A stack of 11 3x3 convolutional layers and a single 5x5 layer with no pooling actually corresponds to effectively a 27x27 kernel, which is obviously large enough to cover the entire board. (Your value of 13 is only the distance from the center of the filter to the edge).
On Thu, Mar 10, 2016 at 10:48 PM, Huazuo Gao <gaohua...@gmail.com> wrote: > According to the paper *Mastering the Game of Go with Deep Neural > Networks and **Tree Search*, the main part of both the policy and value > network is a 5*5 conv layer followed by eleven 3*3 conv layer. Therefore, > after the last conv layer, the maximum "information propagation length" is > (5-1)/2 + 11*(3-1)/2 = 13, which is insufficient for covering the full > board. > > It might not have been a big problem though, as tree search and MC > rollouts should mitigate most deficiencies to a large extent. However, > during the opening, realising the correlation between distant stones would > be quite important, provided that tree search would not help much while MC > rollouts might not provide a unbiased view. > > It seems to me that DCNN are not perfect for Go. Anyway, apparently that's > enough for beating top human level. > > _______________________________________________ > Computer-go mailing list > Computer-go@computer-go.org > http://computer-go.org/mailman/listinfo/computer-go >
_______________________________________________ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
