Some Philosophical Questions about Computers and Go

    Keywords: Software

Please feel free to answer a question, pose a new one, rant about the author, etc.

Table of contents
Table of diagrams
Which is the best move?

How long will it be until a computer is clearly the Go champion?

  • Will Moore's Law hold for the next 50/100/200 years?
    • It's proven very difficult to figure out what Moore actually said, since he's misquoted in so many different ways. So to answer this without reference to Moore's Law: Computers will continue to improve at surprising and impressive rates forever, not just for the next 200 years; the rate of improvement probably cannot be estimated beyond the next two years.
      • Moore's original paper (from 1965) is available at [ext] It expresses what was later called his law in terms of number of components per IC, but subsequently it became more common to measure the linear feature size, or just "speed" instead. The speed increase has been a factor of 2 every 18 months, ever since the 1970s (it was faster than that during the 1960s) and it is generally believed that this rate will persist for the near future. However, in a fairly recent (1990s) video, Moore predicted that the ultimate limiting factor in the rate of semiconductor improvements will not be physical or technical, but economic: the cost of semiconductor manufacturing plants is growing exponentially. That is not currently a problem, as demand is growing too; but if the world's appetite for faster silicon continues at this rate, then the amount we spend on semiconductors will equal the global GDP by some time in the 2020s. So I believe that Moore's Law as we know it will not apply for the next 50/100/200 years, because we will not be able to afford it. -- Gary
    • There is no particular evidence that an increase in computing power is sufficient to make the current programs into world class champions, almost without regard to the computing power available. Given current program skill, the programs would need to look dozens of moves ahead (let's say 30) and try all potentially remotely reasonable moves (lets say 50 moves). Standard alpha-beta routines put that at 50 ^ 15 moves tried and carefully evaluated; that's about 10 ^ 25 or 2 ^ 84. Say a modern computer can evaluate 10 moves a second; further suppose that your future supercomputer is 100 billion times faster (10 ^ 11). This means you can generate a good move in only 10 ^ 13 seconds, or only 300,000 years! Now, relatively standard pruning algorithms for selective searching will significantly cut down the number of nodes tried; to get it down to 1 day per move would require a pruning factor of about 10 ^ 8, which is probably doable. So, if Moore's Law holds for the next 55 years (doubtful), then maybe the standard techniques would produce a professional grade player; I doubt even that would produce a world champion, though. In short, we need more intelligent programming; the go programs are getting better as a result of a lot more than just faster CPUs. A professional quality program within 50 years would not really surprise me; a world champion would. -- evand
      • While I think that 1p is reachable and strong 9p is unlikely unless true AI is developed, increasing computer speeds can buy us correspondingly increasing reading capability for tsume-go, ladders, connections and so on, where the branching factor is typically much less than you show above. Further, not all of the tactical readings would need to be repeated at each move. This increasingly strong tactical foundation could allow a program to advance faster than a human who understands the same level of strategic planning. Designing enough knowledge to read the tactical coup in the game shown under Yi Se-tol would be a real challenge, I think, but the low branching factor in the forced moves makes it at least conceivable. -- Scryer
        • Niklaus: The difference in playing strength between 1p and 9p is rather small, tiny compared to the difference between go programs now and 1p. What makes you think that if 1p level is reachable, we can't go all the way to 9p?
          • The difference in playing strength is small, but the thing to realise is that the difference is a reasonable fraction of the distance to perfect play. If it takes a person as long or longer to go from 1p to 9p as it does to get to 1p in the first place, why should we expect a program to advance any faster? We can already see that programs improve more slowly as their strength increases, I would expect this to be more true at the professional level, not less, especially as the number of people capable of playing go at that level and capable of writing good programs gets smaller and smaller. -- evand
            • I think you are over-estimating the strength of 9p players. Cho Chikun says he feels overwhelmed when he thinks of the vastness of go, i.e. he doesn't think he really understands much on an absolute scale. I'd say the distance from 9p to perfect play is much greater than the distance from 1p to 9p. -- Bob McGuigan
            • Bob makes a good point, and the difference between 1p and 9p has differed over time, nevertheless, my feeling, in my own pathetic road of improvement, is that the knowlege base for improvement is not linear. I am no good at math, but each stone of strength seems to require at least twice as much additional knowlege to attain as the last stone of improvement. If this is the case, and continues into the pro ranks, then the purported one stone difference between 1p and 9p is huge, far larger then 4k to 5k, even though based on play, it is the same one stone difference. [George Caplan}
        • First, All those assumptions are based upon the evaluation of board including a lot of things like deep tactical reading, life and death reading, and full board positional judgement. My point is that in order for a current go program to make a professional quality judgement about the value of something like a strategic attack on a group, it would have to read out all potentially reasonable continuations of the attack, including things like leaning moves and kikashi elsewhere, and try other attacks on the board in attempts to set up a double attack, etc. Current programs do caching that avoids repeating the reading when possible; I just assumed that the cache would not change substantially, which is probably correct to within an order of magnitude or so. Finally, I'd like to point out that the Yi Se-tol game you mention is *not* a tactical coup. Anyone can see that if you keep playing out the ladder, you eventually get that move (not everyone can think of it, though); the problem is one of strategic judgement: having played out the ladder and captured the stones, who is ahead? has b so badly damaged his position by playing a dead ladder that the game is lost, or is the capture big enough to make up for it? This is not an easy judgement to make for a human, and a go program of today could only make it by reading out the ladder, and then reading out a reasonable fraction of the remainder of the game. -- evand
    • Tas: As a nanotechnologist I will say that moore's law can at most go on for another 15 years before we hit fundamental physical limits to the power of computers. Limits such as the speed of ligth, the size of atoms, and the entropi of a bit. However it may be that even a simple quantum computer with somthing like 500 qubits, could solve 19x19 go.
  • When a computer becomes the 19x19 champ, will a person still be able to beat it at 39x39?
    • I hate to think how much importing a 39x39 goban from Japan will cost! And where would I play it?
    • 19x19 has features beyond mere scale; it is unique among candidate board sizes (NxN, where N is an odd number) in that it is closest to the point where territory on the third line and below is equal to territory on the fourth line and above. Any different size would change this balance and hence upset the quality of play. -- Bignose
      • Sorry, I don't buy this theory! If B only plays on the third line and W only plays on the fourth line, then B wins by 15 points.

Bill: And Black makes 8 more plays than White. (Pointed out by Fujisawa Hideyuki, IIRC).

      • If they are playing on a 21x21 board, then W wins by 17 points. The two boards are about equidistant from this fabled point. In any case, I see nothing special about the third and fourth lines. Go is undoubtedly a deeper game the larger the board, and historically the standard board has kept growing in size. The limiting factor is probably the time needed to make all the moves, which of course grows quadratically. -- Saesneg
      • quadratically - It grows a little more complicated than that. The number of intersections grows quadratically indeed. There are, however two contrary influences on the increase of the time needed.
        • Some evidence suggests that the number of moves to dispute the same amount of intersections decreases with the size of the board.
        • The complexity of the individual moves increases with the increased number of possible moves. -- mAsterdam
      • It hasn't grown that substantially. The "classic" old board is 17x17 and the current board is 19x19. The reason for sticking with a 19x19 as opposed to a 21x21 is that the go academics felt anything larger than 19x19 would be near impossible for a human to reasonably read. The combinations and possibilities would be too vast. As for the time it takes to play out, in spite of the increase in possible combinations and moves, I don't believe it would be much greater than a 19x19. People rarely fill every intersection on the board, and with bigger sizes there would likely be larger alive territories. Sure, you can play on any size you want, but players of old have long considered a 21x21 board and still opted to stay with a 19x19. But when you're dealing with AIs, it would be folly to concern yourself with larger boards when you can't even handle a 19x19. -- LeiMagnus
    • I'm going to conjecture that there exists an A(n) algorithm for deciding upon the best move in a board position, where n is the size of the board (in intersections). For those who haven't seen this notation, I'm saying that the average time taken by the algorithm to find the best move (on a board with n intersections) is bounded by n multiplied by some constant (which we may or may not know). To put it another way, for an arbitrarily large board, doubling the size of the (edge of the) board will not make the task of finding the correct move significantly more than four times as hard. --Bildstein
      • More accurately, A(n) describes a function of n, which means that it's the correlating value resulting when the function is applied to n. It's may or may not be equal to n multiplied by some constant. I'd bet my life on it not being n multiplied by a constant, though! For example, say that the board is 19 x 19, and usually a game is over when, say, 50% of the board is covered with pieces (that's after [19 x 19 x 50%] moves). So A(n) might be equal to (n^2)^(n^2 x 50%), or n^(n^2). Have a go for yourself - there's no constant there, and it's increasing faster than what we call exponentially. A 17x17 board would be significantly simpler to calculate than a 19x19 board - I'd say in the order of hundreds of thousands of times less complex, at least. Kirk79
      • I disagree. What I'm conjecturing is that as the size of the board grows, the time it takes to find the best move goes up only linearly. For example, I certainly don't find it takes me hundreds of thousands of times longer to make my move on a 19x19 board than on a 17x17 board. --Bildstein
      • But I doubt that you are doing anything like finding the best move. Unless Go has some pretty sophisticated mathematical properties (I don't really think it does), one should expect that the time taken to verify a winning move should go roughly as the number of possible games. It seems that even an algorithm for determining the winner on a finished board would be A(n^2), with n the linear size of the board.
      • You're thinking too analytically. Just because it takes, say, n^n time to verify that a move is the best doesn't mean that there isn't an A(n) algorithm for finding the best move. Also, you're wrong that what I'm doing when I play go is nothing like finding the best move. Very often, I find the best move. Unfortunately, even more often I don't. But when someone peeps, for example, I know to connect, and it doesn't take me very long at all to figure that out. Why is it beyond the realm of possibility that you could take a million of these little rules and come up with a winning (or at least play-to-draw) strategy? I'm conjecturing that this is indeed possible. And if it is, I think it's then quite believable that it scales up linearly. -- Bildstein
      • Thinking too analytically? How do you suppose computers actually operate? What you are suggesting is aggregating a bunch of heuristics. That's fine, except that the implementation of a heuristic in a program still requires analysis. Especially since it is not always correct to connect in response to a peep, and determining whether it is the best move will indeed depend on more than just the immediate situation on the intersection around the peeping stone. Indeed, how do you suppose a computer would determine what, exactly, a "peep" is except for analysis? I can trivially construct a situation where connecting a peep is not the best move. Surely you are capable of this, too (especially if a peep was not the best move!). And once we admit that, we must admit that answering a peep is often a whole board problem, meaning that the problem scales with the size of the board, and that itself is not linear.
      • By "analytically", I suppose what I was getting at is that it is not always necessary to read out all possibilities until the end of the game to establish that a move is the best move. Here's a trivial example:
Which is the best move?  

Even on this 9x9 board, to read it out to the end would be almost impossible. But you or I can easily say, with great confidence, that a is the correct move for White. But I'll happily admit that you could come up with an example where it takes an incredibly long time to figure out the best answer (indeed, I know of some, often called "the hardest problem(s) ever"). What I'm conjecturing is that the latter is the minority, and on average, the time take to find the best move increases linearly with the board size.

Herman Hiddema: I think the best move for white is pass. Black is dead, capturing these stones only gives him a chance to perhaps live inside to area opened up.

Bildstein: Good point. I suppose I have to admit that even in this trivial example, I failed to find the right move. Anyway, if I can be forgiven for that, let's assume we're playing no pass go.

aceofspades: Sorry to go back to the beginning of your discussion, Bildstein, but didn't you mean an O(n) best-move algorithm, since that does in fact mean a function bounded by a multiple of n, but Kirk interpreted your A to mean an arbitary function?

robw?: This discussion is probably long over, but I have real problems with the phrase "best move". In the case of the 5x5, it seems that someone has proven that there is a best set of moves that will always lead to victory. Is there actually a proof that says that the same thing goes for 19x19? I mean.. lets say you have a group in the corner and you and your opponent are fighting over 20 points. Your opponent makes a move that seriously threatens this group. Let's even say that almost all pros would agree that you should go there. Who is to say that there isn't a shape on the opposite side of the board that has an unavoidable path 20 moves deep that leads to a 40 point gain? Maybe no human is able to read this move... and the number of possible positions is beyond the number of atoms in the universe. It may still exist and could be considered "the best move" by god or whatnot. To expect a human or computer to make the "best move" seems very artificial to me.

If a computer becomes champ, will Go still be interesting?

  • Unlike chess, people could still play handicap games
  • We could decide "against God, I would need four stones."
  • Computers have "solved" many other games that people still find interesting to play against one another
    • Checkers being a prime example. (But a bogus one, in that it has not been solved. The closest an entity has come to Kami no itte in checkers may be Schaeffer's program [ext] Chinook, which has won a human championship but has not solved the game. Solving checkers is a long-term goal of the Chinook project. --Scryer
      • They have solved the White Doctor opening (10-14 22-18 12-16) at 2004-08-02.
      • And in July of 2007 the entire game was solved by the Chinook project. Relevant to this issue, [ext] a science article on the subject began with the sentences "The ancient game of checkers (or draughts) has been pronounced dead. The game was killed by the publication of a mathematical proof showing that draughts always results in a draw when neither player makes a mistake."
  • The methods used by computers to beat humans are usually uninteresting to humans (brute-force problem-space searching) who will still want to improve their own play
  • For me the interesting aspect of having an invincible computer would be to observe games of an even higher quality than professionals achieve. Not that I understand those either. I would like to read games of Invincible v. Invincible, commented by professionals. But I think there is little chance of such a machine in our lifetimes. -- Saesneg
    • Assuming that "Invincible" implies "Perfect Play" then it shouldn't be too interesting (at least no the second time) because it would theoretically be the same game over and over again. Both players always get the maximum amount of territory possible for each game position. It might be rotated or miai might be played our slightly differently, but that should only allow for a few variations (and most likely minor ones at that). --IsaacFreeman
      • Why would the difference have to be slight? Our notion of limiting, at least, seems to have near-perfect play playing out a wide and varied game. There are so many board positions possible that I'd suspect that there would be many qualitatively different, perfect games. Miai could be on the scale of twenty moves or more. Near-perfect play, at the least, seems to attest to this: how many different games have there been with the same score? Countless, essentially. When play becomes perfect, we might be limited somewhat, but it could well be that play is, say, restricted to .1% the variety of current top play, which would still probably be very varied.
      • You could also ask the computer what is the best thing to do now and study diferent openings and situations with a "super pro" always at your disposition.
  • For all those people like me, who for practical purposes will never run out of people who can beat them easily, it won't matter at all if a computer joins them. For the casual player, it's for the joy of playing or self improvement, not being the best. People who can't lose either get very frustrated or very good. --Confused
  • It wouldn't bother me if computers were to become better go players than humans. It doesn't bother me that a race car can travel a mile faster than I can or ever will on my own. I still get satisfaction out of "personal bests". So I'll still enjoy go when computers become the strongest players, though maybe it won't happen in my lifetime. -- Bob McGuigan

Go exposes moral character, so computer Go players will always suck at some wabi/sabi level?

  • Not necessarily. Just like in music, computers can produce pleasing pieces without a trace of soul, which are better than most of the junk produced by humans with a soul - if they haven't sold it already. Only because Go can be used to express one's character doesn't meant that every game has to contains a character expression. -- Confused
  • I am pretty sure that Go will never be "solved" although artificial Go player may one day be better than any human player. Michael Moos?
    • The go board has a finite number of states. (Yes, it's large, but it's finite.) 5x5 go is solved. ([ext] Please explain the size where it becomes unsolvable, and why that number is special.
      • Likely there is a size where go becomes unsolvable (although I don't think it's only 19x19). To solve a game you need time and storage - in a universe that is finite in size and probably has a limited lifespan, you don't have an infinite amount of either. And no, I don't think this will ever be a consideration in computer go, but it's a relevant observation. Mitja Lustrek
        • Who says the universe is finite? But whenever we're dealing with a finite number of moves, it will always be solvable. Finite number of moves in an infinite universe is solvable. Finite numbers of moves in a finite universe is solvable. Only an infinite number of moves in a finite universe is not solvable. Theoretically solvable, of course - practically, no computer will ever be able to solve 99 x 99 Go, for example. Maybe not even 37 x 37! Mainly because there's no practical need to develop such a computer. Kirk79
          • Infinite in infinite may not be solvable either, BTW. But what you don't see is that even 19x19 is so large that we can't hope to solve it by conventional computers. Not even 9x9 is to be solved any time soon. Maybe it is possible with quantum computers, but we can't be sure of that yet. Dankrad Feist?
            • Tas Although some algorithms might reduce the number, a pure brute force solving would require storing of all possible board (and a lot of other things). The number of possible board positions is 3^361 which is somewhat larger than the number of particles in the visible universe (estimated 10^80), and you need many particles to store one bit, and many bits to store a go board. There is no particular size that becomes "unsolvable" but there is a size above which it will never be solved. And that size is definitely lower than 19x19, and probably 9x9 too.
  • Jgo? Concerning what Confused said about music, computers don't make music, humans do (currently). My computer has helped me many times in composing new pieces, as a tool. But all the computer scientists in the world couldn't get my computer to compose the rest of the music I would like to write. There are too many variables for it to consider for it to even come up with a melody on its own. A human composer knows that a new melody must be between too typical and too original. In example, A c-scale is too typical, but not if it is done in an original way. How can that be programed? Also, when computers started exceeding the chess playing abilities of humans it was because the computers have moral character and humans carry the burden of moral character. A computer will win repeatedly because it has no sympathy for a life form that tires. The human strength against computers is that we can find its programing glitch and exploit it, but eventually this feels like a pointless exercise to the human. I do feel ashamed if I lose to my go engine, even if it was because I got distracted. Because of my character I feel no pride when it crashes due to a program failure, even though a program crash is just as valid a loss as me getting distracted and blundering.

When a computer becomes champ, will anyone care?

  • All non-luck games will already have been won by computers. Go will inevitably be the last skill game to fall?
    • Wouldn't its being the last game make it more worth caring about for that?
  • 25-50 million people around the world play Go. They might care :-)
    • A similar number play chess. Did they all care?
      • Klaus: Chess has not been solved yet! Kramnik was (at very least) better than Fritz, and Kasparov as good as Junior. Human chess players still suffer from trying to change their playing style according to their opponent. And, by the way, even people I hardly consider to be chessplayers did care!
  • Many people don't consider the brute-force problem space search methods to be very interesting from a can-computers-think perspective, because there is no creativity shown by such methods.
    • Computer players are not limited to brute-force problem space search methods and neither are geometric prover programs which have already come up with creative new proofs for old theorems.
    • As Kasparov-Deep Blue proved, Chess was a dead end for Artificial Intelligence research (as have other games). So there is no worry of a computer being champ. -- Tim Brent
      • This conclusion seems muddled. That a computer has beaten the best human at a match set of games of chess does not mean that another one can't be programmed to beat the best human at another game, such as Go. Games were not a dead end for AI (Artificial Intelligence), but actually a fruitful area that led to many techniques, such as planning, backtracking, pruning, and blackboards. One can easily imagine that a sophisticated game such as Go will lead to other sophisticated techniques. Further, some existing techniques such as neural networks and genetic algorithms have not been exhaustively applied to Go, yet.
  • "The question of whether a computer can think is no more interesting than the question of whether a submarine can swim." Edsger Dijkstra
    • One of my favorite Dijkstra quotes, and it's being abused here. There are people who say "Sure, computers can play chess, do calculus, navigate city streets, and compose music, but can computers think?" Unfortunately for you, should you try to debate such people, they won't ever give you an operational definition of what it means to think. Press them hard, and you come to the conclusion that thinking is "whatever I can (mentally) do, that a computer can't do". Well forget it, Dijkstra says, such questions are uninteresting. A submarine can locomote, through water, from point A to point B. Do you want to call that "swimming"? Who cares? Similarly, the question of "can a computer play professional-level Go?" is very interesting. The question of "is such a computer thinking?" is pointless. -- John Aspinall
      • I don't think it is being abused. I think it being used in exactly the way you intend.
      • If so, then I apologize for ladling on unneeded criticism. But if so, then why is the quote here at all? If we are discussing Computers and Go, and we are all clear that discussion of "thinking" just adds noise to the signal, then why include a quote about computers and thinking? John Aspinall
      • Whoever this Dijkastra person is seems to be making a trivial point (at least taken out of context). As is implied by Aspinall such questions are dependent entirely on the semantics of words like "think" - essentially the provision of an operational definition of thinking. Consequently it is facile to say they are uninteresting in general - they are exactly as interesting as the meaning of their referents. It doesn't seem trivial to me to debate whether or not what a machine is doing when it composes music etc. is thinking. The really interesting question is, "What would the construction of a professional Go-playing program tell us about thinking, under a given definition of the term?" In the case of the above discussion of brute force methods the answer seems likely to be, "nothing at all" if the definition of thinking adopted implies (as I think it usually does) thinking in the sense that humans do it.
        • Dijkstra [ext] was a moderately famous computer scientist (amongst computer scientists). Obviously his interests and yours differ.
    • unkx80: Abused or not, this statement by Dijkstra is often quoted in artificial intelligence texts. The computer scientists do not seem to have a consensus as to whether a computer (they used the word "agent") is intelligent or not, in fact, there are at least four schools of thought on what intelligence as applied to an agent is. Hence, I do not think think people can ever agree on whether a computer can think, specifically, whether a computer player playing Go is a thinking player. One thing for sure is, some undergraduates will remember Dijkstra for his famous algorithm for finding shortest paths.
      • Rethinking what I wrote above I suppose what I would regard as interesting is what could be learnt by constructing a go-playing agent that didn't rely on brute force to any great degree. Douglas Hofstadter (for example) regards analogies as central to human intelligence. Go seems like a wonderful space for representing analogies generally within. On the subject of quotes from AI books, Russell and Norvig disappointed me greatly in their otherwise excellent book by saying, "Go is a game which seems likely to benefit from future advances in AI."
        • Michael?: But what makes you think human brains don't use brute force when solving problems?
        • Bill: (Later.) Consciously, they do, but not very well. Unconsciously, human solutions are too fast for brute force, besides being more fallible.
        • I didn't say they don't use brute force but that they don't use it exclusively. Otherwise we'd never be able to reliably come up with reasonable answers to big problems in a short space of time. What I wanted to distinguish between is (in the ideal case) a perfect go-playing agent which searched _every_ possible board conformation and made the right move every time as a result (hence 'solving' 19x19 Go) and one which made excellent moves as a result of a combination of some brute force and some heuristic measures. The point is that humans can explain board states with reference to mental states and groupings of pieces. Something which purely enumerated every possible board state (and did *nothing* else) would not be able to do this, since its 'understanding' of the board is based entirely on chains of past and future play and contains no higher-level grouping at all. Finally, think of how you play Go, and how professionals describe it - in commentaries it is common to read many moves ahead but no-one ever describes a fuseki in terms of every possible continuation to the very last move - even if we assume that humans enumerate the search space to some degree and do so faster than modern computers (and therefore play better) the space is still far too big for us to be doing it to any great extent.
  • 1) Humans will be the ones to create the computer that can beat any human, so in a higher sense it's a moot point. 2) The game is complex enough, for people at least, that there is a strong psychological aspect of the game. I can't study myself to dan-level, it has to be something I'm good at, otherwise it wouldn't be an achievement. A computer, or person, that has one unchanging, monolithic strategy by which he/she/it always wins is boring. IOW, we won't care if a computer is the best, because we won't play him much. 3) In keeping a sense of goodwill towards the concept of playing a game, winning shouldn't be ultimate goal. Those of us that truly like to play, ergo, won't care if a computer is the best. Those of us that do want to win more than they want to *play* won't care either. Only one of us on Earth is in the unfortunate position of having no human bet\ter than him/herself. I'll leave it to him/her to resolve that issue. --Baz
  • ivoSF?: i think its more likely that an AI with true inteligence is developed that learn to play go using its human like but(asumed) superior inteligence and thus get beter at it then humans, then that a specialized program beat humans.
  • Bildstein: I'm a lot stronger than The Many Faces of Go, but sometimes I play handicap games against it to practice my handicap go. Some people say that playing against computers is bad for your go, but if a computer program got to professional level, I'd certainly like to have the program so that I could play against it and learn from it.

Intuition is likely unreliable in these matters

As a strong computer programmer (with an interest in AI), and a weak Go player, I'd be very reluctant to make judgements about computer Go. It's obviously a very hard problem, but I think that for someone's intuition about computer Go to be especially valuable, they'd need to be an upper-level dan player and have a talent for innovative AI solutions. There's a couple of possibilities, none of which are mutually exclusive, and it isn't easy to guess which might apply:

  • Modest strategic abilities and decent heuristics might be prove formidable in the hands of a machine that could read out millions of board positions per second. The computer wouldn't necessarily have to have great intuition--just good enough to prune the search tree to point where brute force becomes formidable (this appears to be what happened with Deep Blue; it's not so much a great chess player as an adequate player with inconceivable reading ability).
  • Computers could prove formidable for a few games because of novel play style, but would not stand up in long-term competition against people who specifically worked to attack the computers' weaknesses (I gather this is already the case, for single-digit kyus). Similarly, Deep Blue would not likely stand up to Kasparov in series of a few hundred games without heavy, ongoing intervention by the programmers.
    • As a side note, even when Deep Blue challenged Kasparov, humans were making changes between matches. There were some mistakes that it kept making and humans fixed it before the next round.
  • Dan-level play could prove truly deep, requiring AI advances that bring computers to near human-level intelligence in several important areas. Even so, I wouldn't be surprised to see this happen within my lifetime--like Bill Joy, I sometimes wonder if humans aren't in danger of obsoleting ourselves.

Rarely are interesting AI problems solved by search. It's often a leaden, uninspired technique. You've got to look for either a sudden insight or a methodical understanding of the actual problem.

A lot of supposedly hard "AI" problems get partially solved when somebody comes up a key insight--for example, the MIT "eigenfaces" algorithm boosted face recognition from 20% to 90+% accuracy by applying undergraduate statistics cleverly. Spam filtering went from horrible to 98+% accuracy with the discovery of a clever hack inspired by Bayesian probabilities.

Other times, progress in "AI" problems comes from dutiful enumeration--and tweaking--of hundreds of rules. For example, look at the sudden jump in bot quality in Counter Strike: Condition Zero. This was based on a few clever algorithms and lots of hard work, carefully enumerating hundreds of human-like behaviors and limitations, and testing their interactions extensively. The end result is fairly similar to playing against real humans, but only because the author was an expert Counter Strike player.

Thus, any workable solution to computer Go is likely to be surprising in nature. Certainly no low-level Go player understands the game well enough to know what problems need to be solved to play at a consistent 1p level, and few professional dans have spent enough time studying AI to have a good intuition for what novel techniques might be available. Even then, a few hundred dan-level AI hackers could overlook important tricks for a decade or two, only to be surprised by an AI "tesuji".

erislover: I think the last sentence there about an AI tesuji is critical to understanding the philosophical issue here. We have a baseline of professional strength as a measure of go-playing ability. Indeed, amateur players emulate pro play to the extent that they can, study joseki (which is just another name for "what pros play"), and so on. Only late in the endgame when the search space shrinks considerably can computers overtake even professionals in their calculating ability.

This sets us up for a very bad problem: a novel advancement in a certain area might be overlooked for the lack of advancement in the rest of the game. A correct algorithm for some particular line of play might surely be excellent, but the mistakes made elsewhere could overshadow it. Unless the assumption that pro players are playing correctly is correct, we could easily miss strong routines by artifically selecting them away. Perhaps opening on tengen is always the right move in an odd-sized board. Perhaps the ideal opening is always two to five moves from a corner, and this just happens to be the center point on smaller boards. Who is to say?

Since go is a zero-sum, perfect information game there is a perfect line of play. But since it is incalculable, it might as well not even exist, IMO. Faith does not really move mountains and if such play is computationally unknowable then it seems quite likely that game-theoretic approaches to play are flawed from the start. As I often like to ask, suppose a powerful being approaches you and claims to be god. What is the difference between a finitely powerful being that is more powerful than you can imagine and an infinitely powerful being? I think we face such a problem here.

If computers do become as strong as professionals, it will be because we've used professionals as the ruler the entire time. It will have absolutely nothing to do (except coincidentally) with perfect play.

How about a hybrid (human/computer) mix?

axd: The discussion has always focused on machines playing on their own. But how about the computer tracking variations, the human directing these and taking decisions? Possibly disgusting evolution, but cannot be avoided. Some day someone will come up with software that "looks over the shoulder" of a go server player (eg analyse the screen on which the go client is running), and assists in decision making. Unplug the human from the machine and his/her rating will drop, that is most probable. But what we are interested in is in the impact of such players on the GO scene. Should they be quarantined in their own arena?

Quarantined? No. Why should they? I don't care whether my opponent uses a computer to help him (except for playing tournaments or things like that). I will have fun playing. If he still has fun is his problem (I wouldn't, that's for sure). And even if people care what their opponent uses, how could players like these be quarantined? It is simply not possible. Dolgan

KarlKnechtel: I have mentally toyed with the idea of a Go tournament where players may bring along any computerized aids they like, with the restriction that they must have written the software themselves. Might be interesting :)

aceofspades: Um, I know I'm not stating anything specific, but I think some of the most interesting "Philosophical Questions about Computers and Go" are about rating - that is, rank-absolute levels. Much of that discussion relates to computer players, and an ideal tournament of a bunch of Turing machines playing each other.

daniel: I have actually been working on a project similar to this; However the point of the project is to teach people to play. essentially the idea is to first ease the new player into the concepts like territory; influence; etc. then slowly remove those crutches by providing challenges in the form of unlockables (the crutches would only be available against the AI) which could reward with new challenges which would earn points of some kind to purchase accessories; eventually it would reach challenges like difficult tsumego; one color go, multiple opponents; only things that I think would help produce a better player. it also implements the newest version of my "personality" AI engine where the decision weighting code with the newest version using equations to control that (allowing for dynamic behavior adjustment) which should at least solve some problems associaed with learning from a computer opponent.

See also [ext] topic on rgg: The guy who stopped Go Programming Development ...

Some Philosophical Questions about Computers and Go last edited by on June 20, 2023 - 20:18
RecentChanges · StartingPoints · About
Edit page ·Search · Related · Page info · Latest diff
[Welcome to Sensei's Library!]
Search position
Page history
Latest page diff
Partner sites:
Go Teaching Ladder
Login / Prefs
Sensei's Library