Seven chessboard knowledge of the Go board can be used to 19 road it?

[tanguocn]

Whether to make a more general Go program, the length and width of the chessboard can be freely defined, and the Go software is then trained to play chess on any size board. First train 7 chess, then 9, then 19?

[kwea123]

It sounds interesting to train a Go AI on rectangular (not squared) boards!

[pasky]

This could be absolutely possible in the convolutional part of the network, you'd need to retrain the fully connected stages at the very end though. (And possibly the additional network layers if you add more convolutions as the board grows.)

In Nochi's source code inherited from Michi, the goban size is not designed to be configurable at runtime, but it could be refactored to change this.