Slashdot Mirror
Neural Network Chess Computer Abandons Brute Force For "Human" Approach
An anonymous reader writes: A new chess AI utilizes a neural network to approach the millions of possible moves in the game without just throwing compute cycles at the problem the way that most chess engines have done since Von Neumann. 'Giraffe' returns to the practical problems which defeated chess researchers who tried to create less 'systematic' opponents in the mid-1990s, and came up against the (still present) issues of latency and branch resolution in search. Invented by an MSc student at Imperial College London, Giraffe taught itself chess and reached FIDE International Master level on a modern mainstream PC within three days.
the big Computer tournaments are run by TCEC at chessdom.com - there it would be paired against other engines, of whom Komodo and Stockfish have been pretty much dominating every year since season 2 -
truth is, all computer chess is computer vs. computer nowadays - the losses come from different evaluations of positions - then the programmers try to correct it, etc - but since all engines are running the same hardware with resources, the best performers should win -
you can follow Season 8 (round 1b right now) here
http://tcec.chessdom.com/live....
----------
ah honey, we're all resplendent - Bill Mallonee
For comparison, GnuChess also plays at an International Master level. The article says this chess engine is much slower than GnuChess.
Humans are able to play chess at a high level because they are able to brutally prune the decision tree.....a grandmaster can quickly eliminate most moves as useless (although he/she will probably think of it in reverse terms: saying he/she quickly identified the important moves in the position). A computer that could combine that kind of pruning with the massive searching power would be ridiculously powerful. Better than our current computers by an order of magnitude.
"First they came for the slanderers and i said nothing."
Well, yes, that kind of is the issue. The computation chess masters make, the actual thoughts, could be handled on a 1950 computer no problem.
The question is how. It isn't brute force, though they do delve into plies ass desired. The real trick is knowing which handful to explore mentally. And if it were just pattern matching against known games, it would ne done by computer already that way, too.
(-1: Post disagrees with my already-settled worldview) is not a valid mod option.
The question is how. It isn't brute force, though they do delve into plies as desired.
As you mention, grandmasters know when it is appropriate to search through the move-tree, when to look deeply into a position. They prune the search tree very hard.......so the question is, how can we get computers to know which branches are ok to prune, and which aren't? Computers still aren't very good at that.
Incidentally, it amazes me how often Tal would say, "and in this position, I decided to calculate every variation all the way to mate." There are not many people who can keep the moves that clear in their head (and sometimes, the calculation was too deep, and he just guessed).
"First they came for the slanderers and i said nothing."
I'm old enough to remember when the MCP could only play chess!
Liberty - Security - Laziness - Pick any two.
It only plays at around the level of GnuChess, so don't be impressed.
You should be impressed. Not by it's level of play (which is not impressive), but by the fact that it:
1. Taught itself to play
2. Reached FIDE Master Level in THREE DAYS.
To be honest, I'm not even sure why this is a story.
See #1 and #2 above.
See subject.
Despite the fact that computers can now beat even the best human players at chess, I've always been of the opinion that beating a human at chess was not really a solved problem, because where chess programs do so by exhaustively examining millions of board combinations to make even a single move, a grand master chess player will generally contemplate but the tiniest fraction of that amount.... and they can still play chess pretty damn well. If a computer only considered as many board combinations as a grandmaster did, but still otherwise using the same chess playing algorithms as what are typically used today, even a rank amateur chess player could probably beat it.
To me, the problem of making an a chess playing algorithm that can beat a human being should really be figuring out exactly what it is that the best chess-playing humans do when they play that enables them to play as well as they do *WITHOUT* the ability to consider every combination.
Personally, I'm betting cracking that nut is more than halfway to achieving general AI.
File under 'M' for 'Manic ranting'
Are we sure it did not just learn how to install and launch GnuChess?
Violence is the last refuge of the incompetent. Polar Scope Align for iOS
What's the difference?
The rules were programmed in advance.
For example, you couldn't put it in front of scrabble and expect it to do something reasonable.
"First they came for the slanderers and i said nothing."
No, that is what existing high level chess programs do, and exactly what this one doesn't do. Please go and learn about what a neural network is before commenting on this story again.
I thought Claude Shannon https://en.wikipedia.org/wiki/... wrote the first program to play chess. Among other things he dabbled in.
Sigh. The reason that humans experts are no longer competitive is because human experts prune where Deep Ply fears to trust static analysis. Pitted against a relentless algorithm which resists intuitive pruning, grand-master human pruning leaks a full pawn or two per game.
It's damn amazing how well grand-master level pruning actually works, but don't mistake this for flawless chess. Beautiful? Maybe. Flawless? Not even close.
When it was still somewhat competitive between man and machine, the human chess players would think they were pressing an overwhelming advantage, only to discover themselves mired in tiny, unanticipated tactical disadvantages move after move after move after move. "The damn thing keeps finding these fiddling resources!" If you weren't careful, you could easily lose from what had initially appeared to be a won position (and it probably would have been, against a human opponent blind to all those fiddling resources).
The trick for the competitive chess programmer was to achieve the right balance in the static evaluator so that tangible material gains didn't consistently outweigh less tangible advantage of tempo. Matthew Lai in his paper does not seem to grasp this essential trajectory of computer chess. He seems to think it's remarkable that his Oldsmobile displays more rigidity on the impact sled than the lunar lander, when it's pretty clear to everyone else involved that no Oldsmobile ever made was going to win the space race. The ply-based chess engines had their static evaluators hand-tuned by experts over many decades within a space gram clock-cycle budget.
Until he actually defeats all these programs on existing commodity hardware at existing tournament time controls, he's comparing watermelons to kiln-dried coconut flakes.
It's the same problem with new technology. It isn't enough to merely be better in some personally favoured dimension of merit. Your immature new thing has to be better enough to actually pass the mature old thing on its own terms.
Got a better substrate than silicon? Yeah? What's your defect density cranking out 10,000 wafers per month? Oh, you haven't actually developed all that quality-control infrastructure yet, but you figure you can do it at half the price once you work out the final kink from your strained bullerene crystal lattice?
Awesome progress, pal, but I think I'll invest my own Bitcoin elsewhere.
For the record, I've long believed that the trade-off moving from depth to sophistication wouldn't prove particularly steep (for the right sophistication). But any gradient that's a net loss (no matter how small) provides pretty much no immediate competitive incentive for anyone to invest any real effort hoeing that row.
The great thing about neural networks is that they don't actually require much real effort. The machine itself does most of the work in 72 hours. And then what have you got? A RISC chip that never actually kills x86 (because those idiots were busy touting microcosmic instruction set efficiency long after the real game had shifted to streamlining the cache hierarchy, where's there's no low-hanging ideological shortcut to help you overcome the first-mover fat-payroll advantage).
I have seen something else under the sun: The race is not to the swift or the battle to the strong, nor does food come to the wise or wealth to the brilliant or favor to the learned; but sunk cost and legacy happen to them all.
Go ahead, give it the scrabble manual. See what it does. Give it a reading teacher, it won't matter. It is incapable of doing more than playing chess. That's what it was programmed to do.
"First they came for the slanderers and i said nothing."
Sigh. The reason that humans experts are no longer competitive is because human experts prune where Deep Ply fears to trust static analysis. Pitted against a relentless algorithm which resists intuitive pruning, grand-master human pruning leaks a full pawn or two per game.
lol yes, but that's why we consider computers stupid, even though they can still win at chess through brute force. The fact remains that the vast majority of chess moves in a given position are bad, and the computer program that learns to prune them first will have a huge advantage.
"First they came for the slanderers and i said nothing."
Giraffe's magic is supposedly in the decision-pruning algorithms. Surely some of the concepts involved would be transferable to other games, like Scrabble.
Maybe you won't be satisfied until an actual humanoid robot is moving pieces by itself, having bought the chess set from a local shop.
But Von Neumann was something special.
A polymath and a polyglot, his early work with chess is not to be scoffed at.
Happiness in intelligent people is the rarest thing I know.
Ernest Hemingway
Chess was AI until the computers started doing it well, then it became "not AI". So AI is defined as whatever humans do better than computers at the current time, a list which is getting smaller and smaller. I guess someday there will be nothing left.
This is the kind of argument you see from people who think a chatterbox thinks. There's been a clear understanding of the difference between "hard AI" and "weak AI" for decades. Chess playing computers are clearly weak AI. This isn't an insult, it's the way these things are categorized.
"First they came for the slanderers and i said nothing."
Maybe you won't be satisfied until an actual humanoid robot is moving pieces by itself, having bought the chess set from a local shop.
Mate, if you're going to say it can teach itself, then it better be able to actually teach itself. I have no objection to this chess program as a clear demonstration of weak AI.
"First they came for the slanderers and i said nothing."
What's the difference?
The rules were programmed in advance
A person who would be considered having taught themselves to play would still have to have access to the rules from somewhere.
Your comment makes it sound like you define "teach itself to play" as needing to reinvent the entire game independently
Actually you have roughly 20 million neural networks in the brain.
ph'nglui mglw'nafh Cthulhu R'lyeh wgah'nagl fhtagn
Talking about the phrase "Teach itself" is a mere semantic dispute. I would rather discuss what the AI actually does.
Very well said. I will think more deeply about that in my future conversations.
"First they came for the slanderers and i said nothing."
The point is not to win. Chess supercomputers already do that. The point is to write a program that can play well using limited resources, and maybe learn something about how humans do it.
You said yourself: "It's damn amazing how well grand-master level pruning actually works."
I suppose it would be more impressive if it learned how to play without knowing the rules. On the other hand, that's a little unfair, no?
Although, there are types of neural networks that have learned to play things like Mario Kart by watching human players play.
Oh, I see why people are confused with what I wrote. I meant, the rules for filling in the database were programmed in advance, the guidelines. The AI can't change its program.
"First they came for the slanderers and i said nothing."
It uses a neural network to recognize good moves. If you don't consider training a neural network "changing it's program" then I've got some bad news regarding your own autonomy.
Why are you so certain that a neural network matches a human brain?
"First they came for the slanderers and i said nothing."
Well, yes, that kind of is the issue. The computation chess masters make, the actual thoughts, could be handled on a 1950 computer no problem.
The question is how. It isn't brute force, though they do delve into plies as desired. The real trick is knowing which handful to explore mentally. And if it were just pattern matching against known games, it would be done by computer already that way, too.
What?
FTFY... (although perhaps a few players I know might be thinking about it the original way it was written)
Autonomous cars are not AI either.
When playing some of the better conventional chess engines I am astounded at how often i get out past fifty moves and discover that I am in a really fatal position. Trading teat for tat with the pieces and having something that looks like a real chess game on my part does not equal being in a good position for the end game. I am impressed and grateful for the fine work that has been done on Crafty,Rebel and Sjeng.
You make some very important points in your post: for your new product to take over, it needs to do everything the old product does, and then do something better. However, take this into account:
1) The team that built Deep Blue were IBM employees, and had so they had different resources available. I doubt this student (I call him kid) had a grandmaster available to help him fine-tune his evaluator, or a fab to build custom silicon for his chess-playing machine. Also, it is very instructive to watch the documentary "Game Over" to learn a few things about how IBM used the game against Kasparov to push up their share price. That should gave some idea of the resources they have thrown at the project.
2) The same Deep Blue team were coming from the CS department at Carnegie-Mellon Univ. where they did their Ph.D. on computer chess, and studied with a prof that spent a lot of his career on this subject. They were grown-ups with a lot of experience in the field, and much wiser than a young student.
3) The current computer chess champion (Komodo) again had its evaluator fine-tuned by a grandmaster: https://en.wikipedia.org/wiki/...
4) Most of the top chess programs have been written by programmers that have written other chess engines before. Their "success" is their 3rd of 4th re-write of a chess engine, and no amount of talent can replace that kind of experience.
Given all these points (and a lot more that can be identified along the same lines) I would say this kid did a good job.
"Models" not "matches", and the results (sometimes literally) speak for themselves.
And did you exchange a walk on part in the war for a lead role in a cage? - Pink Floyd.
There are a few blocks with "input" and "hidden layer 1"/ hidden layer 2. What does that mean? Absolutely nothing
At some point you have to stop explaining subject specific phrases in an article, "hidden layer" means something something to people who have a basic understanding of the subject, google it if you don't.
And did you exchange a walk on part in the war for a lead role in a cage? - Pink Floyd.
20 million neural networks in my brain... all interconnected...
It means I have a neural internet in my brain! That's fucking cool!
To some extent neural nets do model what happens in a human brain, but they also do things that we're fairly sure human neurons dont, most notable being back propagation, or at least not in the format we do it with neural networks. Thats not to say there are analogous mechanisms, in fact there *must* be one (how else to explain the elasticity of inputs). But there are critical differences.
Now that doesnt mean of course that a computer neural network is stupider. In fact cell for cell our neural networks out perform the shit out of biological neurons , its just the brains have so much more , both in terms of mechanisms and sheer neuron count + connectivity.
Excuse the Unicode crap in my posts. That's an apostrophe, and slashdot is busted.
Talking about the phrase "Teach itself" is a mere semantic dispute. I would rather discuss what the AI actually does.
The more important semantic dispute is whether you should throw ever use the unqualified phrases "AI" or "Artificial Intelligence" about a limited computer program.
I can see that "weak AI" is acceptable as a technical term, as it is clearly differentiated from anything to do with intelligence as generally understood.
To have a right to do a thing is not at all the same as to be right in doing it
This is weak AI, not strong AI.
"First they came for the slanderers and i said nothing."
Neural networks are anything but resource-efficient.
Compared to how a tradition chess engine searches the game tree, neural networks are extremely resource efficient.
The top engines use an algorithm called PVS() (Principle Variation Search), which is just a variant of AlphaBeta() that includes a null-window aspiration search. They do some advanced stuff such as pruning and extensions, but in the end the core of it is still AlphaBeta(). These engines still have to search through literally billions of positions on each move in order to beat the top humans.
Searching through and evaluating billions of positions vs pushing values through a neural network. One of these is inefficient, yes. You've picked the wrong one.
"His name was James Damore."
Why are you so certain that a neural network matches a human brain?
Maybe because a neural network, for a fact, matches the human brain.
This is well understood. What isnt well understood is the learning mechanism, but we do know for certain that it is to a large degree a timing-dependent hebbian learning process. To be quite specific, Hebb's Rule is a good predictor of neuron activation in brains. People did fucking science.
At least become casually acquainted with the subject before acting like a know-it-all. Clearly you don't know shit. What possesses people like you to act like you fucking know something when you know for certain that you don't is beyond me. Come on guy... you know you are wholly ignorant on the subject, so why are you acting like some fucking knowledgeable person about it? You do know that its wrong to do that, right? Its not just wrong, its dishonest. That makes you a dishonest fuck.
"His name was James Damore."
Thats not to say there are analogous mechanisms, in fact there *must* be one (how else to explain the elasticity of inputs).
The neurons in a brain to a large extent use some form of hebbian learning. We know this for a fact because Hebb's Rule is proven to be a good predictor of neuron activations.
"His name was James Damore."
It doesn't even "recognize" good moves. It used Stockfish's evaluation algorithm to build up its own database. It did not "learn" anything new, because Stockfish still runs circles around it.
Support microSD: in a post 9/11 world, it is unwise to carry your data on media that you cannot comfortably swallow.
Come on guy... you know you are wholly ignorant on the subject, so why are you acting like some fucking knowledgeable person about it? You do know that its wrong to do that, right? Its not just wrong, its dishonest. That makes you a dishonest fuck.
Oooh, insults, you sound so intelligent when you insult me.
Maybe because a neural network, for a fact, matches the human brain. This is well understood.
No it's not lol. They match some aspects of neurons, but not all of them. We don't even entirely understand what neurons do. It's unlikely we even know all the different types of neurons that exist.
Getting back to the chess-playing neural network in this story..........it is a specific, chess-playing neural network. As a result, it clearly belongs in the subset of weak-AI. Neural networks in general may match a human brain (something we don't know, which you so elegantly try to cover up with insults), but this particular one clearly doesn't.
"First they came for the slanderers and i said nothing."
You are just an ignorant fuck that likes to pretend that he knows something, even when you know exactly zero.
It's unlikely we even know all the different types of neurons that exist.
There are 8 types you ignorant fuck.
"His name was James Damore."
in your earlier post, you stated quite clearly that the learning process isn't understood. There's plenty more about the brain that's not understood. If you'd like an introduction to the topic, I can give you some book recommendations, but your rage is rather entertaining.
Shall I try to enrage you some more? Did you know there are over 50 types of neuron in the retina alone?
"First they came for the slanderers and i said nothing."