Applying ESP and Region Specialists to Neuro-Evolution for Go

Go is one notable board game where computer competence still trails behind that of human experts. In the past, neural-network-based approaches have shown promise. In this paper, the ESP variant of the SANE neuro-evolution algorithm was applied to go, and an alternate network architecture featuring subnetworks specialized for certain board regions was implemented. ESP produced simpler networks that performed just as well as the more complex ones produced by SANE in other studies. Having region-specialist subnetworks improved the already great performance marginally. However, both the simple network and the network with specialists failed to scale up to board sizes larger than 7x7.