|
Author
|
Topic: Kasparov's recipe
|
Danpech
Member
Member # 163
|
posted 03. April 2003 15:43
I posted this muse over at the www.arn.org forums, but, I figure there are plenty of people here who do not often visit those forums, so I am posting it here as well.
How is Garry Kasparov, the world's chess master, able to compete against awesome brute-force computing? (see http://www.arn.org/docs2/news/manversusmachine040103.htm ).
As math prodigies know, there is more than one way to skin a calculation.
The most elegant (i.e., most over-all efficient, in time, energy, etc.) computation is always used, no matter what moves the other player makes. But, there is no single most elegant computation. Rather, it shifts as the game proceeds, and depends on what move the other player makes. This is how the ultimate reprogrammable computer would operate.
How this elegance is maintained (people often fail in doing so, but brute-force, artificial computing cannot even get that far) is by the same means that a sentient being selects what random nature cannot: say, prime numbers.
Hmm... Prime numbers have an algorithm: 'Cannot be divided by any number but itself and one'. How could nature have come up with big bugs like us, who can think up rules like that (I'm trying to think naturalistically, here)? Could nature have evolved selection-rule-thinker-upers? But, how would one know, at each move in the game, as to what computation is most elegant, unless...
My train of thought runs out of steam.
Anyway, no brute-force computing can match the ultimate reprogammable computer. The one is, by definition, ultimately superior to the other. Perhaps only biology will do. Perhaps biology is the most elegant substrate. Surely, there cannot be two equally elegant substrates? (There goes 'Strong' AI's multiple realizability.) A car is 'faster and better' than human bipedal locomotion, but a car cannot invent the car. The most elegant machine of all is the machine that can invent (at least mentally) anything that is invent-able. Can such a machine improve on its own production? I don't think so, and here's why: All-purpose elegance is irreducibly complex.
Perhaps the first time such a machine was produced, that was the only time it could be produced (if it was the product of non-sentient natural stuff---that stuff which that very machine can make sense of). How, then, could it have been the product of RM&NS? No chance could have been involved, otherwise it would not be the most elegant machine. (All this seems to me to imply that perfect human physical and mental health is IC, so that once it is corrupted, humans cannot engineer a way to regain it.) If you are not the foundation of reality, then why would you think that you can comprehend the foundation of reality? To hope to prove wrong the implicit claim of theism that it is ultimately incomprensible to you?
If, once you find yourself stood upright, you cannot remain standing, then you must either find a seat or eventually fall down. Some think it proud presumption to believe that we are made by a God. But, proud presumption is not native to such a belief, while it is native to the belief that we can engineer a machine that is better than us (which essentially includes the belief that we can engineer ourselves into something better than what we are). Just because we can conceive of something that is logically better than us (entirely better, not better in just one aspect) does not mean that we can produce it. After all, the foundation of reality produced us.
[ 03. April 2003, 16:06: Message edited by: Danpech ]
IP: Logged
|
|
Micah Sparacio
Member
Member # 6
|
posted 03. April 2003 16:33
quote:
Perhaps biology is the most elegant substrate. Surely, there cannot be two equally elegant substrates?
Danpech, I agree with the first statement ("perhaps" indeed) but disagree with the second. Two chairs can be made out of wood and metal respectively and both serve the function of being a chair just fine (aeshetic tastes aside for the moment). So I don't think that you've destroyed the multiple realization argument by any means.
Regarding your overall post, I'm getting several vibes regarding your main thesis. On the one hand, you've got an argument about how reprogrammable machines would be the ultimate chess players. I'm not sure that this is the case seeing that we do in fact have reprogrammable computers and time has shown us that brute force with a bit of heuristic does the job much better in silico. Perhaps neural networks will attain a level of sophistication and performance to provide an added advantage.
The reason brute force (even without heuristics) works so well in chess is that chess is a fairly restricted domain with a finite set of rules, and is thus perfectly computable. Many other human endeavors aren't nearly as conducive to simulation.
What I'm getting at here is that a reprogrammable computer would still be a computer and would still play chess like a computer which is fundamentally different than the way Kasparov plays chess. Also, I don't see why you need a reprogrammable computer to deal with contingency. Again, contingency in a chess match is highly constrained and can be generalized into heuristics. Chess playing algorithms are able to deal with unforeseen moves by their opponent.
On to the main argument of your post: the proud presumption of assuming that we can produce a machine better than a human being.
I understand your concern, but human beings have always had a desire to improve the world in which they live. This is one of the primary characteristics that differentiates us from many other animals. This desire to improve our environment and even our individual condition, need not be a bad thing nor a proud presumption. For example, is it presumptious of me to desire a future time where my lower back pain (I'm only 24!) could be eliminated? How about the pioneers of the medical revolution?
Finally, I think I understand the impetus for your post and probably agree with you for the most part. There is a general movement in Academic-type worlds to reduce the significance of the human being in a multitude of ways. Attempts to behaviorize, computationalize, informationalize, robotasize, etc. often DO appear to be presumptious and often are. It is unfortunate that ideology drives so much of what goes on in the world.
Still, your main thesis that presumption is native to any desire to "make us better humans" is faulty. Slippery slope notwithstanding, one can hope for a better world while still being pretty damned impressed with the current one.
[ 03. April 2003, 16:52: Message edited by: Micah Sparacio ]
IP: Logged
|
|
Danpech
Member
Member # 163
|
posted 03. April 2003 20:24
I just now lost the draft of my reply, so I'm going to try to recall the jist of what I wrote.
In reference to substrates, I was appealing the the evolutionist thinking. I did this in order to argue for the idea that there cannot be two versions of the ultimate reprogrammable computer (and for the idea that humans cannot engineer a better human computer). I think that a supernatural presuppostion readily allows any material to be embued with sentience (what is matter, anyway?), but that it is far from sure that materialism equally allows any material for this.
By 'computer' here, I mean it in the original (but now disfavored) sense of what the human mind does, which is as opposed to a 'simulated computer'. If I recall, this is how Turing used it.
By the 'ultimate reprogrammable computer', I mean the human mind. In comparing this to a simulated computer I'm referring to the whole computing package: function, size, energy requirement, computing time, etc.. As a parallel, I have the idea that human bipedal locomotion is superior, in terms of elegance for total human bipedal requirements, than anything that we can construct, even though we can make a bipedal machine that can out-do us in a given, limited range of requirements. Recall the statement that "a mouse is the best at being a mouse".
If chess is perfectly computable, then, even though a human generally requires more time to decide his move, a very simple computation could, in principle, be faster than brute-force calculation. And, at each move by the opponent, the criteria for computational elegance (without respect for time) may change, and this is what I am supposing that a human chess player seeks to attain (even though he can get bogged down by the subjective aspects). This is what I mean by the ultimate reprogrammable computer. I have the idea that the human mind is far more flexible than is any simulated computing that we can come up with. So, the more complex the task, the more efficiently a human can do it than can a machine, even though we can make machines to do it dependably while we go do something else.
I hope, now, you see what I mean by the proud presumption that we can produce a machine that is better than a human. The ability to effectively treat an illness does not mean that the over-all life of a healthy human can be improved, in the long run, with cyborg technology, bioengineering, etc..
quote:
Still, your main thesis that presumption is native to any desire to "make us better humans" is faulty. Slippery slope notwithstanding, one can hope for a better world while still being pretty damned impressed with the current one.
My thesis is only that perfect human health is IC, and is non-duplicable/unimprovable by alternate means (and it is far from sure that it is duplicable/improvable, which is what I refer to by my comment about presumption). If this is so, then it would seem odd that human intelligence is alternately duplicable. I hang a lot of things here on if's. But, I say that if all I say is true, then humans cannot artificially duplicate full functional human intelligence.
Thus, I say, tentatively, that we cannot make a chess playing machine that once-and-for all trumps humans as far as the computing package is concerned. My thought is that Kasparov's chess computing is far more thrifty per package (including, but not limited to, output) than is any machine we can make. Provided that the person is in perfect health and top form, the machine could never win a game if a draw is mathematically possible. (Not that we need be concerned that a machine can be made to do certain things better than can a human.)
IP: Logged
|
|
warren_bergerson
Member
Member # 262
|
posted 04. April 2003 14:35
Danpech,
Quote: I did this in order to argue for the idea that there cannot be two versions of the ultimate reprogrammable computer (and for the idea that humans cannot engineer a better human computer).
The ultimate pragmatic issue. Is the human mind in its current form the ‘ultimate reprogrammable computer’ as you suggest or is it the recently evolved, primitive, inefficient, ‘Wright Brothers’ version of a fundamentally new type of computing machine.
As I have discussed before, there is considerable evidence that the volume of ‘creative output’ of the human mind is continuing to increase at a rapid rate suggesting that the human mind continues to develop or evolve at a rapid rate. There is therefore at least a reasonable basis for believing that the mind can be made more efficient and more productive.
IP: Logged
|
|
Rex Kerr
Member
Member # 632
|
posted 04. April 2003 19:52
It's not clear to me how one comes to the conclusion that "no brute-force computing can match the ultimate reprogrammable computer". Both are equivalent to universal Turing machines, so the claim is presumably one of resources? But then if you define the ultimate reprogrammable computer as one that is maximally efficient, then the claim is trivially (and uninterestingly) true: the maximally efficient computer is as efficient as or more efficient than any other computer.
Also, note that the human brain with on the order of 10^14 synapses each potentially performing an operation at about 100Hz has a processing capability of something like 10^16 operations/sec; Deep Junior ran on an 8-processor system that would have computed on the order of 2*10^10 operations/sec. So in terms of chess, Deep Junior is a much more efficient use of operations than the human brain. (Then again, the human brain runs at about 15W, while the 8 processor system probably ran at about 1000W....) Anyway, the bottom line is that it is quite unclear what conclusions one can draw about efficiency from chess matches! (Elegance is subjective.)
Humans don't have "all-purpose elegance" in any deep sense. We can address almost any problem, but for complex algebra and calculus, any human is outclassed by orders of magnitude by, say, Mathematica, in terms of speed, or computations/watt, or whatever other measure you want to use. Mathematica is not always the best at putting equations in a form that humans find appealing; but then, we aren't always the best at putting equations into a form that Mathematica finds computationally efficient. So perhaps this "all-purpose elegance" capability is IC, or perhaps not, but we have no examples of it.
There are other odd questions in here, such as: "If you are not the foundation of reality, then why would you think that you can comprehend the foundation of reality?" Why would you think you could comprehend the foundation of reality even if you were the foundation of reality? And what does it mean for a foundation of reality to comprehend something? "Comprehending" usually involves forming abstracted models of how something works; if you generate the properties of something via your own high-level descriptions, that is not normally considered comprehension. Also, given that comprehension involves abstracted models, I don't see why one would suspect that high-accuracy models couldn't be found by extremely simple systems. After all, the whole point of abstraction is that you don't have to keep track of every tiny detail. It may not be possible; but it may. We have tried for a fair while ourselves, and have made some pretty good (albeit incomplete) progress.
IP: Logged
|
|
Danpech
Member
Member # 163
|
posted 06. April 2003 18:16
Hmm...
Maybe I 've been trying to compare apples and oranges.
Or, maybe not. Maybe I'm comparing sparks and stars.
You see, Kasporov is actually doing far more calculations, and far more efficiently system-wide, than any computer we have ever built. Even if he is not playing chess. The sheer complexity of simply managing a walk across the street in downtown Manhattan makes any simulated computing machine to date look like a two-dimensional gyroscope.
If one could speed up all chess games a hundred-fold, and have a person observe the patterns long enough, he could learn to play chess so well that he could beat Deep Jr. while stoned drunk.
Or, am I out of my mind? I've done things like this with patterns, although not nearly to the extent as I've just described, and not with chess (and not drunk). I've also done some things far more complicated than would seem to some people to be humanly possible, but not all idiot-savants are savant merely in regard to mental gymnastics.
In regard to my impression that humans are inherently superior at chess over man-made chess players, I'm willing to grant that I might be out of my right mind, but not by a lot.
[ 17. April 2003, 03:23: Message edited by: Danpech ]
IP: Logged
|
|
Danpech
Member
Member # 163
|
posted 17. April 2003 01:37
I think that an Idiot-Savant whose expertise is chess would be a truer example of the human chess playing potential against a machine. In comparison, I think that Kasparov is merely a practiced chess player. Imagine if you could play chess as naturally as you negotiate a busy crosswalk?
There is more than one way to play chess.
[ 17. April 2003, 01:40: Message edited by: Danpech ]
IP: Logged
|
|
yersinia
Member
Member # 324
|
posted 17. April 2003 02:29
An idea occurs to me:
1) Consider tic-tac-toe. After a bit of learning which moves to make, you can "solve" the game such that you never lose and always win or draw. Two such players will always reach a draw.
2) Consider checkers. Apparently the grand master humans (most of them are old guys) were considered to have effectively "solved" that game (although I guess as a computational matter it has not been completely solved yet)
3) So, perhaps the best humans and best computers have both reached a close approximation of "solving" chess, even though a full solution (calculating every single move) isn't possible.
Just a thought...
IP: Logged
|
|
Danpech
Member
Member # 163
|
posted 17. April 2003 03:39
Is chess severely more complicated than checkers? Is chess fundamentally more complicated than tic-tac-toe?
In any case, chess has patterns. It seems to me that only when a computer makes use of these patterns, instead of being made to rely solely on brute-force computation, that it can be said to approach the computational thrift, in chess-playing, of even the merely practiced human players. There is more than one level of patterns, and I think the Savant would be aware of the 'simpler' (i.e., more thrifty) patterns.
[ 17. April 2003, 03:46: Message edited by: Danpech ]
IP: Logged
|
|
gedanken
Member
Member # 594
|
posted 17. April 2003 19:31
Machines have bcome very advanced checkers players a long time ago, yet at chess only recently.
At tick tack toe, I have personally detailed complete strategies for both defense and offence. It is a very simple game. (By the way, first player can win if the second player doesn't make proper move on the first play. But if done properly, second player can always draw -- it gets booring since it is always the same game. (As Yersinia noted))
The "Idiot-Savant" is an interesting case. This is testable, as chess has a long history. So are there Idiot-Savant chess players? If so, what level did they achieve? My suspicion is that chess takes real intelligence and that the "Idiot-Savant" is not broad enough typically in thinking ability. Once again testable -- who knows of historical cases?
The human chess player is not inherently better (with criterion being rate of winning). This has been demonstrated with machines now at about a fair level with the best of humans. Soon an order or two improvement will occur in machine, but human will improve little in the game -- remember the game is very old.
[ 17. April 2003, 19:39: Message edited by: gedanken ]
IP: Logged
|
|
RBH
Member
Member # 380
|
posted 17. April 2003 21:04
Gedanken wrote quote:
Machines have bcome very advanced checkers players a long time ago, yet at chess only recently.
A checkers program written by Arthur Samuel in the 1950s was a notable achievement in machine learning. This site gives a short account of it, though it misses one very important capability the program had. When Samuel, its creator, was no longer good enough for the program to learn any more by playing against him, Samuel arranged for the program to play against versions of itself and learn from that. It ultimately played at a very high level.
John Holland has a very good description of the program and its implications for machine learning (still relevant today) in his Emergence: From Chaos to Order. Holland was a friend and colleague of Samuel in the early days of computing. That book is also useful in understanding how simple agent-based systems can produce complex behaviors and structures. Holland is the 'father' of genetic algorithms.
RBH
IP: Logged
|
|
gedanken
Member
Member # 594
|
posted 19. April 2003 02:24
A minor side note, I shook hands with Samuel once in 1980, at a Stanford robotics lab.
IP: Logged
|
|
Douglas
Member
Member # 641
|
posted 02. May 2003 12:21
gedanken,
Are you sure it was Samuel?
IP: Logged
|
|
gedanken
Member
Member # 594
|
posted 02. May 2003 13:14
Absolutely. I was standing next to the DEC-10 (KI series I believe) that was housing the second editions of Donald Knuth's The Art of Computer Programming, which finally started to come out in the last 5 years or so. (Knuth took a 10-15 year detour into typsetting, inventing TEX, etc. and writing on the subject. I did not meet him, only touched the computer with his book. The operating system on the computers was "WAITS", the Western Adaptation of MIT's "Incompatible Timesharing System".) Pictures of Samuel are linked here, including one from SAIL lab in 1970. He apparently still worked in the lab through about 1986, if my subtraction is correct, and it looks like he worked on MetaFont, which was Knuth's font generation software to work with TEX typesetting system at SAIL. The introduction is a little unclear, as it is stated that he logged into a computer at SAIL up through age 90, which would have been 1991. I hadn't known until reading this that Samuel had worked on MetaFont, but that would have put him at the lab, including other projects at SAIL. Hey, there's interesting tidbits in that WAITS link. The SAIL computers had a disk based video display that shared bitplanes between time-sharing users, so they could do video graphics. When someone was doing full image display, they could get assigned several bitplanes to map greater dynamic range, but when not they only got black/white mono plane. Custom terminals had video-feeds directly from the comptuer, separate from the terminal keyboard serial connection. Early WYSIWYG development occurred on that system. I didn't knkow they called them "buckey bits", but the keyboards had "control" and "meta" keys for the EMACS editor -- forerunners of the "ALT" or apple keys on PC keyboards.
We (Tom Binford and I) went up and saw the Stanford robot cart and I believe Shakey in an old storage room (neglected, not like in this picture).
I saw large fat cables that linked early versions of Xerox "Star" computers all over the labs -- that had interactive graphics and mouse interface. The cables were "Ethernet". (First PC based Ethernet card came out in '82, a couple of years later. Apparently Jobs had seen the Xerox star the year before that, prior to developing the Mac which spawned the mouse/windows style revolution. The Star was introduced commercially in 1981, the next year after I'd seen it. The pictures don't show the computer cabinet that was about the size of a two-drawer file cabinet, but the width of a standard equipment rack, which set beside your desk. Amazing how much computer history came out of ideas that started at Stanford and MIT AI labs.)
[ 03. May 2003, 10:19: Message edited by: gedanken ]
IP: Logged
|
|
Douglas
Member
Member # 641
|
posted 28. April 2007 13:58
gedanken,
quote:
Absolutely. I was standing next to the DEC-10 (KI series I believe) that was housing the second editions of Donald Knuth's The Art of Computer Programming, which finally started to come out in the last 5 years or so. (Knuth took a 10-15 year detour into typsetting, inventing TEX, etc. and writing on the subject. I did not meet him, only touched the computer with his book.
Sorry for the delay in replying, but I was joking about the possibility that you might have met a really advanced robot, rather than Samuel.
IP: Logged
|
|
|
Printer-friendly view of this topic