Developing superior artificial board-game players is a widely-studied area of Artificial Intelligence. Among the most challenging games is the Asian game of Go, which, despite its deceivingly simple rules, has eluded the development of artificial expert players. In this paper we attempt to tackle this challenge through a combination of two recent developments in Machine Learning. We employ Multi-Dimensional Recurrent Neural Networks with Long Short-Term Memory cells to handle the multi-dimensional data of the board game in a very natural way. In order to improve the convergence rate, as well as the ultimate performance, we train those networks using Policy Gradients with Parameter-based Exploration, a recently developed Reinforcement Learning algorithm which has been found to have numerous advantages over Evolution Strategies. Our empirical results confirm the promise of this approach, and we discuss how it can be scaled up to expert-level Go players.
|Title of host publication||20th International Conference on Artificial Neural Networks|
|Publication status||Published - 08.11.2010|