5 ms·

The answer is actually: We don't know. Some people suggest comllex visual pattern matching is at play, but this seems really unlikely when you get to know the s

The answer is actually: We don't know. Some people suggest comllex visual pattern matching is at play, but this seems really unlikely when you get to know the space because: 1. Good Go players can often reconstruct an entire game just by looking at the board, if they have some idea how it started. Given that good go players can substantially alter the board in the course of play (called "playing under the stones" in many books), this suggests more than simple visual recognition. 2. Some go players can even play "one color go" which is pretty amazing to watch. Its basically a game of who can keep every move in their head. This isn't a silly stunt, some people really practice this. Personally, I think that actually Go is more like a contextual NL problem than a vision definition problem. The existence of things like "joseki" and the fact that small board games play out in such a radically different way than big board games suggests that a variety of human cognitive shortcuts are at play. It is absolutely the case that with just a few months of modest practice almost anyone can beat the pants off the best go playing computers.

mquander 11y ago
> It is absolutely the case that with just a few months of modest practice almost anyone can beat the pants off the best go playing computers. It sure isn't. The strongest computers are a few stones worse than professionals. You can count on your hands and toes the number of people in the United States that can beat the pants off the best go-playing computers.
Someone 11y ago
https://en.m.wikipedia.org/wiki/Computer_Go#Performance https://en.m.wikipedia.org/wiki/Computer_Go#Performance: "In 2009, the first such programs appeared which could reach and hold low dan-level ranks on the KGS Go Server also on the 19x19 board." I know virtually nothing about what it takes to become a low-level dan ranked player, but I would think it would take more than "just a few months" to "beat the pants off" them. Back to the subject at hand: I think we will solve mathematical go before we solve chess (where 'solve' is used in the mathematical sense, so that, for example, we can prove "chess is a win for white, in 53 moves", and mathematical go is as described in https://math.berkeley.edu/~berlek/cgt/gobook.html; https://math.berkeley.edu/~berlek/cgt/gobook.html; its difference with regular go variants is the way half stones are counted). Reason is that both games, even with extensive pruning, are too complex for an full search of their game tree, and go has a simpler structure, making it easier to reason about it without doing that exhaustive search. [I also doubt I'll live to see either happen]
- gizmo686 11y ago
  It has been demonstrated that Go (when generalized to an arbitrarily large board, and an arbitrary initial configuration of stones) is PSPACE hard. This means that an analytical solution would either need to be an optimization of brute force, or exploit some additional structure that results from starting with an empty go board. As an added wrinkle, such a solution may involve assuming that White [0] plays optimally. Ironically, this means that it may still be possible to beat someone who has fully analyzed and "solved" go, by making incorrect moves that puts the board in a position where their analysis only says that a winning move exists. [0] I am assuming that Black has the winning strategy.
  - jacquesm 11y ago
    This was my standard trick in chess when I still played a lot. If I would play against someone strong on openings but a weak player otherwise make an irrational opening move totally upsetting their apple-carts. If they were strong players they'd take advantage but if they were only strong on book knowledge you'd take them to the cleaners. Knowing your opponent in chess can be as important as knowing the game.
  - Someone 11y ago
    I'm not convinced proofs for arbitrarily large boards say much about solving the 19x19 board. There may be a generalization of tic-tac-toe for n-dimensional boards that are hard to analyze, but that does not imply that tic-tac-toe is particularly hard to analyze (to be fair, tic-tac-toe is an example where one basically has to do an exhaustive search to show it is a draw) A full analysis of go will have to include some "if N is less than X" clause for an X at least equal to 20, making it less beautiful, but I don't see that as a big hindrance (and X need not be optimal in a proof for a 19x19 board) In particular, the proof I know of (http://www.cs.bu.edu/faculty/gacs/courses/cs535/papers/Lichtenstein_Sipser_go.pdf http://www.cs.bu.edu/faculty/gacs/courses/cs535/papers/Licht...) uses building blocks so large that you can, maybe, fit 2 on a 19x19 board. Also, I think it is clear white doesn't have a winning strategy, as black can pass at the start of the game.

mquander 11y ago

> It is absolutely the case that with just a few months of modest practice almost anyone can beat the pants off the best go playing computers. It sure isn't. The strongest computers are a few stones worse than professionals. You can count on your hands and toes the number of people in the United States that can beat the pants off the best go-playing computers.

Someone 11y ago

https://en.m.wikipedia.org/wiki/Computer_Go#Performance https://en.m.wikipedia.org/wiki/Computer_Go#Performance: "In 2009, the first such programs appeared which could reach and hold low dan-level ranks on the KGS Go Server also on the 19x19 board." I know virtually nothing about what it takes to become a low-level dan ranked player, but I would think it would take more than "just a few months" to "beat the pants off" them. Back to the subject at hand: I think we will solve mathematical go before we solve chess (where 'solve' is used in the mathematical sense, so that, for example, we can prove "chess is a win for white, in 53 moves", and mathematical go is as described in https://math.berkeley.edu/~berlek/cgt/gobook.html; https://math.berkeley.edu/~berlek/cgt/gobook.html; its difference with regular go variants is the way half stones are counted). Reason is that both games, even with extensive pruning, are too complex for an full search of their game tree, and go has a simpler structure, making it easier to reason about it without doing that exhaustive search. [I also doubt I'll live to see either happen]

gizmo686 11y ago
It has been demonstrated that Go (when generalized to an arbitrarily large board, and an arbitrary initial configuration of stones) is PSPACE hard. This means that an analytical solution would either need to be an optimization of brute force, or exploit some additional structure that results from starting with an empty go board. As an added wrinkle, such a solution may involve assuming that White [0] plays optimally. Ironically, this means that it may still be possible to beat someone who has fully analyzed and "solved" go, by making incorrect moves that puts the board in a position where their analysis only says that a winning move exists. [0] I am assuming that Black has the winning strategy.
- jacquesm 11y ago
  This was my standard trick in chess when I still played a lot. If I would play against someone strong on openings but a weak player otherwise make an irrational opening move totally upsetting their apple-carts. If they were strong players they'd take advantage but if they were only strong on book knowledge you'd take them to the cleaners. Knowing your opponent in chess can be as important as knowing the game.
- Someone 11y ago
  I'm not convinced proofs for arbitrarily large boards say much about solving the 19x19 board. There may be a generalization of tic-tac-toe for n-dimensional boards that are hard to analyze, but that does not imply that tic-tac-toe is particularly hard to analyze (to be fair, tic-tac-toe is an example where one basically has to do an exhaustive search to show it is a draw) A full analysis of go will have to include some "if N is less than X" clause for an X at least equal to 20, making it less beautiful, but I don't see that as a big hindrance (and X need not be optimal in a proof for a 19x19 board) In particular, the proof I know of (http://www.cs.bu.edu/faculty/gacs/courses/cs535/papers/Lichtenstein_Sipser_go.pdf http://www.cs.bu.edu/faculty/gacs/courses/cs535/papers/Licht...) uses building blocks so large that you can, maybe, fit 2 on a 19x19 board. Also, I think it is clear white doesn't have a winning strategy, as black can pass at the start of the game.