New Algorithm for Computer Go?
Found via Pejman:
An interesting article on Computer Go. Go appears the least solvable of all of the traditional games. World-class computer opponents exist for Chess, Checkers, and Backgammon. The idea of training Neural Networks using a database of professional games and evaluating the position probabalistically seems genius. Who knows, this project may produce the first master level computer go opponent.
Neural nets worked for backgammon, after all. Last year I tried the free version of Jellyfish with staggering losses. Of course, I'm not a master player, but the fact that master Backgammon players consult with their (snarky) computers after a match says volumes.
I don't play Go much, despite it's depth and elegance. strong players often can't explain what makes a good move to weaker players. Learning involves study and bashing your head against the wall, and while I like studying games as much as anyone, I found the progress depressingly slow.
Update:
[Continuing laocoon's comment], I would imagine that the Neural Net would be used to generate candidate moves for strategic positions (particularly openings and for 'quiet' moves in general), the 'probability' function would be used to evaluate positions (using the NN to determine moves, likely counters, etc) and a tactical engine would handle localized fighting. [Even my old Sargon chess program in the early 80s always searched through the end of a capture sequence]. Endgames can also be solved handled analytically once the parts of the board cease to interact.