Others involved:
Funding: EPSRC Research Studentship £20,800
Board games such as chess, GO and naughts and crosses have been an important area of AI research since the pioneering work of Turing and Samuel. Turing realised that a computer could be programmed to play a game, and Samuel investigated how a program could learn to play draughts. This research focuses on the development a computational model of how people learn to play board games. Whilst much research has been done on modelling how performance improves in playing a single game, little research has been done on developing a general model which can learn a variety of board games. Board games are a good domain to do machine learning research in as other domains are much too complex to learn. A good model of how people learn to play games would have important implications in both education and in the development of more general machine learning systems.
The system to learn board games is based on a four layer model of the knowledge used in playing games:
1. Primitive features of the board position.
2. Concepts about the board position.
3. Strategies of how to play
4. Meta-strategies of how to choose between strategies.
The primitive features are built into the system as LISP code. Concepts are learned by examples given to the system. The strategies and meta-strategies are learned from the user in a rule based notation.
Graham Bevan is registered for a full-time PhD at the University of Glamorgan.
Page by Kevin H. Sewell
Last updated 24/October/95