3 ms·

This seems like a really interesting area of research. As far as I can tell from my (limited) background on chess/go/etc algorithms, they seem to include lots o

This seems like a really interesting area of research. As far as I can tell from my (limited) background on chess/go/etc algorithms, they seem to include lots of hard-coded scenarios (like a bunch of chess openings) and brute force searching. Are there any teams working on more, I don't know, human-seeming algorithms? Everyone seems to agree that humans are good at these games from pattern recognition on a large corpus of game experiences. I'm out of my element here and realize I'm approaching "Even though I don't understand what you do, I'm going to assume it's easy for you to implement my suggestion" territory, but what if you encoded positions as general shapes rather than stone-by-stone? What if you played through lots and lots of games and built up Bayesian probability models of how likely certain combinations of shapes lead to victories? I mean, when a human go player looks at a board position and sizes it up what are they doing? I guess if we aren't particularly good at computer vision and understanding how the human mind works yet, then there are hard problems here, but it seems like a neat sort of interdisciplinary approach. edit: It looks like NeuroGo[1] linked to from an article in a different comment is more akin to what I'm thinking. It's just not very good.... ha. [1] http://webdocs.cs.ualberta.ca/~emarkus/neurogo/neurogo1996.html http://webdocs.cs.ualberta.ca/~emarkus/neurogo/neurogo1996.h...

exDM69 15y ago
> As far as I can tell from my (limited) background on chess/go/etc algorithms, they seem to include lots of hard-coded scenarios (like a bunch of chess openings) and brute force searching. When playing chess at an advanced level, you need to know a fair number of different openings (that even have names). Chess AI programs have a huge library of openings, endings and other game situations that are evaluated. In contrast, Go does not have a "library" of openings the same way chess does. There are simply too many ways a game can progress. However, in Go, there are many "micro patterns" that repeat over and over again, and you have to learn some of these even at a beginner level. There are numerous recurring formations, like the "Ladder", where one player inevitably loses and it should be recognized after 3-5 stones are in a specific form. So Go has these patterns that recur, but they aren't as clear and simple as in chess. While a ladder can be recognized after just a few stones, the outcome may depend on a stone on the other side of the board, that may be have been there for 100 turns. This is what makes pattern recognition tricky for Go AI.

exDM69 15y ago

> As far as I can tell from my (limited) background on chess/go/etc algorithms, they seem to include lots of hard-coded scenarios (like a bunch of chess openings) and brute force searching. When playing chess at an advanced level, you need to know a fair number of different openings (that even have names). Chess AI programs have a huge library of openings, endings and other game situations that are evaluated. In contrast, Go does not have a "library" of openings the same way chess does. There are simply too many ways a game can progress. However, in Go, there are many "micro patterns" that repeat over and over again, and you have to learn some of these even at a beginner level. There are numerous recurring formations, like the "Ladder", where one player inevitably loses and it should be recognized after 3-5 stones are in a specific form. So Go has these patterns that recur, but they aren't as clear and simple as in chess. While a ladder can be recognized after just a few stones, the outcome may depend on a stone on the other side of the board, that may be have been there for 100 turns. This is what makes pattern recognition tricky for Go AI.