Indirect encoding of neural networks for scalable Go

Abstract

The game of Go has attracted much attention from the artificial intelligence community. A key feature of Go is that humans begin to learn on a small board, and then incrementally learn advanced strategies on larger boards. While some machine learning methods can also scale the board, they generally only focus on a subset of the board at one time. Neuroevolution algorithms particularly struggle with scalable Go because they are often directly encoded (i.e. a single gene maps to a single connection in the network). Thus this paper applies an indirect encoding to the problem of scalable Go that can evolve a solution to 5 x 5 Go and then extrapolate that solution to 7 x 7 Go and continue evolution. The scalable method is demonstrated to learn faster and ultimately discover better strategies than the same method trained on 7 x 7 Go directly from the start. © 2010 Springer-Verlag.