwisatz haderach. It always comes back to Dune these days.
posted by thebestusernameever at 6:37 PM on September 14, 2015 [1 favorite]

I wonder if it's named Giraffe because it's preposterously specialized.
posted by Wolfdog at 6:39 PM on September 14, 2015 [3 favorites]

Greetings Professor Falken. Kudos for spending $5 to make a single comment.

I look forward to more wisdom from 1983 Puget Sound.
posted by pjenks at 6:42 PM on September 14, 2015 [1 favorite]

"John! Good to see you. I see the wife still picks your ties. "
posted by Chrysostom at 7:22 PM on September 14, 2015

Well...I've been wondering for a while now what's gonna wipe out the human race, Skynet, or apocalyptic environmental collapse. Looks like (yay?) Terminators for everybody!
posted by sexyrobot at 7:25 PM on September 14, 2015

Apocryphon: I loathe how chess is constantly trotted out as a metaphor for every activity signaling intelligence. This machine could use a moral teacher. "The lesson is, if anyone who looks on to the world as if it is a game of chess, deserves to lose."

"Moral teacher" and "AI self-teaching" are fairly orthogonal ideas. It's like you're complaining the sushi-building robot can't carry a tune.

And chess is trotted out as a familiar, brain-intensive and complex activity. The researchers could talk about how many pattern recognition rules the computer develops per millisecond, or some other direct metric, but that would make for a fairly meaningless article for those of us not highly familiar with the topic.
posted by IAmBroom at 9:08 PM on September 14, 2015

Someone, somewhere is priming their startups pitch for funding to sell this as middleware to game studios.

And yea, that's a multimillion dollar company idea right there. Especially for stuff like RTSes or better yet MOBAs as a solid training tool for players trying to get competitive.
posted by emptythought at 10:36 PM on September 14, 2015

Chess is so cliche, play poker or backgammon instead uggggggg
posted by Apocryphon at 10:55 PM on September 14, 2015

Or cribbage! If there's a cribbage bot out there that knows how to curse or crow a tenth as well as my father, I haven't found it.
posted by Earthtopus at 11:49 PM on September 14, 2015 [1 favorite]

Earthtopus, your post is the only place on the internet (according to a quick web search - mefi gets indexed pretty often, so you already show up) that uses those terms in the context of cribbage. Please to expand :)
posted by pharm at 1:20 AM on September 15, 2015

I should be very surprised if no one has ever mentioned cursing in the context of cribbage.
posted by Wolfdog at 2:53 AM on September 15, 2015 [1 favorite]

Specifically, the search 'cribbage "throw a tenth"' finds one result on duckduckgo, ie this page & none at all on Google, which is forced to fall back on invented searches that I might have meant instead. As an occasional cribbage player I am therefore intrigued to know what the term means to Earthtopus!
posted by pharm at 3:08 AM on September 15, 2015

crow means to brag or talk up one's victory, and Earthtopus would be surprised if a computer could do that a tenth as well as his father does.
posted by Wolfdog at 3:52 AM on September 15, 2015 [2 favorites]

Holy mis-parsing Batman!
posted by pharm at 4:13 AM on September 15, 2015 [1 favorite]

Indeed, my intent was (curse or crow) (a tenth as well), but clearly now I have to work 'crow a tenth' into my next game *somehow*... I just need to decide what it means first.
posted by Earthtopus at 5:37 AM on September 15, 2015 [1 favorite]

Chess is so cliche, play poker or backgammon instead uggggggg

The best backgammon programs outpaced the best humans a while ago.

Poker is an active area of current AI research, more so than chess.
posted by dfan at 7:36 AM on September 15, 2015 [2 favorites]

> Yes, I'm old. Old enough to remember the MCP when he was just a chess program!

I'm so old I have no idea what you mean by MCP; to me it means "male chauvinist pig." Care to elaborate?
posted by languagehat at 8:02 AM on September 15, 2015

languagehat: "I'm so old I have no idea what you mean by MCP"

I assumed this was a reference to the Master Control Program from Tron.
posted by mhum at 8:17 AM on September 15, 2015

Thanks! I even saw Tron, but, well, that was in 1982.
posted by languagehat at 9:06 AM on September 15, 2015

Surely, it was Cindy Morgan's finest role, aside from Caddyshack.
posted by Chrysostom at 9:15 AM on September 15, 2015

Poker is an active area of current AI research, more so than chess.

I would think that the boom in online poker is related to it. Don't you love it with pop culture and entertainment money makes certain avenues of scientific inquiry more lucrative? Either way, I'm glad, because chess is so overused in culture. Have AI play xiangqi.
posted by Apocryphon at 9:49 AM on September 15, 2015

I would think that the boom in online poker is related to it. Don't you love it with pop culture and entertainment money makes certain avenues of scientific inquiry more lucrative?

Probably it is fun for AI researchers who are poker players to be researching a game they enjoy, but more importantly, poker is a really interesting game from a mathematical viewpoint and is harder than chess, at least as far as beating the best human players goes. Also, making optimal decisions in the face of limited information is applicable to a lot of (more) real-life problems. People who are interested in the mathematics of poker might be interested in, well, The Mathematics of Poker.
posted by dfan at 4:28 PM on September 15, 2015