On Jun 29, 2011, at 11:40 AM, Don Dailey wrote:
And I think this classifier is a Neural Netowrk that is built on the fly. It could be built in advance, but I think the power of it is that it's built from scratch before each move so it's much more relevant for the exact position that is being searched. I suspect it would be much less useful if it were generated in advance based on thousands of games.
It is not built from scratch each move; the weights learned during the playouts for previous moves still apply.
It sounds crazy to me that it works at all as it has no real knowledge of the position. However if you were given 10 or 20 moves without being able to see the starting position, you probably could deduce a lot of interesting things about the starting position. The most obvious is that you know some of the points which were not occupied. And if you assume the moves are reasonable moves then you can also deduce certain things that are probably true and false.
Note that the two previous moves on the actual board will probably not recur in succession later in the playout. The network's reaction to that move sequence is therefore mainly based on the first move in each playout.
However a NN can probably see patterns in this we would not even notice or consider. So I would not mind seeing the experiment expanded to longer move sequences.
See the paper. Responding to two moves works better than responding to one or three. (We were surprised by the latter result; we thought deeper history might slow things down, but we didn't expect it to harm performance.)
Peter Drake http://www.lclark.edu/~drake/ _______________________________________________ Computer-go mailing list [email protected] http://dvandva.org/cgi-bin/mailman/listinfo/computer-go
