Go remains a challenge for artificial intelligence. Currently, most machine learning methods tackle Go by playing on a specific fixed board size, usually smaller than the standard 19x19 board of the complete game. Because such techniques are designed to process only one board size, the knowledge gained through experience cannot be applied on larger boards. In this paper, a roving eye neural network is evolved to solve this problem. The network has a small input field that can scan boards of any size. Experiments demonstrate that (1) The same roving eye architecture can play on different board sizes, and (2) experience gained by playing on a small board provides an advantage for further learning on a larger board. These results suggest a potentially powerful new methodology for computer Go: It may be possible to scale up by learning on incrementally larger boards, each time building on knowledge acquired on the prior board.

PDF, PS

Proceedings of the Genetic and Evolutionary Computation Conference (GECCO-2004). New York, NY: Springer-Verlag, 2004