Evolving Hierarchical Neural Networks to Play Go (1998)
The ancient Chinese game of go has fascinated people for thousands of years. More recently, people have attempted to create computer systems that play go. They have not met with great success. Today, the best computer go players cannot beat advanced human go players. Richards et al. (1997, in press) applied the SANE (Symbiotic, Adaptive Neuro-Evolution) system developed by Moriarty (1997) to playing go. He was somewhat successful, and this research is an attempt to extend their work, by extending the SANE architecture, and testing the extension in the game of go. The extension this thesis presents adds another level of hierarchy and modularity to the basic SANE algorithm. As currently implemented, it works approximately as well as the basic SANE algorithm, but does not improve on it. Several possible explanations for this are posited, and potential improvements are identified.
View:
PDF
Citation:
Technical Report HR-94-01, Department of Computer Science, The University of Texas at Austin, 1998.
Bibtex:

Todd Greer Undergraduate Alumni