Comment Re:Obligatory (Score 0) 368
Actually, there is a backgammon AI that consistently beats human grandmasters. You're correct that it uses neural networks, and that the author, Gerald Tesauro, has done some pretty cool things, but you're under-representing its success. Without digging through my old IEEE paper stacks, the first thing I can find is from more than a decade ago when TD-Gammon was already playing grandmasters to a tie http://www.research.ibm.com/massive/tdl.html.
I also happen to thing backgammon is more interesting in chess because it isn't fully deterministic, which also explains why neural networks would meet with more success than a decision tree approach. Most of the AI involved in playing strong chess involves looking as many moves ahead as possible (I know there is more to it than just that, but thats the thrust of the technique), and of course as computers get faster you can look further ahead. Eventually you're able to look far enough ahead to see all possible outcomes of the game from any position. To me this barely qualifies as AI. That gives the computer more than just perfect information about the state of the game, it has perfect information about every state of every game. It is less a prediction than a simple A to B to C map of a win. Thats why games that are non-deterministic (like backgammon or any game with random dice rolls) or games where the players are given incomplete information (think poker or blackjack, few boardgames deal with incomplete information because it would require a neutral referee and probably be cumbersome and boring) are far more interesting for the field of AI.
I also happen to thing backgammon is more interesting in chess because it isn't fully deterministic, which also explains why neural networks would meet with more success than a decision tree approach. Most of the AI involved in playing strong chess involves looking as many moves ahead as possible (I know there is more to it than just that, but thats the thrust of the technique), and of course as computers get faster you can look further ahead. Eventually you're able to look far enough ahead to see all possible outcomes of the game from any position. To me this barely qualifies as AI. That gives the computer more than just perfect information about the state of the game, it has perfect information about every state of every game. It is less a prediction than a simple A to B to C map of a win. Thats why games that are non-deterministic (like backgammon or any game with random dice rolls) or games where the players are given incomplete information (think poker or blackjack, few boardgames deal with incomplete information because it would require a neutral referee and probably be cumbersome and boring) are far more interesting for the field of AI.
