24 ms·
There's an interesting trend in the computer go community away from rule-based AI, and towards a more stochastic approach. This means that where the leading dev
by prospero 18y ago
There's an interesting trend in the computer go community away from rule-based AI, and towards a more stochastic approach. This means that where the leading developers were all 1 dan or better (very talented amateurs, by the metric of the article), it's become possible for talented programmers with a relatively limited understanding of the game to make major contributions.
Go is a complicated game, and the human response to that is to create rules of thumb and build a strong intuitive understandings of different positions. It's arguably less elegant to analyze a position by running hundreds of thousands of playouts, but it's finally a problem that resembles real computer science (though most academic papers are still of the "we tried this, and it did okay" form). It's a pretty exciting time to be involved, and I'd suggest that anyone with even a passing curiosity check it out.
- Alex3917 18y agoThe archives of this mailing list has some good discussions about the stochastic approach: http://computer-go.org/pipermail/computer-go/ http://computer-go.org/pipermail/computer-go/ Apparently the strongest computer AI is now able to beat 1k players consistently.
- prospero 18y agoIt's worth noting that it can beat 1k players on a 9x9 board. On a 19x19 board (the standard board size), they are both weaker and less consistent.
- dood 18y agoOne of the strongest current programs, MoGo, recently beat a 5 dan player on 9x9 using a monte carlo approach. http://en.wikipedia.org/wiki/Computer_go#Monte-Carlo_methods http://en.wikipedia.org/wiki/Computer_go#Monte-Carlo_methods
- prospero 18y agoI'd seen that, but I haven't seen any published results of 1k players being routinely beaten at 19x19. If I've missed something like that, I'd be grateful for a link.
- nazgulnarsil 18y agoagreed: 9x9 is to real go what checkers is to chess.
- schtog 18y agoYes chess isn't really interesting out of an AI-perspective since a supercomputer can beat a human with brute force + heuristics. In Go you'd need something more like real AI.
- geebee 18y agoChess could be interesting from an AI perspective if you limited the amount of brute force searching allowed. Computer chess and human chess are, by and large, completely different games. I'm not sure I agree with this article about patterns in chess. The real masters think strategically as well as tactically, and use the structure of the pawns and pieces, combined with intuition, to guide their moves. Ultimately, I suspect that go and chess fire up similar regions of the human brain.
- pon 18y agoThey did MRI on Z.Polgar while she was playing. What they found out was chess thinking happens in a small region of the brain. The region is the same humans use when trying to recognize faces. Training modifies that machinery to recognize chess board positions too.
- geebee 18y agoHuh, that surprises me. I would have expected a chess master to use more of her brain than I do (not that I know how my own brain would look all wired up, but I have trouble believing that I don't think much more narrowly about the game - kind of how people who actually understand math approach factoring a set of equations differently from people who need a set of rules they can follow closely).
- eru 18y agoPerhaps they only looked at what parts of their brain experienced chess players used more?
- prospero 18y ago