01-29-2016, 04:59 AM
The wonderful thing about learning algorithms is that they can spot "patterns" so subtle that we don't even notice we've learned them ourselves, and could never express them in words.
We are *FAR* beyond attempting to score the board by any easily-expressible rule.
Go, like all turn-based games, has a self-organizing property. If you can make even a loose approximation of a "this board is better for white than that board" function, then using that board-evaluation function as the "leaf node" in a min-maxing tree one or two levels deep provides a much stronger approximation. And assuming there is any underlying pattern better than min-maxing, that stronger approximation is a pattern that a learning algorithm can learn. The learned pattern can be used as a better board evaluation function, rinse repeat.
I'm going to state confidently that there is NO turn-based game at which humans can consistently beat computers any more. Not once you've sicced a learning algorithm on refining the board-evaluation function and given it time to run.
That said, it will suck up a lot of computer power to learn iteratively better board valuations. It could take a warehouse full of machines months to do it. But once it's done, a single reasonably-powerful computer can use the learned pattern to beat any human master.
We are *FAR* beyond attempting to score the board by any easily-expressible rule.
Go, like all turn-based games, has a self-organizing property. If you can make even a loose approximation of a "this board is better for white than that board" function, then using that board-evaluation function as the "leaf node" in a min-maxing tree one or two levels deep provides a much stronger approximation. And assuming there is any underlying pattern better than min-maxing, that stronger approximation is a pattern that a learning algorithm can learn. The learned pattern can be used as a better board evaluation function, rinse repeat.
I'm going to state confidently that there is NO turn-based game at which humans can consistently beat computers any more. Not once you've sicced a learning algorithm on refining the board-evaluation function and given it time to run.
That said, it will suck up a lot of computer power to learn iteratively better board valuations. It could take a warehouse full of machines months to do it. But once it's done, a single reasonably-powerful computer can use the learned pattern to beat any human master.